A model of feature extraction for well logging data based on graph regularized non-negative matrix factorization with optimal estimation

Reservoir oil-bearing recognition is the process of predicting reservoir types based on well logging data, which determines the accuracy of recognition. However, the original well logging data is multidimensional and contains potential noise, which can influence the performance of sequent processing, such as clustering and classification. It is crucial to obtain key low-dimensional features and study an accurate automatic recognition algorithm under unsupervised condition. To solve this problem, we propose a feature extraction method named graph regularized non-negative matrix factorization with optimal estimation (GNMF-OE) according to the characteristics of well logging data in this paper. Firstly, the low dimensional embedding dimension of high-dimensional well logging data is modeled and estimated, which enables the method to obtain the appropriate number of features that reflect the data structure. Secondly, local features are optimized by structured initial vectors in the framework of GNMF, which encourages the basis matrix to have clear reservoir category characteristics. These two approaches are meaningful and beneficial to construct an appropriate basis matrix that discovers the intrinsic structure of well logging data. The visualized experimental results on real datasets from Jianghan oilfield in China show that the proposed method has significant clustering performance for reservoir oil-bearing recognition.