Variational Autoencoders for P-wave Detection on Strong Motion Earthquake Spectrograms

TL;DR

This work addresses robust P-wave detection in strong-motion data by reframing arrival detection as self-supervised anomaly detection using variational autoencoders. Through an extensive grid search over four VAE architectures, it shows that enforcing global context with a self-attention bottleneck yields the best detection performance (AUC up to 0.875, 0–40 km AUC ≈ 0.91), at the cost of reconstruction fidelity, while skip-connections improve reconstruction but can erode discrimination due to overgeneralization. The study systematically analyzes how latent capacity and Transformer hyperparameters shape stability and performance, revealing that a hybrid of convolutional inductive biases and attention provides a robust, practical alternative. These findings suggest architectural biases that prioritize global structure over pixel-perfect reconstruction are essential for reliable, self-supervised P-wave detection with potential for real-time earthquake early warning applications.

Abstract

Accurate P-wave detection is critical for earthquake early warning, yet strong-motion records pose challenges due to high noise levels, limited labeled data, and complex waveform characteristics. This study reframes P-wave arrival detection as a self-supervised anomaly detection task to evaluate how architectural variations regulate the trade-off between reconstruction fidelity and anomaly discrimination. Through a comprehensive grid search of 492 Variational Autoencoder configurations, we show that while skip connections minimize reconstruction error (Mean Absolute Error approximately 0.0012), they induce "overgeneralization", allowing the model to reconstruct noise and masking the detection signal. In contrast, attention mechanisms prioritize global context over local detail and yield the highest detection performance with an area-under-the-curve of 0.875. The attention-based Variational Autoencoder achieves an area-under-the-curve of 0.91 in the 0 to 40-kilometer near-source range, demonstrating high suitability for immediate early warning applications. These findings establish that architectural constraints favoring global context over pixel-perfect reconstruction are essential for robust, self-supervised P-wave detection.

Figures (8)

• Figure 1: Detection performance versus reconstruction quality for the four VAE architectures tested in our experiments. Detection is measured by the Area Under the Receiver Operating Characteristic Curve ($\mathit{AUC}$, higher is better), and reconstruction is measured by the Mean Absolute Error ($\mathit{MAE}$, lower is better). Models that reconstruct more accurately tend to detect anomalies less, effectively demonstrating a trade-off relationship.
• Figure 2: Unified diagram of the proposed VAE architectures. Components marked with an asterisk (*) are optional, defining four configurations: Basic-VAE (baseline encoder--decoder), Skip-VAE (incorporating long-range skip connections), Attention-VAE (incorporating the self-attention bottleneck), and Hybrid-VAE (incorporating both). The encoder and decoder are symmetric, where each standard residual block consists of two convolutional layers with ReLU activation and a local skip connection.
• Figure 3: Original and Noise- and Artifact-Detected Earthquake Accelerogram Record. Top Panel: Original seismic record with noise and detected P and S phases. Middle Panel: Simulated FBM noise. Bottom Panel: Final augmented signal combining the original seismic data with the synthesized noise.
• Figure 4: Normalized Cross-Correlation Analysis. (a) No significant peak is observed, minimizing false positives. (b) A distinct peak corresponding to the P-wave arrival is present. (c) Peaks are distinct for both P- and S-wave arrivals, demonstrating robustness under overlapping signals.
• Figure 5: Comparison of Receiver Operating Characteristic ($\mathit{ROC}$) curves for the four VAE architectures. The Attention-VAE achieves the highest performance, followed by the Hybrid, Skip, and Basic models.