Technical newsheet
Industrial IoT & MonitoringDevice profile
ANCHOR Ships Gap-Free Maritime IoT Over Breaking Links With oneM2M and MEC
ANCHOR pairs oneM2M exactly-once delivery with MEC handover across a simulated Korea Strait, moving real model-boat telemetry between Busan and Kyushu edges in 0.23 s without losing readings.
By Grace Kim5 min read916 words
Features
- Busan-to-Kyushu MEC handover: 0.23 s from zone observation to first acknowledged store
- 10,000 successful sends produced exactly 10,000 requests in the mocked-transport test
- Anomaly model warned 32 s after injected drift; a fixed rule fired at 432 s
- Isolation Forest scoring runs in 13–15 ms per 10-second window on Raspberry Pi 5
- Disaster-aware federated learning excluded ship-001 for 30 rounds during a simulated typhoon

A maritime IoT system called ANCHOR delivered ship sensor data to shore without losing a single reading across a simulated Busan-to-Kyushu edge handover that completed in 0.23 seconds, from the moment the relay observed the new zone to the new edge acknowledging its first store. The team behind the project combined oneM2M resource semantics with ETSI multi-access edge computing (MEC) to close a basic gap in ship telemetry: a vessel at sea drops in and out of base-station coverage, and if readings from those dead intervals vanish, shore operators never learn what happened.
How does it deliver each reading exactly once?
Every sensor event gets an event ID built from the vessel, device, boot ID and sequence number. When a new vibration content instance (CIN) appears on the ship, a subscription notifies the ship relay, which commits the record to SQLite before attempting delivery. If the send fails or the relay restarts, the queued record stays for another attempt.
The oneM2M resource structure makes the retry safe. The ship relay uses HTTPS to create a named CIN at the receiving CSE; if the reply is lost, the relay can later retrieve that exact resource and verify what was stored. In a deliberate lost-acknowledgement test, Busan stored engine event ship-001:engine:boot:2 but the reply was dropped. The ship retried through Kyushu with the same resource name and body, Kyushu confirmed storage, and the ship cleared its queue. When both edges forwarded event 2 to shore, the second CREATE returned 4105 (resource exists); the forwarder retrieved the CIN by name, compared content, and matched. The three journals held exactly events 1, 2 and 3 at shore with original bodies intact.
Request accounting is deterministic: 10,000 distinct successful sends in a mocked-HTTP test produced exactly 10,000 requests, plus one RETRIEVE only when a retry hits 4105.
What hardware carries it?
The physical testbed is a model boat on a river in Seoul carrying three ESP32 nodes feeding a Raspberry Pi 5 gateway (ship MN-CSE), with a second Raspberry Pi 5 on shore as the IN-CSE:
- ESP32-S3 with MPU6050 and NEO-6M GPS measures bridge roll and pitch and reports position.
- ESP32-WROOM-32E with ADXL345 and MQ-2 samples engine vibration at 200 Hz and reports a summary plus gas-sensor voltage.
- ESP32-C6 with DHT11 and MC-38 magnetic switch tracks cargo temperature, humidity and hatch state.
Devices connect over MQTT secured with TLS, each with its own certificate; the relay onward uses HTTPS. Access control is per-request: the bridge identity CBridge can subscribe to commands but is denied when it tries to insert one — boards listen, the ship core issues.
What does edge handover buy?
Processing runs on MEC hosts close to the fleet, so collision warnings, typhoon avoidance and federated-learning participant selection happen near the ships rather than at a distant data center. The Korea Strait scenario in the ETSI MEC Sandbox splits the sea into three zones — kr-busan (mep1), jp-kyushu (mep2), and vsat open sea with no MEC host — served by LTE-M cells and satellite points of access. The relay polls MEC 013 location every two seconds and switches destination as the vessel crosses zones; the ferry-hakata twin appeared on mep2 about 1 second after it attached to the Fukuoka cell, while the old path on mep1 returned 404.
Disaster simulation exposed one implementation lesson: because degradation was originally added to current values, latency stacked to 801 ms at the Busan station across repeated disasters. The fix reads the original scenario values before applying degradation.
What did the AI components measure?
On-board anomaly detection with Isolation Forest models scores a 10-second window every second in about 13–15 ms on the Raspberry Pi 5; training widens the normal band by each sensor's rated accuracy (±2 °C, ±5 %RH for the DHT11). In a synthetic engine-temperature drift experiment, the model warned 32 seconds after onset where a fixed rule took 432 seconds. A retrained model hot-swapped onto the ship 25 seconds after publication, without restart.
The Fleet AI scanner on MEC reroutes only when predicted closest approach falls under 0.2 nm within 8 minutes on two consecutive scans. In 45 minutes of normal traffic it logged seven danger warnings and no actual approach within 0.25 nm. In one reroute, predicted closest approach was 0.0 nm; measured distance after the intervention was 1.75 nm.
Disaster-aware federated learning excluded ship-001 from round 13 of run t5 after a typhoon core covered its Busan cell, ran two participants for thirty rounds, and readmitted the vessel in round 43. The final model version listed three vessels and 180 trees.
What limits remain?
An ESP32 buffers at most 32 unsent records in RAM, so a power cut before the Pi receives a sample can lose it. The federated model trains on synthetic 30-second windows, not the 10-second windows the physical ship uses, and the FL coordinator re-reads the whole update container each poll — one run stalled near 970 updates. The team's next measurement is a full voyage comparing ship and shore IDs and reporting the delay distribution alongside any missing records.
The adoption question the prototype raises: will fleet operators accept SQLite-journaled, exactly-once oneM2M delivery as the reference pattern for regulated maritime telemetry, where the shore journal — not radio availability — defines the record?
via hackster.io (Source)
Filed under
- maritime-iot
- onem2m
- multi-access-edge-computing
- sensor-telemetry
- exactly-once-delivery
More from Grace Kim
Show full bio
Correspondent covering companies, deals and strategy in Instruments, test and measurement at Testbench Report.
118 articles
Application notes
- oneM2M Made Indispensable: Swarm Project Ties Robot Swarms to 5G Edge Handover
- Object Controllers Move Railway Signalling Maintenance Off Site
- STMicroelectronics Puts Neural-Network Inference On-Chip in Industrial Vibration Sensor
- ISA-95 Five Levels Map the Path From Shop Floor Sensors to ERP
- Trackside Sensors Feed Control Room Decisions in Rail Monitoring Push