A state-conditioned recognition framework for TBM cutter using mechanism-guided historical statistics and deep observations

This study tackles the reliable recognition of the tunnel boring machine (TBM) cutter damage state under complex and varying operating conditions, where vibration responses are broadband, strongly nonstationary, and dominated by randomly excited impacts, making stable fault characteristic frequencies difficult to extract. Using a scaled multicutter rock-breaking bench platform, we show that partial wear induces a fundamental transition in the cutter-rock contact from rolling indentation to sliding scraping, thereby reshaping the energy distribution and impact-related statistics. Guided by this mechanism, we design physically interpretable state proxies to represent history-accumulated statistical signatures of damage. We further propose a mechanism-guided state-conditioned recognition framework that decouples instantaneous observations from historical-state information: A deep observation encoder (deep convolutional neural network (DCNN) or Transformer) learns window-level dynamic-response features from raw multichannel vibration segments, while a mechanism-guided state estimator computes state proxies from historical statistics and uses them as a state-conditioned constraint for decision making. To align offline learning with online monitoring, the estimator employs global statistics during training to establish stable reference baselines but switches to historical statistics during testing to avoid future information leakage. Cross-condition evaluations demonstrate improved robustness to distribution shifts, achieving peak accuracies of 92.17% (DCNN) and 88.91% (transformer), with time-evolution results indicating progressive stabilization as historical statistics accumulate. The ablation results verify the complementary contributions of different proxies and show a marked drop when state variables are removed (down to 64.73%), confirming that performance gains stem from state-conditioned constraints rather than increased model complexity. Overall, the proposed framework provides an interpretable and practically oriented route toward more stable cutter-state monitoring under varying operating conditions.