PA-HTGNN: A physics-aware heterogeneous temporal graph neural network for PDC drill bit wear anomaly detection incorporating bit structural parameters

Accurate monitoring of PDC drill bit wear is critical for improving drilling efficiency and reducing non-productive time. However, existing methods still face three major challenges: the lack of continuous time-step-level wear labels during drilling, the absence of explicit modeling of the modulation effect of bit structural parameters on wear evolution, and the limited integration of physical efficiency indicators with data-driven models. To address these issues, this paper proposes a physics-aware heterogeneous temporal graph neural network (PA-HTGNN). First, a weakly supervised labeling strategy is developed based on clustering of rock-breaking efficiency indicators, including mechanical specific energy (MSE), depth of cut (DOC), drilling impedance (DRIMP), and rate of penetration (ROP), to convert run-level international association of drilling contractors (IADC) dull grades into continuous time-step-level labels. Second, drilling operational parameters, six bit structural features, including bit diameter, blade number, cutter diameter, cutter density, total nozzle area, and back rake angle, and global state information are jointly modeled as a physics-aware heterogeneous graph, in which MSE, DOC, and DRIMP are encoded as differentiated edge attributes through a static mask matrix. Third, a two-layer edge-conditioned graph convolution module is employed to extract spatial coupling features by adaptively modulating message passing with physical edge attributes, and a two-layer gated recurrent unit network with an additive temporal attention mechanism is further used to capture the sequential evolution of wear states. The cleaned field dataset comprises 363 usable drilling runs. The primary evaluation uses 70 runs from 15 wells under a leave-one-run-out cross-validation (LORO-CV) protocol, while an additional 36 runs with held-out bit types are reserved for cross-type generalization testing. Under the leave-one-run-out cross-validation protocol, PA-HTGNN achieves an accuracy of 96.14% and an F1-score of 0.9485, outperforming AutoML, Bi-LSTM, ANN, and random forest across all four evaluation metrics. Ablation experiments show that structural parameter modeling provides the largest contribution, with the accuracy decreasing by 4.18 percentage points after its removal. Cross-type transfer tests further show that incorporating structural parameters increases abnormal-class recall from 73.3% to 86.7%, reducing missed detections for unseen bit types. Node-importance analysis across different lithologies reveals patterns broadly consistent with the rock-breaking mechanisms of mudstone and sandstone formations, supporting the physical interpretability of the model. The proposed method provides a physically interpretable and generalizable framework for online drill bit wear anomaly monitoring, with potential applicability to field drilling optimization and tripping decision support.