Developing a Predictive Wear Model for Intelligent Tool Change Systems

With the development of smart manufacturing technologies, managing materials, workforce, and equipment in machining processes has become crucial to ensure the reliability of key components. Tool wear, as the weakest and most damage-prone element in the OUPN system, affects both product quality and machining efficiency, and is unavoidable due to thermodynamic interactions during cutting [1]. Standards such as ISO 3685 [2] provide guidelines for tool performance at constant cutting speeds, while Taylor’s equation is essential for predicting tool life and production costs, particularly for hard-to-machine materials [3]. Since variable cutting speeds are common in industrial settings, the study presented in [4] aimed to develop a method for predicting tool life under such conditions.

Predicting tool wear or damage, along with estimating its remaining useful life (RUL), is essential for effective monitoring of the machining process. This can be done by using reliability function models based on tool wear behavior, data-driven models utilizing signals from the process, workpiece, and tool, or hybrid models that combine both approaches [5].

In reliability function models, wear is described statistically by establishing a reliability function from empirical data. These models assume that tool degradation follows a specific probability distribution, with parameters estimated from the full wear dataset. Methods such as Gaussian process regression, hidden Markov models, Bayesian models, and adaptive hidden Markov models are commonly applied. A key step in these approaches is collecting accurate wear data and choosing a suitable distribution.

In evaluating tool surface degradation and reliability related to catastrophic failure, it is crucial to identify when a component of the machining system vector first exceeds a critical level. This is complex due to the system’s multidimensional nature. The time to such failure is typically treated as a random variable, with its probabilistic characteristics derived from the system’s statistical properties.

Many methods for tool wear assessment have been developed [6], with most focusing on identifying critical wear indicators, as summarized in Tab. 1. The data in Tab. 1 were obtained from research by D&H Innovations Ltd. and D&H Engineering Ltd. between 2021 and 2023. Certain wear types involve accelerated degradation of the cutting blade beyond technologically justified limits. However, such occurrences are rare, happening in less than 1% of all tool changes, a rate even lower than other operational events like retooling. This low incidence highlights the reliability of existing wear patterns and supports the feasibility of developing accurate tool wear models. These models can be confidently implemented in production to improve monitoring, optimize tool usage, and reduce unexpected downtime.

Fig. 1.

a)Reasons for premature tool replacement, b) forms of tool wear

Monitoring and prediction of tool wear has been the subject of many studies, where possible research methods have been indicated. In [7], a literature review is presented on tool wear, its description, monitoring, and RUL prediction in the context of big data. The authors proposed research directions in the area of tool wear. Most often, taking into account the critical value of tool wear, the reference value model was used, which is shown in Fig. 2.

Fig. 2.

a)Measuring reference data, b) off-line identification of reference value model

Sensor data can refer to any measured quantity that shows a correlation with the amount of wear [8]. For example, in [9] the development of a wear measurement system based on the AE sensor was developed with an appropriate data processing system adapted to their properties. In works such as [10] and [11], the focus was mainly on developing signal processing for the tool wear inference system. In [12], methods for processing data to obtain the best RUL prediction are discussed in detail. In [13], the authors analyze changes in tool geometry, focusing on wear, design modifications, and operational factors, while presenting methods to describe these changes through measurements and descriptive metrics.

Abrasive wear depends on the physical properties of the interacting material pair – the tool material and the workpiece material, the stereometric features of the cutting edge, and the dynamic properties of the machining system. Changes in these factors over time lead to variations in the wear rate.

By understanding how the wear rate changes over time – that is, the tool wear intensity I = dVB/dt – it is possible to determine the tool life at a given time as the inverse of the tool wear intensity 1T=∫0VBcrit1I(VB)dVBT = \int_0^{V{B_{crit}}} {{1 \over {I(VB)}}} dVB

VB – tool wear indicator,

VB_crit – critical value of the tool wear indicator.

The relationship between tool wear intensity and the actual tool life can be modelled using various functions. For constant tool wear intensity, the simplest and most commonly used approach is a linear model. It works well under stable cutting conditions with wider workpiece tolerance limits and effectively captures the linear portion of the wear curve, based on time or the number of workpieces processed. Tool replacement is then triggered after reaching the defined threshold.

Until the critical value of the tool wear indicator is reached, the tool position is continuously adjusted according to the progressive wear model. However, in cases of accelerated wear – after exceeding the critical value – even corrective measures may fail to keep up with rapid tool degradation. Therefore, to prevent sudden tool failure caused by cumulative cutting effects, the prediction range is intentionally limited to avoid accelerated wear.

The initiation of the tool change process can also occur in the case of a loss of stability in the manufacturing system. Process stability is determined based on measurement results and process stability data. The initiation of the tool change process carried out by the Intelligent Tool Change System, is based on information provided by the management computer, which monitor for any process trends in Statistical Process Control charts. The emergence of such trends signals a potential loss of process stability, leading to the application of artificial intelligence techniques to evaluate the state of the process and tools.

A review of AI integration into CNC systems is provided in [14], while [15] separates the tool wear process into monitoring and prediction phases. An advanced method using deep learning is described in [16].

From the analysis conducted, the modelling of wear progression over time in modern systems takes the form of a recursive model, determined according to the scheme presented in Fig. 3. Wear values are predicted recursively based on previous values, considering the measurement data collected from the process.

Fig. 3.

Recursive model of tool wear

In the application for tool wear prediction in the Intelligent Tool Change System, the model must be supplied with real-time data from the process. Under these conditions, the recursive model can predict the wear of a cutting tool over time, as tool wear is a gradual process that depends on the number of workpieces processed. Consequently, the model recalculates wear parameters with each new measurement.

The recursive model prediction relies on the quality of the acquired data, as the model update and tool wear prediction primarily depend on new data collected during the machining process. The model evolves with each additional measurement, adapting to real-time changes in the tool's wear state.

The cutting tool wear model occurs at various levels of the cyber-physical model. The basis for choosing the modelling method is the purpose for which the model is to be developed. The cutting tool wear model is a model of tool deterioration, which consists in the fact that the condition of a cutting tool becomes increasingly worse and gradually causes the tool to lose its ability to perform in line with expectations. However, the wear model is associated with the possibility of developing an inverse model and indicating the usefulness of a given tool, how useful it is in removing machining allowance. The definition of RUL is closely related to the wear model. On the other hand, the tool usefulness model is the possibility of introducing corrections to the process. It follows from the considerations that many basic assumptions should be taken into account when building the model. They are systematically listed in Tab. 2.

The development of models of blade wear over time and determination of RUL were carried out using monotonically changing values of wear indicators, which have a direct impact on the value of blade shortening and surface precision. Determination of the reference value of critical wear enabled the model to be adjusted to the actual values of blade wear in given cutting conditions. The biggest limitation in the use of models with a reference value is the problem of the dispersion of wear intensity and tool life. Studies show the randomness of factors that cause deterioration of cutting properties of tools, where, apart from abrasive wear, other wear mechanisms, such as diffusion or chemical wear, begin to dominate. Defining the tool life in the conditions of randomness of the tool wear process requires determining the time of its reliable operation until the blade blunting criterion (reference value) is achieved. Problems related to determining the time of the first exit of the process from the allowable area were considered in the theory of stochastic processes. Effective solutions can be obtained from the equation of the reliability function, which depends on the physicochemical and strength properties of the blade material. These in turn depend on the temperature and machinability of the processed material. Additionally, the wear of the blade depends on its load and the dynamics of the cutting process.

The linear wear model of the form as Eq. (2) was analysed 2y(t)=y(t=0)+Aty{\rm{(}}t{\rm{)}}\;{\rm{ = }}\;y{\rm{(}}t\;{\rm{ = }}\;{\rm{0)}}\;{\rm{ + }}\;At

where the parameter A refers to the intensity of tool shortening over time and the first-order inertial model as expressed in Eq. (2) 3y(t)=y(t=0)exp(−tT)+K(1−exp(−tT))+C0y(t) = y(t = 0)exp\left( {{{ - t} \over T}} \right) + K\left( {1 - exp\left( {{{ - t} \over T}} \right)} \right) + {C_0}

The parameter T in equation (3) for T > 0 can be interpreted as the period where for t=T we reach 63% and for t=3T we reach 95% of the critical value of the wear. The parameter K can be interpreted as the gain, the maximum value in the steady state. Knowing the current value of y(t) and the value of y(t) for t=3T we can estimate the RUL.

The goodness of fit of the models was assessed using the NRMSE index. The goodness of fit values are within the range 〈–∞; 1〉, where “0” – perfect fit to the reference data (zero errors) " – ∞" – poor fit, “1” – the model values do not fit better than the fit of the reference value with a linear model. Tab. 3 presents the results for the average NRMSE values for the goodness of fit of the reference data for the linear and first-order inertial models for all experiment points. The results indicate the advantage of the first-order inertial model over the linear model. The fit was significantly better, and the average NRMSE value was half as small.

Fig. 7 shows an example of a single cutting process, in which the collected reference data of the tool wear for 75% of the critical value. The reference data were modelled with a linear model and a first-order inertial model. For this specific cutting case, the linear model seems to be a better fit than the first-order inertial model. The fit was similar for the linear model and the first-order inertial model. However, statistically for the entire experiment, the results for the first-order inertial model indicated its advantage over the linear model.

Fig. 7.

Example of application of linear and first-order inertial models in the estimation of tool wear

Despite the good fit of the models, it can be observed that the measured values and the values estimated by the models, when compared to the reference value under narrow tolerance conditions, do not ensure complete compliance of the items with the specification. To perform a more comprehensive analysis, Fig. 7 presents the measured value of Ddmax for selected experimental points (10% of the critical wear value) along with the cumulative value of tool position corrections. The value of Ddmax and the cumulative value of the tool position correction, adjusted for the current measurement value, remain in high agreement, which was confirmed by a statistical test for equality of means. The plots in Fig. 8 have been supplemented with estimated cutting edge reduction values for the linear model. The linear model was determined using the least squares method based on all the measured data.

Fig. 8.

Tool wear cases as a function of the number of workpieces, highlighting the Ddmax values

The chart in Fig. 8a indicates that a tool position correction was made for the fifth and seventh workpieces. The corrections were significant, amounting to 10 μm each. Introducing the correction based on the measured dimension values of the workpiece enabled the production of subsequent workpieces in compliance with the specification. Similarly, the situation unfolded in the remaining cutting trials, as shown in Fig. 8b–8d.

In this particular case, all workpieces were measured, and the correction was made in the subsequent step. This is an ideal situation, as statistical process control is used in production, and corrections are applied with a certain delay relative to the current measured value. Therefore, there is a need to predict the tool position correction values for the production of subsequent workpieces.

If the linear model identified based on multiple repetitions of the experiment were applied, it would be evident that its use in production would not guarantee an accurate reflection of the wear progression. The corrections calculated from this model might not ensure the production of workpieces within the specified dimensional tolerance. For example, the intensity of tool wear, according to Fig. 8a and 8c, is much greater than in the case of the experiments shown in Fig. 8b, where tool position corrections were made only for the 14th workpiece. The linear model suggested an earlier correction; however, despite the delay, the intensity of tool wear was small enough that the workpieces remained within the accepted tolerance range.

The analysis of cases involving tool position corrections and wear progression indicates the need to train the models with reference values during the machining process in order to predict wear values. This, in turn, would enable the effective application of tool position corrections.

In the case of the Intelligent Tool Change System, the model is developed based on historical data of the process trace and tool wear. The maximum cutting edge reduction and the number of workpieces machined with the given tool are determined. As indicated in section 4.2, these values do not guarantee achieving the required workpiece accuracy due to the cutting edge reduction but serve as baseline data for developing the appropriate model selection strategy.

The adaptive linear model and first-order inertial adaptive model were developed in two versions, the results of which are shown in Fig. 9 and Fig. 10.

Fig. 9.

comparison of tool wear measurements and data generated using the linear model, as well as the adaptive linear model (a) and first-order inertial adaptive model (b), as reference data for tool position correction for the first 10 pieces

Fig. 10.

A comparison of tool wear measurements and data generated using the linear model and the adaptive linear model (a) and first-order inertial adaptive model (b), as reference data for tool position correction, considering a smoothing window for the first 10 pieces

The adaptive linear model (Fig. 9a) and first-order inertial adaptive model (Fig. 9b) more accurately represent the tool position corrections than the linear model. The smoothing of the curves by averaging predictive tool position data from the current and previous models allowed for the model to be smoothed. The use of measured values for model fine-tuning means that the developed recursive model could be applied to predict wear values for subsequent workpieces.

The strategy for determining and fine-tuning the models was verified using experimental data from a real production process in the Intelligent Tool Change System. For the manufacturing process being carried out, based on workpiece measurements taken every 5 pieces, corrections were made using the adaptive linear model. The model was fine-tuned as new measurements were received and predicted the cutting edge reduction values based on a monotonic function model with an increment of 5 μm. The prediction and actual measurements were in high agreement. For the sample data from Figure 7, the goodness of fit index NRMSE for the model was 0.2674, and after averaging, it was 0.2187 (Tab. 4).

The use of the first-order inertial adaptive model for all experiment points yielded better results than the adaptive linear model. For the sample data from Fig. 7, the goodness of fit index NRMSE for the first-order inertial adaptive model was 0.1890, and after averaging, it was 0.1783 (Tab. 5). The prediction allowed for real-time tool position correction, ensuring that all workpieces were produced within the specified dimensional tolerance.