Abstract
Distinguishing human footstep signals from environmental noise is a critical challenge for Unattended Ground Sensor (UGS) systems, particularly as the signal-to-noise ratio (SNR) degrades with distance. This paper proposes a novel Parallel Recurrent Neural Network (PRNN) architecture to enhance seismic target recognition. The PRNN integrates a time-domain feature network, optimized for running signals, and a frequency-spectrum feature network, optimized for walking signals. Experimental results demonstrate that this dual-branch approach achieves a human recognition accuracy of 98.3%, significantly outperforming traditional single-model machine learning methods.
Introduction
Unattended Ground Sensor (UGS) systems are vital for monitoring remote areas, relying on seismic waves to detect intruders. However, these systems often suffer from high false alarm rates due to low SNR and the high variability of human movement signatures. A person walking generates a different seismic profile than one running, making it difficult for a single algorithm to generalize across all scenarios. To address this, we present a Parallel Recurrent Neural Network (PRNN) that processes temporal and spectral features simultaneously.
System Architecture
The proposed system follows a structured pipeline: data acquisition, feature extraction, and target recognition via the PRNN. The core innovation lies in the parallel processing of seismic data to capture both rhythmic temporal patterns and distinct spectral signatures.
Feature Extraction and Processing
Before classification, raw seismic signals undergo specialized preprocessing to highlight relevant characteristics for each neural network branch:
- Time Domain Path: Raw signals are processed through a fast bandpass filter (25–50 Hz) followed by peak value region extraction. This emphasizes the rhythmic, high-amplitude bursts typical of running.
- Frequency Domain Path: The Welch algorithm is employed to calculate the Power Spectral Density (PSD), identifying the spectral "humps" (150–350 Hz) that characterize walking movements.
The Parallel Recurrent Neural Network (PRNN)
The PRNN utilizes two specialized Long Short-Term Memory (LSTM) branches:
- Time Domain Feature Network: Trained to identify the temporal dependencies and rhythms of footfalls.
- Frequency Spectrum Feature Network: Trained to recognize the "color" or frequency distribution of the signal.
The outputs of these branches are fused using a logical OR operation. This synergy ensures that if either specialized branch detects a human signature, the system triggers an alert, effectively lowering the omissive report rate.
Experimental Results and Validation
The model was validated using a custom dataset collected from MEMS seismic sensors buried in hard soil. Data from five individuals moving within a 30-meter radius provided over 16,000 samples for training and testing.
The PRNN approach demonstrated an "absolute advantage" over single-network designs:
- Accuracy: 98.3% (vs. 91.6% for time-only and 73.7% for frequency-only models).
- False Alarm Rate: Reduced to approximately 0.8%–1.0%.
- Reliability: The omissive report rate was kept under 1%, ensuring high security for UGS applications.
Conclusion
By leveraging the complementary strengths of time-domain and frequency-domain analysis through a parallel LSTM architecture, the PRNN significantly advances the state of seismic target recognition. Future work will focus on adapting the model to diverse terrains (such as sand or mud) and enhancing its ability to track multiple simultaneous targets.
Summary updated