Explainable Bayesian-optimized LSTM for predicting multivariate tool wear in tribological systems

Accurate prediction of tool wear is required for condition monitoring and process optimization in tribological systems as their operating conditions and material treatments vary. This study presents an explainable, Bayesian-optimized Long Short-Term Memory (LSTM) framework for multivariate prediction of wear rates in untreated and cryogenically treated WC–Co cutting tools. Experimental wear data were obtained from ASTM G65-compliant pin-on-disc tests. Tests are conducted under multiple load and sliding speed conditions. The model uses applied load, sliding speed, and sliding time as sequential inputs to simultaneously predict six volumetric wear rate outputs. Bayesian hyperparameter optimization using Optuna was executed over 20 trials. This yields an optimal three-layer LSTM architecture with a minimum validation mean squared error of 4.62 × 10−5. To enhance interpretability, Shapley additive explanations were employed, revealing sliding time as the dominant factor governing wear progression across all tool states. The proposed framework offers an effective data-driven approach for multivariate tool wear prediction with the integration of explainable deep learning in tribological monitoring applications.