DAS signal modeling using the generative adversarial neural network technique

Andrey V. Timofeev

2025

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.17586/2226-1494-2025-25-1-87-94
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

A new method of generating model sets of Distributed Acoustic Sensing (DAS) signals of different classes is proposed. Statistical characteristics of model signals are quite similar to real DAS-signals of corresponding classes and can be used for sharp improvement of DAS-signals processing quality by machine learning methods. The proposed method is a modification of the Generative Adversarial Network (GAN) technique. The novelty of the approach lies in the introduction of an additional external control loop for the performance of the generative network which includes a classifier trained on an available (small) corpus of real DAS signals. A method for generating model sets of DAS signals based on GAN technology is proposed, and it differs from the classical technology by the presence of an additional external quality control loop. An optimality criterion for the generating system is formulated, the optimum of which is achieved by step-by-step reconfiguration of the GAN neural network structure. Reconfiguration is based on the Nelder- Mead optimization method. A software implementation of the proposed solution architecture on the Python platform is developed and tested on real data. Results are presented proving the practical efficiency of the proposed method. In particular, the proposed method allowed to increase the capacity of the training dataset and, thus, to increase the resulting reliability of the classification of target DAS signals. The developed approach is promising for use in cases where the capacity of the datasets provided for training is insufficient to ensure highly reliable classification.