Channeling Success with TinyML
Incorporating tinyML into LoRa communication helps IoT devices avoid interference and packet loss when transmitting data to edge computers.
Collecting data from geographically dispersed networks of sensors to support applications like environmental monitoring, smart agriculture, and infrastructure monitoring has never been easier, thanks to advances in Internet of Things (IoT) technologies. Given the wide range of applications that they support, it is no wonder that tens of billions of IoT devices now dot the globe. However, there are still a number of challenges that need to be solved in this area, with communication issues being near the top of the list.
Data collected from large, distributed sensor networks is of little value if it cannot be reported to either edge or cloud computing systems for further analysis. But transmitting this data is often easier said than done, given that a large number of the world’s IoT devices find themselves in areas where Wi-Fi, cellular, and other common communications networks are unavailable.
This has led many researchers and engineers to experiment with alternative communication methods, such as LoRa. It is easy to get started with LoRa transmissions, but it is often unreliable due to its unlicensed nature and limits on how often devices can transmit data. Better techniques are sorely needed for tasks such as channel selection, which is essential in avoiding interference and packet loss.
A trio of researchers at the Technical University of Braunschweig in Germany have put forward a potential solution to this problem that harnesses the power of tinyML. These highly-optimized artificial intelligence algorithms are capable of running on even highly resource-constrained computing systems, such as those found in IoT devices. In particular, the team developed a predictive model that was shown to be capable of reducing interference and improving the reliability of LoRa communications.
The team focused on improving the effectiveness of transmissions between IoT devices and edge servers, especially in the most challenging environments — densely populated areas, where competition for unlicensed frequency bands leads to interference and frequent packet collisions. To address this, the solution introduces a frequency-hopping mechanism driven by tinyML algorithms implemented directly on IoT devices. This mechanism dynamically identifies underutilized sub-frequencies within the available frequency band, balancing network utilization, minimizing collisions, and ensuring stable transmissions.
The model further predicts the best combination of LoRa transmission parameters (such as bandwidth, coding rate, spreading factor, and base frequency) to optimize data rate and range based on real-time environmental and network conditions. For instance, a spreading factor of 12 may be used to maximize range in low-interference scenarios, while a factor of 7 might enhance data rates in densely populated areas. Furthermore, the model enables IoT devices to adapt their communication strategies dynamically, ensuring consistent performance.
To evaluate their approach, the team created an experimental setup with three LoRa-enabled IoT devices that were positioned at varying distances (1m, 15m, and 30m) from a gateway to simulate real-world conditions. The devices transmitted data packets periodically over different frequencies while metrics such as signal strength, signal-to-noise ratio, and packet delivery ratio were measured at the gateway. The hardware utilized included Heltec LoRa modules for the IoT devices, Adafruit Feather microcontrollers for the gateways, and Raspberry Pi boards as edge computing nodes.
The new tinyML-based approach achieved up to 63 percent higher signal strength and 44 percent better signal-to-noise ratio compared to a random channel-hopping method. Additionally, an excellent packet delivery ratio demonstrated the model's ability to consistently select channels capable of transmitting all packets successfully. These improvements validate the tinyML model's effectiveness in learning patterns from channel usage data and optimizing LoRa communication in the IoT-edge continuum.