Selected as global nominees
Take a look at our team and whole project here: https://www.spaceappschallenge.org/nasa-space-apps-2024/find-a-team/sylar/?tab=project
Project Summary
For our project, we adopted a traditional function analysis methodology to detect seismic events, drawing on our signal processing expertise rather than implementing the commonly suggested convolutional neural networks (CNNs). Through the application of targeted filters and the derivation of a characteristic function, we achieved accurate identification of seismic activity onset. This approach makes our solution particularly well-suited for resource-constrained environments such as space missions, where power availability and energy harvesting remain persistent challenges. By avoiding the computational overhead of AI-driven models, we deliver a practical, reliable, and energy-efficient solution for real-time seismic monitoring.
Project Demonstration
Project
https://github.com/sylar-lab/SylarSismo-Analyzer
Project Details
THE CHALLENGE AND THE TEAM
Our challenge centered on developing analytical software to process authentic seismic data from two landmark space exploration missions: the Apollo lunar missions and the Mars InSight Lander. The objective was to detect seismic events—subtle vibrations and ground movements that hold critical information about the internal structure of the Moon and Mars—amid substantial background noise.
Our team assembled a diverse yet complementary skill set: a software engineer with expertise in audio signal processing, a hardware engineer specializing in fundamental signal component filtering, a radio-frequency engineer experienced in low signal-to-noise ratio (SNR) environments, and an entrepreneur in journalism who contributes innovative, unconventional perspectives to problem-solving.
OUR APPROACH
Upon initial examination of the challenge materials—from the datasets provided to the terminology in submission guidelines—it became evident that the competition framework anticipated neural network and edge AI implementations as the primary solution path.
However, following thorough evaluation and discussion, we determined that our collective technical foundation positioned us to pursue a classical functional analysis methodology instead. We believed this approach would prove equally effective, if not superior, when leveraged through our specific domain expertise. Consequently, we concentrated on employing multiple filtering techniques to construct a characteristic function capable of precisely detecting seismic event onsets.
Through iterative development, we recognized that strategic isolation of relevant signal components would enable accurate identification of seismic initiation points without the architectural complexity of neural networks. This direction reinforced our commitment to foundational engineering principles and validated our conviction that elegant, streamlined solutions—grounded in rigorous technical understanding—often outperform more elaborate alternatives.
The result of this philosophy is demonstrated below: a 400-line Python implementation processing NASA’s miniSEED data format:
[IMAGE 1: FINAL RESULT VISUALIZED IN THE DIFFERENT STAGED]
So, how did we achieve the characteristic function and why is it so unique in addressing this challenge?
THE PROJECT
To provide comprehensive insight into the characteristic function construction process, we developed visualizations documenting each critical stage required to establish optimal filtering and triggering parameters, determined through bisection methods.
The first visualization presents normalized raw data. This baseline plot enabled us to assess the fundamental signal structure and ambient noise characteristics of the unprocessed seismic data, establishing a reference point prior to any signal conditioning or filtering operations.
[IMAGE 2: RAW DATA PLOT AND DATA ENVELOPE]
Next, to determine the appropriate bandpass filter and remove all frequencies that do not fall within the typical range of seismic events, we plotted the Fast Fourier Transform (FFT) of the signal. By analyzing the distribution of frequencies through these FFT plots, we were able to identify the dominant frequency bands associated with seismic events.
[IMAGE 3: FAST FOURIER TRANSFORM]
From this analysis, we saw that the frequency ranges of seismic activity vary between the Moon and Mars due to differences in their geophysical properties. As a result, we designed a separate filter for each case—one tailored for lunar seismic events and another optimized for Martian seismic activity.
[IMAGE 4: FITTED PARAMETERS]
We then applied the bandpass filter to the raw data, which allowed us to generate a much cleaner spectrogram, an evident indicator of seismic activity.
The filtered spectrogram made it easy to visually discern the presence of seismic events, but it wasn’t useful for the functional analysis we wanted to perform since it’s made of scattered points.
We took the next step by converting the spectrogram into a more structured representation: a spectral content plot
[IMAGE 5: SPECTROGRAM AND RELATIVE POWER PLOT]
The spectral content plot (red line) was derived by computing the root mean square (RMS) of the vertical spectrogram components. This transformation enabled time-resolved quantification of “relevant” frequencies—specifically, those frequencies retained following bandpass filtering operations. In essence, this process converted the multidimensional frequency-time data from the spectrogram into a univariate signal that clearly delineates temporal instances of significant spectral energy.
While this yields another oscillating function with considerable variation, it represents a crucial advancement: we now possess a mathematically tractable model of seismic activity amenable to further analysis. To this spectral content signal, we applied a median filter with a 15-sample window [blue line], effectively suppressing residual noise artifacts and outliers attributable to environmental interference.
[IMAGE 6: CLOSE UP OF MEDIAN FILTERING THE NOISE SPIKES]
The median filter works by sliding a window—the length of which varies based on the planet and is defined in the MOD_PARAM section that was previously shown — across the spectral content plot. For each window, it replaces the central value with the median of all the values in that window.
As a result, we were able to eliminate short, sharp impulses that didn’t correspond to genuine seismic activity, leaving behind a cleaner and more accurate representation of the seismic events we were trying to detect.
PS: look at how happy we were when we found out the peaks were finally gone:
[IMAGE 7: THE MOMENT THE PEAK WAS GONE]
The final step in our process was constructing the characteristic function, which we based on the previously refined spectral content function, and finally apply a threashold
[IMAGE 7: CHARACTERISTIC FUNCION AND THRESHOLD-TRIGGERING]
Here’s a rewritten version:
To achieve this refinement, we implemented two critical modifications to the spectral content signal.
First, we applied a second median filter, utilizing the global median computed across the entire data frame. This frame-wide median established a dynamic baseline for signal smoothing, effectively attenuating residual low-amplitude fluctuations. By referencing the global median, we successfully mitigated low-frequency drift and long-term biases that could otherwise mask seismic event detection.
Following median filtering, we performed amplitude normalization to enhance peak prominence relative to the background signal level. This normalization facilitated the establishment of a robust detection threshold capable of reliably discriminating genuine seismic activity from residual noise.
The resulting characteristic function yields a normalized signal in which each peak corresponds to a seismic event. This framework enables automated onset detection through a fixed threshold value (configurable within the MOD_PARAMS section), ensuring consistent performance and accuracy across diverse datasets.
CONCLUSIONS, INNOVATIONS AND ADVANTAGES
Our classical signal processing methodology offers substantial advantages over convolutional neural network (CNN) implementations, particularly regarding computational efficiency and power consumption.
CNNs, while highly effective for pattern recognition tasks, impose significant computational and memory demands that render them suboptimal for resource-constrained systems. Conversely, our approach employs a sequence of computationally efficient operations—fast Fourier transforms (FFTs), median filtering, and normalization—all readily optimizable for low-power hardware architectures.
The algorithmic simplicity of our solution enhances system reliability and ensures consistent seismic event detection without the risks of model overfitting or the overhead of continuous retraining. Notably, our preliminary CNN experiments revealed critical limitations: the constrained NASA dataset induced severe overfitting to specific signal characteristics, resulting in substantially elevated false detection rates.
In summary, our solution achieves an optimal balance between detection precision and energy efficiency, positioning it as the pragmatic choice for real-time seismic monitoring in power-limited environments such as planetary exploration missions.