CNN标签数据迭代优化及在小断层识别中的应用

Iterative optimization of label data in CNN algorithm and its application to small fault identification

标签数据的代表性及普适性是影响CNN算法精度及泛化程度的重要因素之一,目前基于CNN算法的断层识别方法大多使用理论标签数据.本文提出一种理论标签数据和实际标签数据迭代优化方法,最终得到了基于实测地震数据具有良好代表性又融入了基于理论数据具有较强普适性的标签数据,在这种标签数据训练下的CNN模型具有更好的泛化能力.针对渤海深埋潜山的低信噪比资料,应用该方法识别潜山内幕断缝系统取得较好效果,证明了方法的有效性和适用性,为潜山内幕小断层及断缝系统识别提供了一种高效可靠的方法技术.

The representativeness and universality of the tag data is one of the important factors affecting the accuracy and generalization degree of Convolution Neural Network(CNN)algorithm. At present, most of the fault identification methods based on CNN algorithm use theoretical tag data. In this paper, an iterative optimization method of theoretical tag data and actual tag data is proposed. Finally, a tag data with good representativeness based on measured seismic data and strong universality based on theoretical data are integrated into the CNN model trained by such tag data. The CNN model has better generalization ability. Aiming at the data of deep buried hill in Bohai Sea with low SNR, the application of this method to identify the internal fracture system of buried hill has achieved good results, which proves the effectiveness and applicability of the method, and provides an efficient and reliable method and technology for the identification of small faults and fracture system inside the buried hill.