Xi-Net: Transformer Based Seismic Waveform Reconstructor

Anshuman Gaharwar, Parth Parag Kulkarni, Joshua T. Dickey, Mubarak Shah

Abstract

Missing or erroneous data is a major problem in today's world. Collected seismic data sometimes contain gaps due to a multitude of reasons like interference and sensor malfunction. Gaps in seismic waveforms hamper further signal processing to gain valuable information. Plethora of techniques are used for data reconstruction in other domains like image, video, and audio, but translation of those methods to address seismic waveforms demands adapting them to lengthy sequence inputs, which is practically complex. Even if that is accomplished, high computational costs and inefficiency would still persist in these predominantly convolution-based reconstruction models. In this paper, we present a transformer-based deep learning model, Xi-Net, which utilizes multi-faceted time and frequency domain inputs for accurate waveform reconstruction. Xi-Net converts the input waveform to the frequency domain, employs separate encoders for time and frequency domains, and one decoder for getting reconstructed output waveform from the fused features. 1D shifted-window transformer blocks form the elementary units of all parts of the model. To the best of our knowledge, this is the first transformer-based deep learning model for seismic waveform reconstruction. We demonstrate this model's prowess by filling 0.5-1s random gaps in 120s waveforms, resembling the original waveform quite closely.

Introduction

Seismic sensors are critical for monitoring planetary activity, yet the data they produce is frequently compromised by "gaps"—missing signal segments typically caused by sensor malfunctions or environmental noise. These gaps, often lasting between 0.5 to 1 second within a 120-second recording window, significantly impede subsequent signal analysis. While image inpainting is a mature field, seismic waveforms present unique challenges: they are high-frequency 1D signals where long-term dependencies are vital. Traditional Convolutional Neural Networks (CNNs) often struggle to capture these long-range patterns effectively. Furthermore, treating a waveform as a 2D image for inpainting is computationally inefficient, as the signal data occupies less than 1% of the pixel space in a standard plot.

Methods

The proposed architecture, Xi-Net, is named after the Greek letter Xi ($\Xi$) due to its three-pillar structure: a Time-domain encoder, a Frequency-domain encoder, and a joint Decoder.

The Xi-Net architecture featuring dual encoders for time and frequency domains and a shared decoder.

Key technical innovations include:

  • Multi-Domain Fusion: The model processes raw signals alongside frequency data generated via Discrete Time Fourier Transform (DTFT). To handle complex numbers within a transformer framework, the real and imaginary components are stacked as separate "channels," analogous to RGB channels in computer vision.
  • 1D Swin Transformer Blocks: The model adapts the Shifted Window (Swin) Transformer mechanism for 1D data. This includes 1D Window Shifting to capture cross-window connections and 1D Patch Merging/Expanding for hierarchical feature extraction.
  • U-Net Style Architecture: Skip connections pass high-resolution features from both encoders directly to the decoder, preserving fine-grained details that might otherwise be lost during the bottleneck phase.

Experimental Setup

The study utilized a dataset of 13,696 seismic signals provided by the Air Force Technical Applications Center (AFTAC), with 11,000 samples for training and 2,696 for testing. Signals were upsampled by a factor of 2 and cleaned using a Butterworth Bandpass filter. To simulate real-world conditions, random gaps ranging from 0.5 to 1.0 seconds were inserted into the 120-second waveforms.

Key Results and Discussion

Xi-Net outperformed the 1D CNN baseline across all evaluated metrics. Notably, the Mean Range Difference (MRD) improved from 0.2758 to 0.1861, and the Discrete Fréchet Distance (DFD) was reduced to 0.182.

Qualitative results showing (a) full reconstruction, (b) pattern detection, and (c) cases with no reconstruction.

Qualitatively, the model demonstrates three levels of performance:

  1. Full Reconstruction: Matching the original signal almost perfectly.
  2. Pattern Detection: Recovering the general oscillatory shape even if the exact amplitude varies slightly.
  3. No Reconstruction: Failing to capture the missing segment, often due to extreme signal complexity.

Beyond accuracy, Xi-Net offers a massive leap in efficiency. It reached peak performance in just 7 hours of training on a single NVIDIA V100 GPU, whereas the 21-layer CNN baseline required 27 days to achieve comparable results.

Conclusion

Xi-Net represents the first pure-transformer approach to seismic waveform reconstruction. By leveraging both time and frequency domains, it achieves superior accuracy and significantly faster training times than convolutional alternatives. Future work will focus on reducing "edge jitter" near reconstructed gaps and exploring Generative Adversarial Network (GAN) integrations to further enhance the realism of the generated signals.

Summary updated

Abstract

Missing/erroneous data is a major problem in today's world. Collected seismic data sometimes contain gaps due to multitude of reasons like interference and sensor malfunction. Gaps in seismic waveforms hamper further signal processing to gain valuable information. Plethora of techniques are used for data reconstruction in other domains like image, video, audio, but translation of those methods to address seismic waveforms demands adapting them to lengthy sequence inputs, which is practically complex. Even if that is accomplished, high computational costs and inefficiency would still persist in these predominantly convolution-based reconstruction models. In this paper, we present a transformer-based deep learning model, Xi-Net, which utilizes multi-faceted time and frequency domain inputs for accurate waveform reconstruction. Xi-Net converts the input waveform to frequency domain, employs separate encoders for time and frequency domains, and one decoder for getting reconstructed output waveform from the fused features. 1D shifted-window transformer blocks form the elementary units of all parts of the model. To the best of our knowledge, this is the first transformer-based deep learning model for seismic waveform reconstruction. We demonstrate this model's prowess by filling 0.5-1s random gaps in 120s waveforms, resembling the original waveform quite closely. The code, models can be found at: https://github.com/Anshuman04/waveformReconstructor.
BibTeX
@inproceedings{gaharwar2023xi,
  author = {Anshuman Gaharwar and Parth Parag Kulkarni and Joshua T. Dickey and Mubarak Shah},
  title = {{Xi-Net: Transformer Based Seismic Waveform Reconstructor}},
  booktitle = {ICIP},
  year = {2023},
  url = {https://openreview.net/forum?id=BBAd4kIbEP},
}
Figure