EFFICIENT NEURAL NETWORK ALGORITHMS FOR MICROCONTROLLERS IN AUDIO DETECTION SYSTEMS

This study focuses on the development and deployment of optimized neural network algorithms for real-time audio-based drone detection on resource-constrained microcontroller systems. Addressing the growing need for efficient drone detection, the research aims to design lightweight models that balance accuracy, low latency, and energy efficiency for edge devices such as STM32 and ESP32 microcontrollers. The study evaluates neural network architectures, including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and lightweight models like MobileNetV2 and TinyCNN. GRU-based RNNs achieved the highest accuracy of 98%, while MobileNetV2 offered a balance between performance and efficiency. Optimization techniques such as quantization and pruning were applied, enabling quantized MobileNetV2 to achieve inference speeds of 45 FPS and energy consumption of just 3 mW per inference. These results underscore the practicality of deploying such models in real-world scenarios. Integration of these models onto microcontrollers was facilitated by frameworks like TensorFlow Lite and STM32Cube.AI. Field tests demonstrated the system’s robustness in diverse environments, including noisy urban areas. Detection accuracy exceeded 90% within a 100-meter range, even under adverse conditions. The results highlight the system’s potential for low-power, autonomous surveillance applications. Key contributions include demonstrating the effectiveness of neural network optimization for edge systems, creating a scalable framework for audio-based detection, and advancing lightweight models for energy-efficient tasks. Future research will focus on advanced optimizations, expanding datasets, and integrating multi-modal detection. This study lays a foundation for practical, efficient, and scalable drone detection technologies, addressing key challenges in energy use, accuracy, and real-world deployment.