In modern manufacturing, milling and micromilling processes play a central role in precision production. However, rapid wear of cutting tools often leads to sudden tool breakage, unplanned downtime, and part rejection. Maintenance is therefore essential to ensure efficiency, safety, and cost-effectiveness across industries. Traditional maintenance strategies have gradually evolved into predictive maintenance approaches supported by advanced technologies, creating a strong industrial demand for accurate and reliable predictive solutions. This review systematically analyzes studies from the past decade on predictive maintenance in milling and micromilling processes, with a particular focus on the performance of machine learning and deep learning algorithms integrated with computer vision techniques. The study evaluates model performance, prediction accuracy, and industrial applicability to identify key strengths and existing research gaps. The findings indicate that multi-sensor data fusion, deep learning methods, and hybrid models achieve the highest performance in tool wear monitoring and remaining useful life prediction of cutting tools.

Forecasting the Remaining Useful Life (RUL) of cutting tools plays a key role in intelligent predictive maintenance and downtime reduction in today’s manufacturing. Traditional approaches based on hand-designed features or conventional machine learning models are typically incapable of handling intricate temporal relationships and sensor–interaction. This paper proposes a graph-based deep learning pipeline for tool wear forecasting in CNC milling, employing vibration, spindle load, and cutting force signals. Features from both the time-domain (mean, standard deviation, skewness, and kurtosis) and the frequency domain (FFT coefficients, power spectral density) were extracted, and graphs were formed based on correlation and temporal edges in order to model feature dependencies and sequential relationships. Three graph neural network (GNN) architectures were utilized: graph convolutional network (GCN), graph recurrent network (GRN), and graph attention network (GAT). Models were trained and tested on 120 machining runs, and hyperparameter optimization was performed using Optuna and Bayesian Optimization. The experimental results show that Optuna tuning enabled GRN to achieve the strongest time-domain performance (RMSE = 3.68, \(\text {R}^2\) = 0.9963), while Bayesian Optimization yielded improved frequency-domain performance for GAT (RMSE = 13.50, \(\text {R}^2\) = 0.9487). The designed framework illustrates the power of GNNs for tool-condition monitoring and offers a scalable solution for predictive maintenance, with future work considering physics-informed graph modeling and transfer learning for real-time application.