Abstract
Detecting weak seismic tremors across a network of sensors is a significant challenge in geophysics, primarily because signals arrive at different sensors at different times and with varying intensities. We propose an Asynchronous Subspace-CUSUM procedure that jointly estimates unknown signal waveforms and relative sensor delays. By synchronizing sensor data, our method effectively boosts the signal-to-noise ratio (SNR), allowing for the detection of subtle geophysical events that are often invisible to traditional 'one-shot' detection schemes. We demonstrate the efficacy of this approach through both synthetic simulations and real-world seismic data from Parkfield, California.
The Challenge of Asynchronous Detection
In monitoring scenarios like seismology, a signal originating from a single source does not reach every sensor simultaneously. Because seismic waves travel at finite speeds through heterogeneous ground, different sensors observe the event with varying time delays ($\tau$) and amplitudes ($\alpha$).
Standard detection methods often rely on 'one-shot' schemes, which trigger an alarm as soon as any single sensor detects a change. However, seismic tremors are frequently low-amplitude events buried in background noise. For these 'weak signals,' a single sensor may lack the sensitivity to distinguish the tremor from noise. To improve detection, data from the entire network must be combined, but this requires accurate knowledge of the relative time delays and the underlying signal waveform—both of which are typically unknown.
The Subspace-CUSUM Approach
Our proposed method, the Asynchronous Subspace-CUSUM, addresses these challenges by treating the delays and the signal shape as unknown variables to be estimated jointly. The procedure employs an iterative algorithm that alternates between two key steps:
- Delay Estimation: A Maximum Likelihood approach is used to estimate the relative time offsets between sensors.
- Signal Estimation: Singular Value Decomposition (SVD) is applied to the sample covariance matrix to identify the 'principal component' or the most likely waveform of the signal.
By synchronizing the sensors, the algorithm transforms a collection of weak, asynchronous observations into a single, high-SNR synchronous vector. This 'global' approach exploits the collective information of the network, making it far superior to local detection schemes in low-SNR environments.
Performance Evaluation
We evaluated the Subspace-CUSUM procedure against the standard one-shot scheme using synthetic data from a 50-sensor network. The results demonstrate a clear advantage for our method, particularly when dealing with weak signals.
As shown in the performance plots, the Subspace-CUSUM consistently achieves lower average detection delays for the same average run length (false-alarm rate). This performance gap is most pronounced in the weak-signal regime ($\mu = 0.1$), where the collective processing of the sensor network provides the necessary boost to identify the change-point quickly.
Real-World Application and Results
The algorithm was also tested on real-world data from 13 seismic sensors in Parkfield, California. The detector successfully identified three known earthquakes with high precision. Furthermore, it picked up several low-frequency tremors that were previously cataloged but difficult to detect using standard automated methods. This confirms that the Subspace-CUSUM is not only theoretically sound but also practically effective for monitoring subtle geophysical activity.
Conclusion
The Asynchronous Subspace-CUSUM procedure provides a robust framework for multi-sensor change-point detection. By integrating synchronization and signal estimation into the detection process, it overcomes the limitations of traditional methods in detecting weak, asynchronous signals. Future work will focus on the computational efficiency of the sliding window approach and proving the asymptotic optimality of the asynchronous case.
Summary updated