In this study, we proposed a methodology for predicting tool wear in the turning process using the SVR model. This model maintains stable performance even in small-scale data environments and demonstrates robust characteristics against outliers. We detected changes in machining performance caused by tool wear through an AE sensor and accelerometer. Features were extracted from the acquired sensor signals and utilized in the machine learning model. Prior to training, the extracted features underwent a preliminary screening process based on distance correlation. By optimizing the feature combination using the RFECV algorithm, we achieved a prediction accuracy of R² = 0.95. The analysis revealed that key features influencing the tool wear prediction model included several significant variables. Additionally, we found that evaluating feature importance allowed for more efficient model improvement. Overall, when developing a tool wear prediction model for cutting, it is crucial to utilize various sensor signals, extract features in both the time and frequency domains, and optimize the combination of those features.

We present a 3-PRS compliant parallel manipulator actuated by PZTs. The motion ranges are 400-㎛ translation to the z-direction and 5.7-mrad rotation about any axis on the x-y plane. A mechanical amplifier whose advantage is approximately 5.5 is designed and integrated with flexure revolute and spherical joints in a monolithic way. We evaluated the performance of the system: kinematic and dynamic characteristics. In kinematic point of view, the flexures play the roles of conventional mechanism but any nonlinearity such as dead-zone and backlash, which make it possible to achieve the steady-state resolution less than 0.1 ㎛. The system shows resonance near 86 ㎐ with high magnitude and, therefore, poor transient response. But the settling is faster when the flexures are strained higher.