VECTORALAB EDGE

TinyML & Health Score

1D-CNN INT8 — inference in under 12 ms — fully edge, no cloud

TinyML

On-Device TinyML Inference

A 1D Convolutional Neural Network trained on labeled vibration feature vectors is quantized to INT8 and deployed via X-CUBE-AI. Inference runs entirely on the STM32H750 — no network connection required.

Model Type

1D Convolutional Neural Network (1D-CNN)

Input Shape

float32[32] feature vector

Output

Softmax probability over 3 health classes

Quantization

INT8 (post-training quantization)

Framework

TensorFlow Lite Micro + X-CUBE-AI

Inference Latency

< 12 ms per frame

Model Size

< 32 KB Flash

RAM Usage

< 16 KB during inference

[²] The INT8 model is generated using TensorFlow Lite for Microcontrollers and X-CUBE-AI's Neural ART Accelerator. Accuracy degradation vs float32 baseline is < 1.5 % on the validation set.

92/ 100

Health Score

Health Score — 0 to 100

The planned TinyML model is intended to support continuous condition assessment of the monitored machine. Final scoring behavior and maintenance thresholds will depend on model validation and machine-specific operating conditions.

80–100

Good

Normal operating condition. No action required.

50–79

Degrading

Early wear signs detected. Plan maintenance within scheduled window.

0–49

Critical

Significant fault signatures. Immediate inspection recommended.

[³] Health Score thresholds are configurable via UART command interface. Default thresholds (80 / 50) are derived from empirical testing on benchmark bearing datasets.

Communication

Heartbeat Burst — Host Communication

After each inference cycle, the STM32H750 transmits a compact data burst to the host platform. The burst contains the Health Score, raw feature vector, and a status byte — enabling the VectoraLab platform to log trends and trigger alerts.

UART / SPI / CAN (selectable)Protocol

921600 bps (UART default)Baud Rate

Every frame (~38 ms)Burst Period

Health Score + float32[32] featuresPayload

137 bytes per burstPacket Size

CRC-16/IBMChecksum

Edge Intelligence

Local Decision — No Cloud Required

Every inference decision is made on-device. The STM32H750 maintains full monitoring capability in air-gapped environments, offline deployments, and low-bandwidth industrial networks.

Data Privacy

Raw vibration data never leaves the device. Only the Health Score and aggregated features are transmitted.

Ultra-Low Latency

Fault detection within one frame cycle (~38 ms) — no round-trip to a cloud API.

Air-Gap Ready

Operates without any network connection. Fully functional in SCADA-isolated or OT network environments.

Low Power Mode

STOP2 sleep between frames reduces average power to < 15 mW in intermittent monitoring configurations.