Skip to main content

Release Notes: Vehk Axle Platform v2.3.0

· 2 min read
Vehk Engineering Team
Platform Creators

We are excited to announce Vehk Axle platform v2.3.0—this is the cloud fleet product (dashboard, ingest, device tooling, and prediction services), distinct from the SDK package line vehk-axle which versions on its own. This release focuses on expanded device compatibility, a clearer Device Configurator experience, and a hardened, CPU-bound ML stack.

Device Configurator​

Onboard and manage hardware from the dashboard:

  1. Open axle.vehk.in and go to Settings → Device Config.
  2. Use the tabs:
    • Management — live status, SIM/gateway view, OTA, disconnect alerts
    • Devices — all registered devices, status, message volume
    • Add — guided 3-step onboarding
    • Channels — listener status by protocol
    • Diagnostics — telemetry shadow and raw messages
    • Bulk — CSV onboarding
    • Credentials — shared tenant credentials for large fleets

Full field reference and HTTP/MQTT paths: Device Configuration guide.

Expanded IoT Device Protocols​

The Universal Payload Normalizer has been significantly upgraded. We now offer robust, native support for 150+ telematics device schemas, opening the door for massive fleet compatibility without requiring custom development work.

Newly native integrations include:

  • Wialon IPS: Full API coverage for any Wialon-compliant telemetry hardware.
  • Teltonika & Queclink: Direct TCP/UDP parsing for FMB920, FMC130, GV55, and GL300 models.
  • Concox, Jimi IoT, & iTriangle: Native capabilities handling GT06N, WeTrack2, and Bharat101 devices.
  • CalAmp & Ruptela: Full UDP/TCP parsing for Trace5 and LMU series hardware.

Hardened CPU-Bound ML Architecture​

We completely revamped the vehk-ml prediction engine to run exclusively in lightweight, 100% CPU-bound container apps. This cuts down on deployment complexity while preventing OS dependency crashes.

Graceful degradation & heuristics:
When the ML module encounters situations where core libraries (for example libomp or lightgbm) fail to load—or if a vehicle has not streamed enough historical data yet—the prediction system falls back to heavily optimized heuristic algorithms. That keeps prediction and Remaining Useful Life (RUL) surfaces available even when the full model stack is unavailable.

For more technical specifics, see the Axle Engine docs.

Minor fixes and adjustments​

  • Updated VEHK_ML_CPU_ONLY=True defaults to strip unused CUDA footprint in ML images.
  • Upgraded MongoDB archive loaders to support efficient multi-tenant historical merges.
  • Updated standard prediction API payloads to reflect fallback and heuristic usage flags for clients.