Seed-Driven FPGA Edge Node for Adaptive Seismic Intrusion Monitoring

Large-scale seismic intrusion monitoring demands robust feature extraction from weak, time-misaligned signals under strict edge power and memory constraints. We propose a seed-driven hardware-algorithm co-design based on Frequency-Domain Random Convolution Classification (FDRCC). The pipeline processes 1024-sample frames via streaming Fast Fourier Transform (FFT) magnitude extraction, dual-mode Top-$M$selection, and a weight-stationary ridge classifier. Crucially,$K=100$random kernels are deterministically regenerated on-the-fly from a 64-bit seed. This eliminates kernel storage and enables dynamic sensor adaptability by updating only the seed plus$\sim$23.7 kB of ridge weights, avoiding heavy hardware reconfiguration. Implemented on a Xilinx ZCU104 edge sensing node, the accelerator supports aggregated sensor channels (capacity of hundreds, limited to$\sim$120 by I/O throughput). Validated with real-world LGT-4.5 geophone data, the system achieves 89.49% INT16 accuracy (90.3% FP32) with 2.33 W on-chip power at 100 MHz and 0.4 ms deterministic latency.