Solution · AI Sensor Fusion

AI Sensor Fusion Solution
When the Product Senses More Than Light

A sensor product rarely fails because the model was wrong. It fails because the pressure sensor drifts with temperature, because the two sources were never time-aligned, because the analog front end sits next to an antenna, or because nobody defined what "the signal is missing" means. BesTom designs that chain: sensor selection, calibration and compensation, front-end layout, duty-cycled power, data-quality gates, and on-device fusion inference on Rockchip MCU-class and Linux-class platforms.

Physiological Environmental & fluid Inertial & radar Vision-inertial fusion Calibration & compensation Duty-cycled low power
Sensor domains

What we build around

Four domains, and the same discipline in each: know the error budget before choosing the part.

Physiological

PPG, ECG, body temperature, blood oxygen. Contact quality and motion artefact dominate signal quality far more than sensor grade. See wearable vital-sign monitoring.

Environmental & fluid

Temperature, pressure, flow, gas, humidity. Cross-sensitivity is the core problem — a flow or pressure reading that ignores media temperature is a reading you cannot act on.

Inertial, radar & ultrasonic

IMU, mmWave presence sensing, ultrasonic ranging, magnetic. Selection turns on whether you need absolute position or only rate of change.

Vision-inertial fusion

Camera plus IMU for stabilisation, dead reckoning or activity recognition. Time alignment between the two is the whole problem — see Smart Vision.

Distributed sensor nodes

Wireless mesh and battery nodes reporting into one host. Clock drift across nodes and packet loss are the design constraints, not radio range.

Always-on wake layer

An MCU-class core keeps sensing and decides when the Linux-class SoC wakes. Average current is set here, not by the application processor.

Platform selection

Where the fusion runs

The same product can put feature extraction on a microcontroller and fusion on a Linux SoC. Getting that split right is what makes the power budget achievable.

Role in the productPlatform classWhy it sits there
Always-on acquisition & feature extractionMCU-class — e.g. RK2108D (Cortex-M4F + HiFi3 DSP), as used in the Bestom M08D audio moduleSampling, filtering and wake decisions must run at microamp-to-milliamp average. DSP handles filtering and feature extraction without waking the SoC.
Sensor plus camera in one nodeRV1106B / RV1106 (Cortex-A7 + MCU, SiP DDR)The MCU side covers continuous sensing while the A7 side handles light vision — one part instead of two, and no external memory rail to power.
Real-time closed-loop controlRK3506 (3× Cortex-A7 + M0)The M0 co-processor drives real-time I/O (CAN / UART / SPI) while Linux runs the UI and gateway logic. Sensor actuation loops must not wait on Linux scheduling.
Audio sensing & acoustic analysisRK3308 / RK2118 audio platformsMicrophone arrays, PDM and SAI interfaces with an audio-NPU where acoustic features carry the signal. See our RK3308 acoustic AI cases.
Multi-modal fusion & inferenceRK3576 / RK3588 (6 TOPS NPU)When several sensor streams must be fused with a learned model, or fused with vision and a local LLM. See Edge AI Host.
Wireless sensor meshWi-Fi / BLE mesh companion SoC alongside the hostNode-level radio and clock behaviour are designed separately from the application SoC so mesh firmware can be validated on its own.
Exact platform parameters follow the official Rockchip datasheet for the specific part number. The Rockchip SoM matrix lists the ranges we design within.
Engineering

The six places sensor products actually break

None of these appear in a demo. All of them appear in the field.

#Failure modeHow we design against it
1Cross-sensitivity ignored — a pressure or flow reading is reported as truth while the medium temperature moves underneath itCharacterise the cross-sensitivity matrix on the bench, then compensate in firmware. Ship the calibration procedure, not just the coefficient.
2Sources never time-aligned — streams are fused as if they were simultaneous when they are tens of milliseconds apartDefine the timing budget up front: hardware trigger or timestamped capture, and a stated acceptable skew per stream. Validate skew by measurement.
3Analog front end next to a radio — everything works on the bench and the noise floor rises the moment the link is activeSmall-signal routing, ground partitioning and shielding reviewed together with antenna placement and transmit duty cycle, not as separate tasks.
4Long-term drift — accurate in week one, out of tolerance by month six, and no recalibration path existsDecide the drift budget and the recalibration strategy before part selection. If the product cannot be recalibrated in the field, the sensor grade must be bought with money, not firmware.
5No definition of bad data — missing, saturated or contact-lost samples are fed to the model as if they were validA data-quality gate before inference: validity flags, range checks, confidence and a defined behaviour for each failure class. What the product does when it does not know matters more than its best case.
6Power budget built on active current — the datasheet number looks fine and the battery still dies in weeksBudget the average, not the active. Duty cycle, wake settling time and false-trigger rate decide the real number; measure it over a full cycle, not a snapshot.
Deliverables

What you actually receive

StageDeliverableAcceptance criterion
FeasibilitySensor selection report with error budget, cross-sensitivity risk list and platform splitMeasured data from at least one candidate part in your operating range
HardwareSchematic, PCB, analog front-end and RF coexistence reviewNoise floor measured with the radio transmitting, not only idle
FirmwareAcquisition layer, filtering, calibration and compensation routinesCompensated output within the agreed tolerance across the full temperature and range envelope
Data qualityValidity flags, anomaly handling and degraded-mode behaviour specEach failure class is injected deliberately and the product responds as specified
Fusion / inferenceFeature pipeline, model quantization and deploymentAccuracy held within the agreed threshold after quantization, with the false-positive rate reported
PowerDuty-cycle design and measured consumption profileAverage current over a full representative cycle, including wake settling
Mass productionCalibration fixture, one-point/two-point test SOP, drift sampling planLine-level calibration repeatability recorded and signed off
Why BesTom

Why BesTom

Design house since 2008. We are a design service, not a storefront — finished-product volume runs through MeToclaw, the product brand of the same group.

Sensor products already delivered

Wearable vital-sign monitoring, acoustic sensing on RK3308, and multi-sensor industrial measurement — documented under client cases where the client allows it.

MCU and Linux in the same team

The always-on layer and the application layer are designed together, so the power split is decided by engineering rather than negotiated between two suppliers.

Own audio and DSP module

The M08D module (RK2108D + RK962) exists as a building block when the product needs microphones, DSP and voice in the same node.

Calibration treated as a product feature

Bench characterisation, firmware compensation and production-line calibration are designed as one chain with the same error budget.

Shenzhen prototype-to-volume chain

Rapid prototyping plus contracted manufacturing, with DFM, thermal and EMC work in the same loop as the sensing design.

We will tell you what not to measure

If the error budget says a sensor cannot meet the requirement at the target price, we say so before you commit to tooling.

Related

Related Resources