Design and Validation of a Synchronized Heterogeneous Wireless Sensor Network: From Acquisition to Real-Time Inference

Modern AIoT, used in industry and smart cities, depends on fast, reliable data generated by numerous wireless sensors. These sensors often have to contend with the energy limitations of battery-powered devices. In practice, low-power devices face three interrelated challenges: MEMS noise and distortion; unpredictable delays and losses on low-energy links; and limited edge processing capabilities. These factors create misaligned or missing samples that degrade sensor fusion and real-time analytics. This thesis presents a wireless sensor network framework that improves data quality by enforcing precise time alignment while supporting embedded workloads. A central hub ingests two data sources: Bluetooth Low Energy (BLE) motion streams from sensor nodes and HTTP streams that carry labels and lower-rate signals. Time synchronization (TS) runs entirely at the application layer through periodic exchanges of timestamped packets between the hub and the nodes. TS offers three methods: a host timestamp baseline, a basic offset estimator that rejects outliers, and a Kalman filter that tracks offset and drift. The modular stack includes a real-time GUI for monitoring and interactive labeling, structured logging, config-driven replay, and export to CSV and Parquet. Validation uses three complementary use cases. Two share the same sensor hardware: ST SensorTile.box and SensorTile.box PRO nodes streaming at 75 Hz over BLE to an STM32MP257x hub running OpenSTLinux. The first is a controlled bench that quantifies alignment error, end to end latency, jitter, packet loss, and throughput while comparing the three synchronization engines. The second is a real world bicycle study that uses road surface classification as a workload to assess end to end data and synchronization quality; labels are provided by a Flutter app as timestamped HTTP annotations to the hub. The third demonstrates multi-sensor integration by adding additional BLE and HTTP sources, including ECG, EEG, and MetaWear devices, and aligning all heterogeneous streams on a shared timeline. Results show millisecond-level alignment under stable conditions, with the Kalman approach most robust to clock drift and outliers. Packet loss, latency, and jitter remain within limits compatible with real-time processing at 75 Hz per node. Without synchronization, misalignment can exceed one second, confirming the need for active time correction with relevant problems for the reliability and accuracy of Machine learning models predictions. Field trials align sensor streams with external references, such as GPS and camera, and confirm temporal consistency. An optional audio-assisted mode that uses onboard microphones and FFT tone tagging integrates cleanly for event-triggered labeling. The contributions consist of a hardware-agnostic synchronization module for multi-device BLE sensing. A flexible hub merges BLE and HTTP sources on a shared timeline with real-time monitoring, labeling, and logging. It also interoperates with heterogeneous sensors such as IMUs and physiological devices from different vendors. An empirical evaluation demonstrates millisecond-level alignment and sustained real-time operation on commercial edge hardware. In conclusion, this framework provides a practical foundation for synchronized sensing and embedded analytics in next-generation AIoT deployments.