Abstract
Transient Electromagnetics (TEM) is a workhorse technology offering robust insights into hydrogeological structures in the subsurface. Modern instruments yield enormous amounts of data. To lower the complication of the method and getting the technology even more widespread, machine learning approaches offer great opportunities for complete automation of the data processing task. A major time consuming task is to sort good data from data contaminated by interference to buried power lines, fences, etc. (called as couplings). This process is mostly an expensive, highly sophisticated and a manual task. The state-of-the-art machine learning approaches uses supervised methods trained on a subset of acquired data within a localized region, but this is not generalizable to data obtained at other sites making it an in-efficient solution.
As couplings in the data are generally rare events, we consider it as an anomaly detection problem. Therefore, we deploy an unsupervised learning approach where a deep convolutional autoencoder is used to differentiate between the coupled and uncoupled data. The autoencoder is trained to learn a representation (encoding) of the input data for dimensionality reduction. A reconstruction part is also trained that expands the reduced dimension to result the output as close as possible to the original input. Since the autoencoder aims to replicate the input at its output, we rely on a database of synthetic TEM data that corresponds to a wide range of subsurface models. By learning the original representation from the most salient features encoded in the reduced dimension, the autoencoder learns to reproduce the most frequently observed characteristics. When facing couplings in the data, the reconstruction performance degrades and the reconstruction error between the original data and its reduced dimensional reconstruction becomes higher, which is used as an anomaly score to detect couplings.
We test our method on data acquired at several sites by a towed TEM (tTEM) system and show that the proposed approach exhibits high quality data processing within a fraction of a second, which is ready for inversion. The method is generalizable to practically any tTEM data set acquired worldwide. Importantly, the complicated data processing is been eliminated and no highly skilled operator is required.
As couplings in the data are generally rare events, we consider it as an anomaly detection problem. Therefore, we deploy an unsupervised learning approach where a deep convolutional autoencoder is used to differentiate between the coupled and uncoupled data. The autoencoder is trained to learn a representation (encoding) of the input data for dimensionality reduction. A reconstruction part is also trained that expands the reduced dimension to result the output as close as possible to the original input. Since the autoencoder aims to replicate the input at its output, we rely on a database of synthetic TEM data that corresponds to a wide range of subsurface models. By learning the original representation from the most salient features encoded in the reduced dimension, the autoencoder learns to reproduce the most frequently observed characteristics. When facing couplings in the data, the reconstruction performance degrades and the reconstruction error between the original data and its reduced dimensional reconstruction becomes higher, which is used as an anomaly score to detect couplings.
We test our method on data acquired at several sites by a towed TEM (tTEM) system and show that the proposed approach exhibits high quality data processing within a fraction of a second, which is ready for inversion. The method is generalizable to practically any tTEM data set acquired worldwide. Importantly, the complicated data processing is been eliminated and no highly skilled operator is required.
| Original language | English |
|---|---|
| Publication date | 2021 |
| Publication status | Published - 2021 |
| Event | AGU 2021 - New Orleans, United States Duration: 13 Dec 2021 → … |
Conference
| Conference | AGU 2021 |
|---|---|
| Location | New Orleans |
| Country/Territory | United States |
| Period | 13/12/2021 → … |