Volume 66

Research on Kmeans Cluster Analysis and LSTM for Identifying Sparse Drilling Conditions Shuai Ding, Lin Xiao,Qi Wei,Chaodong Tan ,Kun Zhao

Abstract

Utilizing logging dynamic monitoring data for identifying drilling operation conditions is crucial for precise analysis of drilling efficiency and operational enhancement. However, the time series distribution of samples across various drilling conditions exhibits notable non-uniformity, leading to a substantial imbalance in the category labels of drilling conditions. This imbalance issue significantly reduces the accuracy of recognition models due to a bias towards the majority class.In this study, drilling history data from 98 wells were examined to create 14 distinct sets of drilling condition samples. To address the sample imbalance problem, the weights of the samples were determined using the K-means clustering algorithm based on cluster proportions and sample distribution specificity within each cluster. These weight factors were integrated into the loss function to develop a drilling condition recognition model known as K-means Long Short-Term Memory (LSTM), enabling precise recognition of the 14 drilling conditions.Experimental findings demonstrate that the K-means-LSTM model outperforms the LSTM model, with an increase in recognition accuracy of 3.561%, recall by 4.554%, and F1-score by 3.667%. Particularly noteworthy is the substantial improvement in recognition accuracy for break out a single joint and make up a single joint sparse conditions, with increases of 39% and 42%, respectively. Compared to conventional deep learning approaches, the proposed method exhibits remarkable accuracy enhancements and suitability for recognizing sparse drilling conditions, thereby offering robust technical support for fine-tuning drilling operations and supervision.

Keywords drilling operation, condition identification, class imbalance,Kmeans cluster analysis,LSTM