Embedded AI
What the model does and what data it processes
The function of the model and the kind of data it reads decide the architecture, the preprocessing and the hardware. We define both first, together with what the device should do with the result.
· by BCF Embedded engineering team
Functions
Classification
Assigns a class: machine state, gesture, sound type, product defect.
Anomaly detection
Learns normal operation and flags deviations; needs few examples of faults.
Regression and estimation
Estimates a value: remaining useful life, state of charge, a virtual sensor.
Detection and segmentation
Locates objects or events in an image or a signal.
Processed data
Sensor time series
Vibration, current, temperature, IMU. Windowing, filtering and spectral features.
Audio
Spectrograms and MFCC for keyword spotting, sound events and voice activity.
Images and video
Visible and thermal cameras. Resizing, normalisation and the camera pipeline.
Sensor fusion
Several sources combined, such as IMU with radar or environmental sensors.
Preprocessing is part of the model
On embedded devices feature extraction can cost as much time as the inference itself. We implement it in optimised C / C++ (for example with CMSIS-DSP) and keep it identical between the training pipeline and the firmware, so the model sees the same data in the field as in the lab.
Cascades save energy
One device can run several functions in a chain: a cheap voice-activity or motion detector wakes a heavier model only when needed. Most of the time the expensive part stays asleep.
FAQ
Discuss your project
Describe your device and goal. An engineer replies within one business day.