Machine Learning for Tool Condition Monitoring and Predictive Maintenance in CNC Machining: A Systematic Review

Abstract

Computer Numerical Control (CNC) machining is the backbone of precision manufacturing, but the productivity, part quality and machine uptime of CNC machining are dependent on the condition of the cutting tool. Conventional fixed-interval tool replacement either discards valuable tool life or permits the tool to be used past safe bounds of dimension, and progressive flank wear, chipping and breakage also affect surface finish and dimensional accuracy. This systematic review follows the reporting guidelines outlined in the PRISMA 2020 statement, and summarizes the existing research literature between January 2016 and September 2026 addressing the application of machine learning (ML) for CNC tool condition monitoring (TCM) and predictive maintenance (PdM). Studies were classified along four dimensions: machining operation, sensing modality, learning method and maintenance objective. It has been revealed that cutting force, vibration, acoustic emission and spindle current/power are the most important indirect measurement signals; that feature quality has significant influence on the performance of traditional classifiers; and that the deep sequence model, including convolutional and long short-term memory models, is now the most popular model for wear estimation and remaining useful life prediction. The reported accuracy is generally in the range of 85% to 99%, but is rarely comparable because the datasets, tools, materials, and validation procedures vary. The research gaps identified were: lack of or costly labelled data, poor cross-machine generalisation, real-time deployment latency, class imbalance and data quality, limited interpretability, under-used ensemble and hybrid methods, under-representation of small firms and developing economies, and incompleteness in integrating with digital twins and CAD/CAM data. The review suggests a research agenda, in which transferable and explainable models are coupled with design-of-experiments process knowledge, low-cost edge sensing and human-centred implementation, particularly relevant for manufacturers operating in resource constrained industrial environments