Detailed Summary of the Seismic Event Detection System

This project is centered around developing an efficient system to detect and analyze seismic events on planetary bodies like Mars or the Moon. By combining traditional phase-picking techniques with modern machine learning, the system is designed to optimize data transmission from distant space missions while ensuring high accuracy and reliability.

This system blends tunable conventional algorithms with a Convolutional Neural Network (CNN) to provide an end-to-end seismic event detection and analysis workflow. The conventional algorithms handle the initial stages of noise filtering and phase-picking, while the CNN is responsible for the final refinement, ensuring that only the most relevant seismic events are retained.

Initial Picking(Conventional Algorithms)

Finding the Best Bandpass Frequency
The system automatically finds the optimal bandpass filter window for each seismogram. This can be done using either power spectrum analysis or standard deviation analysis, with parameters that can be fine-tuned for different missions or environments.
Applying the Bandpass Filter
The selected frequency window is applied to filter out irrelevant frequencies, keeping the important parts of the seismic signal.
Removing Outliers
Short, high-amplitude noises are removed to prevent interference with the normalization step.
Data Normalization
The data is normalized, placing it on a consistent scale for easier analysis.
STA/LTA Analysis
The system uses STA/LTA to detect potential seismic events by comparing short-term activity to long-term background noise.
Filter Out Picks
If there are too many potential events, it filters them by analyzing the characteristic function, distinguishing between events and noise based on the slope of the rise and fall.

The cleaned and processed data is then cataloged for further analysis by CNN(Convolutional Neural Network).

Final Event Refinement Using CNN

The final step involves passing the detected events through a Convolutional Neural Network (CNN). This lightweight, yet powerful, machine learning model acts as the last layer of filtering, differentiating between real seismic events and noise that survived earlier steps. Though currently in early stages of training, the CNN already delivers impressive results and will continue to improve with more data and tuning. CNN takes a total of 2 inputs, a Seismogram with shape (5565,1) and Auxilary inputs, which were initially set to Standard deviation values before and after the Arrival time.

Efficiency and Scalability

The program is built to be highly efficient in both time and computational resources. It can process a month’s worth of seismic data in under 60 seconds on an average Processor, ensuring rapid data analysis on low-power, space-bound computers. Additionally, the entire workflow for individual events is completed in under a second, ensuring minimal delay between event detection and data transmission. Or it can detect events in last month of data directly on command from Ground Station with Some defined Parameters by which the algorithm is tuned.

Tunable for Different Missions

Similarly with these parameterized variable same program can be deployed on various Missions.

Origionally the I wanted to train CNN inside a GAN( Generative Adversial Network) Specifically a Auxillary Classifier Conditional GAN, and extracting the Discriminator from it as my CNN, inspired from Li et al. (2018) but due to time limitation, a CNN was trained directly on data with some Auxillary inputs to boost accuracy.

Results

In this brief development period, our Convolutional Neural Network (CNN), despite being minimally trained and primarily relying on heuristic methods, achieved results that exceeded expectations. It successfully identified over 80% of seismic events while maintaining a false positive rate of only around 5% from the total false picks generated by the initial filters.

Anyone can explore our GitHub repository to analyze the results, make predictions on their own datasets, or even retrain the CNN. Everything technical about the code itself is there in detail.

Conclusion

This project brings together the best of traditional signal processing and modern machine learning to deliver an innovative solution for seismic event detection in planetary missions. With its rapid processing times, computational efficiency, and customizable parameters, it aims to enhance the capabilities of autonomous space exploration.

For more technical details and updates on the project, explore our GitHub repository.

2024 NASA Space Apps Challenge

Betelwise