<![CDATA[Real-time seismic monitoring in urban and remote environments is increasingly challenged by strong anthropogenic noise and limited computational resources at field stations. Although deep learning-based seismic denoising methods have demonstrated promising performance, their deployment on low-power edge AI platforms for continuous onsite monitoring remains insufficiently explored. This study investigates the feasibility of implementing a Conv2D-based Denoising Autoencoder (DAE) for real-time seismic signal denoising on a Raspberry Pi 5 integrated with the Hailo-8L AI accelerator. The model was trained using synthetic noise injection on seismic waveform data from the CIKJI station to generate noisy-clean signal pairs. The deployment pipeline included signal preprocessing, Conv1D-to-Conv2D adaptation, ONNX conversion, and INT8 quantization into the Hailo Execution Format (HEF). Experimental evaluation on 21,598 seismic windows demonstrated that the Hailo INT8 implementation achieved a Pearson correlation of 0.9645 and an SNR of 14.95 dB, compared to 0.9727 and 17.08 dB obtained by CPU FP32 inference, respectively. Despite a modest degradation in denoising accuracy, the Hailo-8L significantly reduced CPU utilization from 61.4% to 14.5% and maintained stable sub-millisecond inference latency during 24-hour simulated streaming tests. These results demonstrate that edge AI accelerators provide a practical tradeoff between denoising performance and computational efficiency, supporting the development of low-power real-time seismic monitoring systems for onsite deployment in resource-constrained environments. This study provides the first empirical characterization of NPU-accelerated seismic denoising under continuous real-time streaming conditions, establishing practical benchmarks for edge AI deployment in resource-constrained seismic monitoring networks.]]>
