Embedded AI
Embedded AI: machine learning on the device
Models that run where the data is produced - on microcontrollers and edge processors - optimised for memory, power and latency.
· by BCF Embedded engineering team
Why edge, not cloud
Latency
Decisions are made on the device, without a round trip to a server.
Energy and transmission
Sending results instead of raw data saves radio time and battery.
Privacy
Sensitive data such as audio or images can stay on the device.
Four decisions behind every edge AI project
Model sizing
Flash, RAM, compute and energy budget of the model.
Hardware targeting
Which silicon runs it and how we optimise for it.
Learning & distillation
How the model learns from the data you have.
Functions & data
What the model does and which signals it reads.
By data type
Time series and sensors
Vibration, current, temperature: anomaly detection and condition monitoring.
Audio
Keyword spotting, sound event detection, voice activity detection.
Vision
Object detection and visual inspection on processors with enough compute.
By hardware class
The hardware class sets what is possible. We help choose it before the model is designed.
Tiny Edge
- Hardware
- Cortex-M0/M4 class, < 256 KB RAM
- Power
- Coin cell / energy harvesting
- Typical use
- Simple anomaly detection, wake-word triggers
Small Edge
- Hardware
- Cortex-M7/M33/M55 class, 512 KB – 2 MB RAM
- Power
- Small batteries
- Typical use
- Keyword spotting, sensor fusion
Mid Edge
- Hardware
- MPUs with NPU, 512 MB – 4 GB RAM
- Power
- Mains / large battery packs
- Typical use
- Real-time video analytics, gateway AI
Heavy Edge
- Hardware
- Edge GPUs and industrial PCs, > 4 GB RAM
- Power
- Mains, active cooling
- Typical use
- Multi-camera video, larger models
Tools
Frameworks
- TensorFlow Lite Micro
- Edge Impulse
- STM32Cube.AI
Languages
- Python
- C / C++
Quantization and optimization
Most models are trained in floating point and must be converted before they fit a microcontroller. We quantise models (typically to INT8), prune where it helps and measure accuracy, memory and latency on the target hardware - not only in simulation.
Use cases
Predictive maintenance is the closest use case for our focus sectors: detecting changes in vibration or current of pumps, motors and power electronics in maritime and energy installations.
FAQ
Discuss your project
Describe your device and goal. An engineer replies within one business day.