The domestic materials and components industry has achieved rapid scale expansion, along with growth rates and productivity improvements that exceed the manufacturing average. In contrast, manufacturing sites have been experiencing a steady decline in skilled workers, an increasingly aging workforce, and continuously rising requirements for stronger worker safety standards. These environmental changes make the continuation of traditional shop-floor operations largely dependent on tacit knowledge and intuition unsustainable, thereby accelerating efforts to systematize and automate experience dependent factors and to shift decision making toward data-driven practices.

Tool wear prediction in machine tools has emerged as an essential task. Tool wear leads to increased cutting forces, deterioration of surface roughness, dimensional deviations, chatter, and tool breakage, thereby directly affecting product quality, productivity, and equipment utilization. Determining tool replacement timing based on the visual and auditory judgment of experienced operators lacks reproducibility and objectivity, and it is difficult to guarantee optimal tool-life management under diversified and high-speed machining conditions. Accordingly, a framework is required that continuously estimates tool wear through real-time measurement and analysis of process signals such as cutting load, vibration, acoustics, and power, and links the estimation to timely tool replacement. Consequently, tool condition monitoring (TCM), which aims to detect tool wear and breakage at an early stage through real-time acquisition and analysis of process signals such as cutting force, vibration, acoustic emission (AE), and spindle power, has become a key topic in manufacturing sites and research laboratories [1]. In milling operations, the tooth-passing frequency and its sidebands, time-frequency analysis, and machine-learning approaches are widely used as standard tools [2]. Recent studies have further applied multisensorial data, intelligent tools, multisensor fusion, remaining-useful-life prediction, and deep-learning forecasting to tool-wear monitoring [3–12]. In the TCM field, both direct and indirect cutting-force measurement methods have been explored; among them, the dynamometer is widely adopted as a laboratory standard [13–16], because it can accurately measure three-axis cutting forces, yet its high acquisition and operating costs, installation burden, and maintenance complexity limit its continuous deployment in industrial environments [13,14]. To mitigate these constraints, various sensors have been proposed as alternatives. Accelerometers and AE sensors can leverage broad bandwidths and thus show advantages in detecting instability and chatter [17–19]; however, in high-speed machining, limitations related to sensor mounting and wiring, sampling constraints, and variations in response depending on the process and machine make it difficult to stably evaluate long-term wear trends. In practice, sensor durability and long-term reliability under coolant-spraying environments are critical concerns. For rotating measurements, holder-integrated dynamometers have been developed to measure dynamic loads in milling together with spindle rotation [13,20], thereby improving space utilization and reducing wiring constraints; nevertheless, barriers to adoption have been noted, including added structural mass and inertia, calibration requirements and thermal stability issues, and high design and manufacturing costs [13]. As an indirect approach, torque, current, and power-based estimation has also been actively studied [21–23]. Although spindle current and power signals enable low-cost and convenient measurement, they are sensitive to the drive train, controller, and machine-specific constants, requiring machine-dependent calibration, and it has been repeatedly reported that separating high-frequency components and short-period instabilities remains relatively challenging. These limitations highlight the need for a shop-floor-friendly platform that balances ease of installation, sensitivity, and reliability for tool-wear prediction. Previous studies have explored indirect force sensing using piezoelectric elements embedded in dedicated fixtures, clamps, or rotating holders [13,24,25]. Those approaches are valuable, but they typically require custom fixture design, sensor integration during manufacturing, or dedicated calibration of the modified structure. In contrast, the Piezobolt approach considered here uses a load-sensing fastener that can replace a conventional clamping bolt with the same interface. The sensing concept is therefore based on measuring load redistribution through the existing clamping architecture rather than through a specially fabricated instrumented fixture. The practical novelty of the present study lies in experimentally verifying that this minimally invasive installation concept can provide sufficiently informative signals for tool-wear prediction when benchmarked against a laboratory-standard dynamometer.

In this study, a bolt-type piezo sensor was employed to investigate the feasibility of tool-wear prediction. Because the sensor can be installed by replacing a conventional clamping bolt with a load-sensing fastener of the same specification, the proposed approach enables deployment without machine modification, redesign, or additional fixtures. Simultaneous measurements using both a Piezobolt and a dynamometer were conducted to quantitatively compare the wear-related information obtained from the two sensing modalities and to evaluate the prediction performance of flank-wear land width (VB) using support vector regression (SVR).

In this study, support vector regression (SVR) was applied to predict the continuous evolution of tool wear. SVR is a regression technique based on the support vector machine (SVM) framework and introduces an ɛ-insensitive loss function, in which prediction errors within a predefined tolerance ɛ are not penalized, while only errors exceeding this margin contribute to the loss. This property is particularly advantageous for machining applications, as it suppresses the influence of sensor noise and short-term signal fluctuations that are not directly associated with progressive tool wear.

The SVR prediction function is expressed as a linear combination of support vectors and a kernel function, as given in Eq. (4).

Here, SV denotes the set of samples that lie on the margin boundary during training and thus define the regression function. K(x_i, x) is a kernel function representing the similarity between two input samples, and the coefficients c_i and the bias term b are determined during the training process. The ɛ-insensitive margin ɛ does not explicitly appear in Eq. (4); instead, it is incorporated into the loss function during training and indirectly influences the regression function through the selection of support vectors and the distribution of the coefficients c_i.

To flexibly approximate the nonlinear relationship between process signals and tool wear, a radial basis function (RBF) kernel was adopted in this study, as defined in Eq. (5).

The SVR model employed an RBF kernel, and the hyperparameter-tuning procedure was conducted using only the training segment so that information from future wear states was not used during model setup. To preserve the temporal structure of the data, time-series cross-validation based on TimeSeriesSplit was adopted within the training segment. GridSearchCV was then used to explore C over 0.1–316.2 on a logarithmic scale, γ over {scale, auto,10⁻⁴ – 10⁻¹}, and ɛ over {0.003, 0.005, 0.01, 0.02, 0.03}, with the negative root mean squared error used as the selection criterion. In the implemented prediction pipeline, the baseline SVR configuration for final model fitting was C = 10.0, γ = “scale”, and ɛ = 0.01.

In addition, the target variable was transformed using log1p, and the input features were preprocessed through quantile normalization and standard scaling. Because SVR is sensitive to the scale of the input variables, these preprocessing parameters were estimated exclusively from the early machining segment used for training, and the same parameters were subsequently applied to the later prediction segment. This procedure prevents information leakage from future data and reflects a realistic tool condition monitoring scenario.

To avoid ambiguity, the exact candidate variables used in the SVR analysis are stated explicitly. For the dynamometer-based model, the candidate feature set consisted of six resultant-force variables: FxFy mean, FxFy RMS, FxFy P2P, FxFyFz mean, FxFyFz RMS, and FxFyFz P2P. For the Piezobolt-based model, the candidate feature set consisted of 35 time-domain variables obtained from the four individual bolt signals (PB1–PB4) and the derived load-combination variables FxFyFz, Mx, and My, each represented by five statistics: RMS, mean, P2P, min, and max.

The larger candidate set for the Piezobolt model reflects the sensing principle of the proposed approach. Unlike the dynamometer, which directly provides force-related resultant signals, the Piezobolts capture load redistribution at four clamping points. Therefore, wear-related information is distributed across the individual bolt responses and their combined patterns. For this reason, both individual bolt variables and mechanically meaningful combination variables were included so that the model could preserve the spatial and mechanical information contained in the four-bolt response before statistical feature selection was applied.

Before model fitting, highly correlated variables were removed from each sensor-specific candidate set using the training segment only. Feature importance was then evaluated only within the training segment using a permutation-based method, in which each feature was randomly shuffled and the resulting change in prediction performance was used to quantify its importance. Fig. 12 presents the top-ranked sensor-derived feature for each sensor as a compact summary, whereas the full ranking of the sensor-derived candidate features is provided in Table 1.

To examine the effect of feature dimensionality on prediction performance, additional comparisons were performed using full, top 1, and top 3 sensor-derived feature settings. In the full setting, the retained candidate features were used in an SVM pipeline consisting of quantile normalization, standard scaling, SelectKBest, and an RBF-kernel SVR model. In the reduced settings, the highest-ranked one or three sensor-derived features from the permutation-based ranking were used as model inputs. Among these settings, the top-1 setting yielded the most stable future-prediction performance and was therefore used for the prediction results shown in Figs. 13–16. Specifically, only the early machining segment was used for training, and the later segment—where tool wear accelerates—was reserved exclusively for prediction. Among the evaluated chronological splits, the 50% training / 50% prediction and 80% training / 20% prediction scenarios are reported here. For example, under the 50% training condition, data from Passes 1–32 were used for training, while data from Passes 33–64 were excluded from training and used solely for prediction. The SVR-based prediction results for the 50% training / 50% prediction scenario are presented in Figs. 13 and 14 for the dynamometer and the Piezobolts, respectively. In the case of the dynamometer, the predicted wear exhibited an approximately linear increasing trend in the prediction interval, resulting in noticeable deviation from the measured tool-wear values. In contrast, the results obtained using the Piezobolts showed a wear progression trend consistent with the measured tool wear, with particularly close agreement in the region where the flank-wear land width VB exceeded 0.3 mm, corresponding to the end-of-life stage of the cutting tool. As shown in Table 1, the dynamometer ranking is dominated by two resultant-force variables, whereas the Piezobolt ranking is distributed across Fsum-based, moment-related, and individual-bolt variables. This difference reflects the sensing principles of the two sensor systems: the dynamometer directly provides compact force-resultant information, whereas the Piezobolts represent wear-related information through distributed clamping-point load responses and their combinations. For the 80% training / 20% prediction setting, the feature values from both sensors exhibited trends like the measured tool wear. Even in the later stage where tool wear accelerates, the change in slope was captured in a comparable manner. In the case of the Piezobolts, setting a critical wear threshold at a flank-wear land width (VB) of 0.3 mm enabled the identification of the wear limit, suggesting that the Piezobolt-based model may be useful for threshold-based tool-wear monitoring under the tested condition. However, additional validation is required before extending this interpretation to real-time industrial deployment.

Quantitative evaluation of the SVR-based tool wear prediction performance was conducted. The predicted flank-wear land width (VB) was compared with the measured values, and the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) were calculated. The quantitative results are summarized in Table 2. Under the 50% training / 50% prediction condition, the Piezobolt-based SVR model achieved an RMSE of 0.041 mm, an MAE of 0.03 mm, and an R² value of 0.80, demonstrating superior prediction performance compared to the dynamometer-based model, which yielded an RMSE of 0.057 mm and an R² value of 0.52. These results indicate that the Piezobolt signals can effectively capture the overall trend of tool wear, even when a limited amount of training data is available.

In contrast, under the 80% training / 20% prediction condition, the dynamometer-based model exhibited higher prediction accuracy, with an RMSE of 0.033 mm and an R² value of 0.80. This result suggests that, when sufficient training data are provided, the reference sensor (dynamometer) can more stably reproduce the tool wear progression. Nevertheless, the Piezobolt-based model also demonstrated a relatively high prediction performance, with an R² value of 0.71, indicating a certain level of robustness with respect to changes in the proportion of training data.

Overall, these quantitative results provide strong evidence that Piezobolts can serve as an effective alternative sensor for estimating tool wear trends, particularly in environments with limited training data. Furthermore, the robustness of the Piezobolt-based model across different training conditions highlights its suitability for practical tool condition monitoring systems in industrial applications.

This study investigated the feasibility of predicting cutting-tool wear using Piezobolts installed at the workpiece clamping points. As an experimental approach, conventional clamping bolts were replaced with Piezobolts of the same specification, allowing sensor installation without additional structural modification of the machine-tool system. After tightening the Piezobolts to the manufacturer-recommended torque, signals were acquired simultaneously from the Piezobolts and a three-axis dynamometer used as a laboratory reference.

For comparison with the dynamometer measurements, the Piezobolt signals were processed by approximating three-axis cutting-force components from the axial loads measured at the four clamping points. The comparison results showed similar wear-related trends over the entire machining sequence, indicating that the Piezobolts can serve as sensitive surrogate sensors for tool-wear monitoring even though they do not directly measure the cutting forces.

Tool-wear prediction was then performed using support vector regression (SVR) while preserving the temporal order of the measured data. The results showed that the later-stage wear trend could be predicted using only the early-stage data, and that the features derived from the Piezobolt signals provided prediction performance comparable to that obtained from the dynamometer-based features. These findings demonstrate the feasibility of Piezobolt-based sensing as a practically installable approach for data-driven tool-wear prediction under the tested condition.

The present study was limited to one cutting tool, one workpiece material (S45C), one machining operation (slot milling), and one set of cutting parameters. Therefore, the results should be interpreted as an experimental feasibility study rather than as a generalized validation of shop-floor deployment or predictive maintenance in diverse manufacturing environments.

Future work will expand the dataset to multiple tools, workpiece materials, and cutting conditions in order to improve model generalizability and strengthen the practical validity of the proposed method. Additional studies are also needed to examine model transferability, recalibration requirements, threshold robustness for real-time monitoring, and the influence of long-term preload stability, fixture variation, and process disturbances on the Piezobolt response. Within the tested scope, however, the proposed method demonstrated that Piezobolts can provide sufficiently informative signals for tool-wear prediction and can serve as a promising foundation for further development of practical tool condition monitoring systems.