Downhole Microseismic Detection Using Fiber-Optic Distributed Acoustic Sensing Based on Segmentation Model and Connected Domain Algorithm

Abstract

The widespread adoption of fiber-optic distributed acoustic sensing (DAS) technology in oil and gas production, the timely and precise identification of microseismic events within DAS datasets holds importance for enhancing both the efficacy and safety of mining operations. The current DAS microseismic detection methods, including template-matching technology and convolutional neural network (CNN)-based approaches, predominantly face challenges such as high computational complexity, slow detection speed, and low detection accuracy. In response, we introduce the SegDetection deep learning model, a semantic segmentation model that integrates dynamic snake convolution with MobileNetV3 to enhance feature extraction capabilities. The model employs lite reduced atrous spatial pyramid pooling (LRASPP) as its segmentation head network. Subsequently, a two-stage connected domain algorithm is utilized to produce prediction boxes and confidence scores. To enhance the segmentation accuracy of our model, we implement a segmentation correction strategy. In the microseismic detection task using the downhole DAS microseismic dataset in Utah, USA, our proposed method achieved an F1-score of 0.902. After applying the error segmentation correction strategy, the F1-score improved to 0.951. The experimental results indicate that the method proposed in this article exhibits commendable performance in downhole DAS microseismic detection. In addition, the error segmentation and correction strategy introduced significantly enhances the model’s detection accuracy, suggesting its broad applicability to various downhole DAS microseismic detection tasks.