ARCHITECTURE OF A DISTRIBUTED CENTRALIZED ALERTING SYSTEM WITH EDGE-BASED ACOUSTIC UAV DETECTION IN HARD REAL-TIME MODE

This paper presents the architecture and implementation principles of a distributed computerized system for centralized alerting where primary acoustic detection is executed on resource-constrained edge nodes and the control center performs aggregation, visualization, logging, and system management. The target application is early warning based on acoustic signatures of low-flying unmanned aerial vehicles (UAVs) under severe background noise, multipath reflections, and variable propagation conditions. Unlike approaches that stream raw audio to the server, the proposed solution follows an event-driven concept: the edge node continuously processes the audio stream, converts it into time–frequency representations, runs a lightweight neural inference pipeline, and transmits compact detection events with timestamps and confidence values to the center. This reduces bandwidth requirements and supports deterministic behavior required for hard real-time alerting. A key contribution is the combination of physical and algorithmic means to improve robustness. On the sensing side, passive acoustic focusing is used to increase the signal-to-noise ratio (SNR) without increasing computational load or power consumption. The resulting enhancement improves the separability of UAV patterns in the time–frequency domain and stabilizes the decision function under non-stationary noise. On the inference side, a hybrid CNN–RNN model is applied: convolutional layers extract local spectral patterns while the recurrent block captures temporal dynamics typical for UAV sounds. To meet edge constraints, integer quantization (int8) is employed to reduce memory footprint and inference latency while keeping the detection accuracy within acceptable limits.Networking is integrated into the hard real-time contour of the system. The paper proposes a deterministic event delivery strategy based on the uIP lightweight TCP/IP stack, focusing on predictable packet processing time and a minimized RAM footprint, which is critical for embedded nodes. Experimental results confirm the feasibility of the approach for operational alerting scenarios: detection accuracy reaches up to 98.2% in challenging acoustic conditions, passive focusing provides an SNR gain of approximately 12–15 dB, and the selected network stack reduces RAM usage by up to 40% compared with a heavier embedded alternative while improving latency stability (jitter). Overall, the proposed architecture demonstrates a controlled trade-off between detection quality, latency, jitter, and resource consumption, making it suitable for scalable centralized alerting systems with edge intelligence.