Neural network-based Bluetooth synchronization of multiple wearable devices

The increasing use of Bluetooth-enabled wearable devices for physiological data collection, especially in healthcare, necessitates accurate synchronization of multiple devices. Existing Bluetooth Low Energy (BLE) synchronization methods struggle to achieve the required accuracy for applications involving multiple wearables performing group activities. This paper addresses this limitation by proposing an application-level solution that avoids placing undue burdens on hardware. While alternatives such as ANT+, WiFi, or Zigbee could be integrated, these protocols typically increase power consumption and size, and require non-standard auxiliary receivers. BLE, being widely used and providing standardized secure connectivity, is a preferable solution. The paper focuses on using BLE in connected mode via the Generic Attribute Transfer (GATT) protocol, leveraging its encryption capabilities and customizability for command control, along with Real-Time Operating Systems (RTOSs) for efficient resource management in wearable devices. However, challenges remain due to time-sensitive parameters like temperature, power variations, aging, and crystal reference drifts in wearables, coupled with the non-determinism of the Bluetooth stack (anchor points and retries). These factors contribute to unpredictable time shifts during packet transmission, hindering mutual synchronization. To overcome these challenges, a novel neural network (NN) is proposed.

The paper reviews existing time synchronization protocols (TSPs) for wireless sensor networks, including the Flood Time Synchronization Protocol (FTSP), Reference Broadcast Synchronization (RBS), and Timing Sync Protocols for Sensor Networks (TPSN). It notes that while these protocols offer methods for fine-tuning wearable clocks, handling clock ticks at runtime, and applying time delays, they are not suitable for energy-constrained portable systems employing RTOSs. Directly modifying the hardware clock can lead to failures due to the impact on multitasking and scheduling mechanisms within the RTOS. The non-deterministic nature of the Bluetooth stack, including variations in anchor points and retries, is highlighted as another significant challenge in achieving precise synchronization among multiple BLE devices. Existing literature on BLE synchronization often involves low-level parameter handling or requires configuring each wearable individually, lacking a universal solution.

The research employs Kinematics Detectors (KiDs), high-performance motion-tracking devices, as hardware and a custom-designed User Interface (UI) as software. KiDs continuously acquire data at 200 Hz for up to 2.5 hours and support the BLE 4.2 protocol. The UI allows for remote control of multiple KiDs, sending commands such as START, STOP, and MARK (for labeling events) via BLE 4.2. The methodology involved evaluating temporal shifts in unsynchronized KiDs by vertically stacking them on a fixed plane to ensure identical spatial domains, enabling simpler analysis of temporal uncertainties. Data acquisition includes recording time stamps on both the remote system and devices to calculate temporal shifts. These shifts were primarily attributed to frequency drift among devices, Bluetooth latency due to retransmission protocols, and non-deterministic jitter. To address these challenges, a three-layer neural network with 20 neurons was implemented at the application level. The NN is trained using a time-stamped series data representing various sources of time shifts. The trained NN predicts and compensates for these shifts, effectively acting as a virtual clock layer, allowing for the accurate time synchronization of multiple KiDs. The NN was trained using RMSE and tested on multiple platforms. The MARK command enables tagging of captured data with labels, facilitating artifact extraction and simplifying data analysis. Experiments involved multiple KiDs performing a hand motion sequence, with the MARK command used to identify specific tasks within the acquired data. Data sampling frequency was calculated, and latency was measured to quantify the impact of Bluetooth communication on synchronization. The performance of the NN-based synchronization was evaluated across different numbers of KiDs (2, 3, and 4) and across multiple platforms.

The study demonstrates that without the neural network-based synchronization, the average synchronization error across multiple KiDs was approximately 30ms. The frequency drift among devices, the non-deterministic nature of Bluetooth re-transmissions (retry protocols), and the inherent latency in BLE communication are major contributors to these synchronization errors. The implemented three-layer neural network successfully reduced the average synchronization error to 1.25 ms, enabling synchronized operation of multiple KiDs at 200 Hz. The neural network demonstrated a high correlation (R-value = 0.9967) between predicted and actual times. The synchronized network showed a 97.3% success rate across various platforms (PC, Linux, and Mac). Synchronization failures primarily resulted from OS software updates impacting the UI and complete battery exhaustion in the KiDs. The MARK command was shown to be effective in labeling specific tasks within the acquired motion data, facilitating data analysis and artifact extraction. The system successfully demonstrated scalability to four KiDs, showing the capacity for multi-limb, multi-person event-oriented motion capture.

The application-level neural network solution presented here successfully addresses the challenges of synchronizing multiple BLE wearable devices without requiring low-level hardware modifications. This approach avoids the risks associated with altering low-level parameters in RTOS-based systems, which can negatively impact timer-critical subroutines, event synchronization, multi-task scheduling, and time-stamped data accuracy. The ability to achieve sub-millisecond synchronization accuracy using a software-based approach opens new opportunities for research requiring high-frequency, synchronized data from multiple wearable sensors. The system’s capability to collect temporally precise data while handling numerous labels simplifies the analysis of complex movements and multi-agent interactions.

This research presents a novel application-level solution for synchronizing multiple Bluetooth-based wearable devices using a neural network. The system successfully reduced synchronization error to 1.25ms, enabling high-frequency (200 Hz) synchronized motion capture. The integrated labeling mechanism simplifies data analysis, making it suitable for studying complex naturalistic behavior. Future research could focus on further optimizing the neural network architecture, exploring integration with other sensor modalities, and investigating the performance in diverse real-world scenarios.

The study primarily focuses on a specific type of wearable sensor (KiD). While the approach is designed to be adaptable to other BLE devices, its performance with different hardware and software configurations needs further investigation. The synchronization success rate (97.3%) indicates that a small percentage of failures can still occur under certain conditions (OS updates, complete battery drain). Future work could focus on addressing these remaining failure scenarios. The study was conducted in a controlled laboratory setting, which might not fully reflect the real-world variability of wireless signal strength and noise levels.