Drill bit failure detection for drilling processes based on global–local feature extraction and multi-stage incremental learning

In drilling processes, real-time detection of drill bit failure states is essential to mitigate operational risks, reduce downtime, and enhance drilling precision. However, drilling signals often exhibit both long-term degradation and local subtle changes. This coexistence poses great challenges to the accurate detection of drill bit failures. Moreover, models trained on historical data often exhibit significant performance degradation when deployed to new drilling depths. This is because the distributions of drilling process data diverge at these new depths due to lithological heterogeneity. To overcome such limitations, this paper proposes a new drill bit failure detection method for drilling processes by integrating Transformer-Convolutional Selective Fusion Network (TCSFN) with multi-stage incremental learning. The main contributions are twofold: 1) A feature extraction method based on TCSFN is proposed to capture global long-term trend features and local transient fluctuation features; 2) a multi-stage incremental learning strategy is designed for different stages of the drilling processes, and composite losses are devised for these stages separately. Case studies involving real-world data are utilized to demonstrate the effectiveness and superiority of the proposed method.