Exploring UWB Positioning & TinyML Edge AI for Smart Asset Tracking

Why precise asset tracking matters

Businesses lose billions each year because they cannot locate tools, inventory or shipments in real time. Digital asset‑tracking systems help prevent losses, optimize workflows and maintain compliance. Traditional solutions rely on GPS, RFID or Bluetooth Low Energy (BLE) beacons. BLE beacons like Eelink’s devices are cost‑effective and battery‑efficient but typically offer sub‑meter accuracy only at short distances and up to 20–30 m range. For high‑value goods or safety‑critical workflows, centimetre‑level precision is required. Two emerging technologies—Ultra‑Wideband (UWB) positioning and TinyML/Edge AI—address these limitations. To learn more about our asset‑tracking products, visit Eelink’s official website.

Ultra‑Wideband fundamentals

Ultra‑Wideband (UWB) is a radio technology that transmits very short (≈2 ns) pulses across a wide 500 MHz channel. It operates in the 3.1–10.6 GHz frequency range and uses time‑of‑flight measurements to compute distance. Unlike narrowband radios, UWB’s impulse nature enables extremely precise timestamping of signal arrival times, providing centimetre‑level accuracy. The pulses spread energy across a wide bandwidth and appear as background noise to other radios, improving coexistence. UWB topologies such as two‑way ranging (TWR), time‑difference of arrival (TDoA) and phase‑difference of arrival (PDoA) offer flexibility in deployment. In typical line‑of‑sight conditions, UWB can measure distances up to 50 m while covering large areas with fewer anchors than BLE‑based systems.

UWB is also secure. Its physical layer uses AES encryption and scrambled timestamp sequences to prevent relay attacks. These properties make it suitable for digital car keys, secure building access and precision asset tracking.

UWB vs. Bluetooth LE positioning

BLE is widely available and energy‑efficient, but its positioning techniques—RSSI trilateration, angle‑of‑arrival (AoA) or angle‑of‑departure (AoD)—are susceptible to multipath interference and provide coarser estimates. Future BLE channel‑sounding techniques may improve accuracy, but advanced algorithms increase power consumption and reduce battery life. BLE channel sounding can deliver sub‑metre accuracy at short distances but requires more anchors for a given area and still trades range for accuracy. Consequently, many implementations pair BLE for coarse localisation (e.g. finding a general zone in a warehouse) with UWB for fine ranging within the zone.

UWB’s centimetre‑level precision and strong security come with slightly higher cost and power usage than BLE. The choice depends on application requirements: high‑value asset tracking, robotics and augmented reality benefit from UWB, whereas low‑cost tags and simple presence detection often suffice with BLE.

TinyML and Edge AI: bringing intelligence to the edge

Tiny Machine Learning (TinyML) is an emerging field that deploys machine‑learning models on ultra‑low‑power microcontrollers. Instead of sending sensor data to cloud servers, models run locally, providing fast inference with low latency and preserving data privacy. Because inference happens on the device, connectivity is optional—crucial for remote asset trackers. TinyML devices typically have kilobytes or megabytes of memory, which limits model size and complexity. Frameworks like TensorFlow Lite, PyTorch Mobile and Edge Impulse optimise models to fit within those constraints.

Advantages and challenges

TinyML’s advantages are rooted in on‑device analytics: low latency, reduced bandwidth costs, improved privacy and the ability to operate without continuous connectivity. However, developers face challenges: limited memory and compute budgets restrict the type of models (often requiring quantisation or pruning); debugging on remote devices is harder; and there is a lack of standard tooling compared with cloud ML. Nevertheless, the field is rapidly advancing and is supported by frameworks that automate optimisation and code generation.

Case study: predictive maintenance with TinyML

Edge Impulse’s 2022 case study illustrates how TinyML can transform asset monitoring. An engineer collected audio samples from an AC motor using an Arduino Nano 33 BLE Sense board (32‑bit ARM Cortex‑M4 CPU, 256 kB SRAM, 1 MB Flash). They trained a neural‑network classifier to recognise four classes—normal operation, two types of anomalies and background noise—achieving about 94–95 % classification accuracy. The model processed one‑second audio frames and was compressed to run on the microcontroller. When deployed, the device illuminated a red LED when a maintenance issue was detected. This example demonstrates how TinyML enables predictive maintenance without streaming high‑bandwidth audio to the cloud; inference happens locally and only small alerts are transmitted.

Integrating UWB and TinyML for smart asset tracking

Combining UWB’s accurate ranging with TinyML yields smart trackers that not only locate assets but also understand their state. A typical architecture might include:

1. UWB ranging chip (e.g. DecaWave DW3000) providing centimetre‑level distance to anchors.
2. Microcontroller with TinyML (e.g. ARM Cortex‑M4 or M7) running models trained to recognise patterns in sensor data (vibration, temperature, orientation). These models can detect anomalies or classify asset states (e.g. item in transit vs. stationary). The predictive‑maintenance case study shows that complex models can run on devices with 256 kB RAM.
3. Energy management – Advanced algorithms can increase power consumption, so designers must balance model complexity against battery life. Techniques such as model quantisation, sparsity and event‑driven inference help conserve energy. For instance, run inference only when the accelerometer detects motion or when the UWB distance changes significantly.
4. Connectivity – The tracker may use BLE or Wi‑Fi to report status occasionally while relying on UWB for location. MQTT remains a popular protocol for publishing results to dashboards.
5. Security and privacy – UWB uses AES‑based security and scrambled timestamps to resist spoofing; TinyML keeps raw data on the device, preserving privacy.

Precision ranges, model selection and trade‑offs

When designing a smart asset tracker, it is important to understand the trade‑offs between precision, model complexity and power consumption:

• Precision ranges – UWB can localise tags to within a few centimetres in line‑of‑sight conditions and up to about 50 m range. BLE with channel‑sounding may achieve sub‑metre accuracy but typically covers 20–30 m. Choose the technology based on whether centimetre or metre precision is acceptable and whether additional anchors are feasible.
• Model selection – TinyML models range from simple threshold detectors and linear classifiers to quantised neural networks. For vibration monitoring (as in the predictive‑maintenance example), a small neural network is appropriate. For temperature or humidity anomalies, simpler statistical models may suffice, conserving memory and power. More complex models (e.g. convolutional neural networks for vision) consume more energy; consider using them only when necessary.
• Power trade‑offs – UWB chips consume more power than BLE because they transmit wideband pulses. However, transmissions are short and sporadic, so average power can be low. Similarly, TinyML inference consumes energy; by running inference only on events and using optimised models, battery life can remain acceptable.

Best practices for deploying smart asset trackers

1. Hybrid localisation – Use BLE for discovery and coarse positioning and activate UWB only when fine localisation is needed. This conserves power while providing high accuracy when required.
2. Efficient ML pipelines – Use frameworks like Edge Impulse to create and optimise models for microcontrollers. Employ quantisation (e.g. 8‑bit weights) and pruning to reduce model size. Trigger inference based on sensor thresholds rather than running continuously.
3. Adaptive sampling – Adjust sensor sampling and transmission intervals based on asset state. When stationary, reduce sampling frequency; when motion or unusual vibrations are detected, increase it. Many asset‑tracking beacons allow broadcast intervals to be configured via over‑the‑air commands.
4. Security – Protect ranging and data transfer. Leverage UWB’s AES‑based security and encrypt MQTT traffic using TLS. Ensure firmware updates are signed to prevent malicious modifications.
5. Data fusion – Combine UWB distance measurements with inertial‑measurement‑unit (IMU) data (acceleration, orientation) and apply TinyML models to detect asset activity (e.g. forklift carrying vs. idle). The interplay between sensor fusion and ML classification yields more contextual insights.

Conclusion

UWB positioning and TinyML are transforming asset tracking. UWB offers centimetre‑level accuracy and robust security, outperforming traditional BLE in environments where precise positioning matters. TinyML allows microcontrollers to run ML models locally, achieving low‑latency inference, data privacy and independence from cloud connectivity. Combined, these technologies enable devices that know exactly where they are and what state they are in, while running for months or years on a single battery. As UWB chip costs decrease and TinyML tooling matures, smart asset trackers will become ubiquitous in manufacturing, logistics, healthcare and smart buildings. Engineers should carefully balance precision requirements, model complexity and power budgets, choose hybrid localisation strategies and leverage modern frameworks to deploy reliable, intelligent trackers. With these considerations, the next generation of IoT devices can deliver unprecedented visibility and insight into physical operations.