Tool wear prediction through a physics-assisted method using few-sensor monitoring data

Tool wear prediction plays an increasingly crucial role in intelligent manufacturing, significantly enhancing machining quality and reducing costs. Data-driven methods, bolstered by artificial intelligence and sensor technologies, have achieved remarkable advancements in tool wear prediction. Most of data-driven methods utilize multi-sensor data for a more precise prediction. However, its application is constrained by expensive monitoring costs and confined monitoring site. Additionally, the prediction results along the time series fluctuate because data-driven methods only establish mapping relations between sensor signals and tool wear value, neglecting the law that tool wear increases monotonously. To address these issues, a physics-assisted hybrid tool wear prediction method is proposed in this paper to realize the accurate tool wear prediction based on few-sensor monitoring. Initially, cutting force and vibration signals are collected as training data under multi-sensor monitoring condition and sensitive features related to the tool wear are extracted. Subsequently, two distinct models are developed: a tool wear prediction model based on multi-sensor and a force feature prediction model based on vibration features. When the machining process under the same working conditions is monitored by few-sensor, the generated force features derived from the force feature prediction model are integrated with vibration features derived from the monitored vibration signals to realize the feature dimension augmentation, subsequently serving as inputs to the tool wear prediction model based on multi-sensor. Furthermore, a physical model describing the tool wear evolution is formulated by fitting the wear curve. On this basis, the final tool wear prediction result is obtained by fusing the results of data-driven model and physical model using the particle filter algorithm. Experimental results demonstrate that the feature dimension augmentation enriches the monitoring information and improve the prediction accuracy. Moreover, the developed hybrid model outperforms individual tool wear prediction models greatly.