The Model

Our machine learning model uses a convolutional neural network - LSTM to identify seismic events amidst the continuous seismic data collected by seismometers. By using both a Butterworth bandpass filter (BBF) and an auto-encoder, our model, which is trained on the seismic data provided by NASA, can identify seismic events with increased accuracy. It also generates data for the frequency domain (from time domain, using Fast Fourier Transform), and also uses this data to calculate the power at each sample of data. With these two extra benchmarks, our model can better identify seismic events due to their nature (having a frequency within the range of 0.5 - 3 Hz, and having much more power than noise).

Our Objective

We aim to increase the energy efficiency of transferring seismic data from probes on extraterrestrial bodies back to Earth. Currently, due to the large quantities of unfiltered and continuous data, the process is time-consuming and energy expending due to the long travel distance and low transfer bit rate. By filtering this data, and sending only the scientifically useful seismic events, our model would be able to increase the overall efficiency of the data transfer and also the analysis process. The reduced quantity of data which needs to be transferred would reduce the power expenditure of the probes and reduce the time that researcher require to analyze the data (compared to unfiltered data containing noise from the seismometer).

By saving energy during the data transfer process, it would open the opportunity for allocating that saved energy towards collecting higher resolution seismic data. This would in turn further improve the data analysis process by also providing researchers with better data to work with.

upload_image.11:39:23.620744

^{Comparison of 1 day worth of data (raw data against filtered data)}

By filtering the data, only ~ 1/300th of the original file size would need to be sent.

The Design

Our model is entirely software based, with the aim of practical application. As in, our model can (after optimization for the OS of probe) be uploaded as an update towards currently active probes and also used in the development of future probes.

The model was developed in Visual Studio Code using python (Jupyter Notebook) for developing the whole model from CNN - LSTM to the auto-encoder and other data preprocessing functions.

2024 NASA Space Apps Challenge

Acedia