Research Progress on Tool Remaining Useful Life Prediction (RUL) Based on Multi-sensor Signal Fusion

Abstract

Tool Remaining Useful Life (RUL) prediction is essential for maintaining machining quality, reducing production costs, and preventing unexpected tool failures. Because single-sensor monitoring captures only partial and noise-sensitive degradation information, this review examines multi-sensor information fusion for tool RUL prediction. Force, vibration, acoustic emission, and current sensing are compared with respect to sensitivity, installation requirements, cost, and industrial applicability. Data-, feature-, and decision-level fusion strategies are evaluated in terms of information retention, noise tolerance, computational burden, and suitable operating conditions, while time-, frequency-, and time–frequency-domain methods are reviewed as the basis for constructing degradation-sensitive representations. Physics-based, data-driven, and hybrid prognostic approaches are compared in terms of interpretability, data requirements, adaptability, uncertainty representation, and online capability. Widely used public benchmark datasets are further assessed for their coverage of tool types, cutting conditions, and wear modes, revealing limited evidence of generalizability to industrial settings. To integrate these topics, a Wear–Information–Decision Chain (WIDC) framework is proposed to link operating context and degradation mechanisms with sensing observability, information representation and fusion, prognostic inference, and maintenance-oriented deployment. The framework demonstrates that reliable RUL prediction depends on consistency across the entire chain rather than on the final predictor alone. Future research should emphasize open full-life datasets, few-shot and physics-informed learning, interpretable fusion, digital twins, standardized cross-condition validation, and computationally efficient online deployment.