An efficient drilling conditions classification method utilizing encoder and improved Graph Attention Networks

Abstract

Identification of drilling conditions is paramount for ensuring safety and efficiency in drilling operations. Traditional methods, often based on manual analysis of drilling parameters and patterns by experts using empirical formulas, typically lack speed and precision in modern drilling procedures. To address this shortcoming, we have developed a sophisticated classification methodology that combines Encoder layers with improved Graph Attention Networks (GAT), aiming to precisely discern seven prevalent drilling conditions. Using an advanced expert empirical decision tree, we achieved superior accuracy in our dataset annotations. The model’s architecture integrates the K-nearest neighbors (KNN) algorithm for seamless integration between the Encoder layer and GAT networks. Sensitivity analyses pinpoint optimal feature parameters, leading to significant enhancements in model accuracy. Experimental outcomes show our annotation approach surpasses traditional techniques by 15.24%, achieving an accuracy of 96.73%. Additionally, our model boasts a recall rate of 94.22%, indicating high reliability in correctly identifying true positive drilling conditions, which is crucial for operational safety and decision-making in drilling scenarios. Characterized by its real-time capabilities, precision, and scalability, our model demonstrates promising potential for applications in practical drilling scenarios.