1 Introduction

Accurate tool life prediction during machining is crucial to expand the range of autonomous machining operations. Tool inserts are subjected to a significant rubbing action during the metal cutting. Throughout the turning operation, the tool comes into the interface with the metal chips and work material. Due to extreme thermal gradients and stresses close to the tool surface, the condition becomes critical [1]. Many businesses use automation to compete in an environment where technology is developing quickly and the lifespans of products are getting shorter. The fast development of machine learning methods has an opportunity to explore automation. Artificial intelligence (AI) systems are becoming more widespread and sophisticated, lowering prices and increasing scalability in various industries [2]. This change substantially influences how products are manufactured, leading to decreased human error, improved efficiency, and higher productivity [3]. The most significant and crucial manufacturing challenge is assessing and forecasting tool life. Estimating tool life requires a lot of time and resources that are, therefore, fairly expensive. Therefore, it is essential to anticipate tool life to maximize machining productivity and cost accurately. Predicting the life of the tool is a significant problem for industrial automation from the perspectives of the overall time of manufacturing, raw-material incurred, and manufacturing expenses. This issue requires the use of AI and machine learning to be resolved. Machine learning is a crucial component of any program or operation. Prior to a few years ago, making accurate predictions was difficult. Tool life can be precisely predicted with AI and several machine learning software breakthroughs without requiring much effort or data. To increase the precision of the CNC machining operation, it is possible to use various computer vision and prediction methodologies in the industrial sector. This helps in lowering the cost of production by maximizing the use of cutting inserts throughout the manufacturing system [1].

A considerable body of research on tool wear describes techniques for predicting the life of the tool and cutting state. In this direction, many researchers proposed several direct and indirect approaches to track tool wear [1] [4]. To assess wear on the tool, the indirect approach correlates characteristics such as temperature, vibration, power, force, and surface roughness. In contrast, a computer vision system is used in direct ways to track tool wear. Compared to indirect approaches, image processing methods offer benefits like machining without load or measurements made without touch. A few key factors, including lighting, the performance of image processing algorithms, and distortions in the images, must be considered for its reliability and accuracy. To track tool wear, by examining signals from numerous sensors to keep track of turning operations, Karam et al. [5] employed a sophisticated approach to forecast tool insert life. To predict tool wear, Kong et al. [6] investigated the relevance vector machine using the indirect method and evaluated how well it performed compared to ANN and SVM. Ong et al. [7] created a wavelet-based prediction algorithm, investigated how machined component surface roughness data affected the wavelet neural network and used neural networks to evaluate the average error. Kong et al. [8] adopted indirect techniques. The precision and effectiveness of neural networks and Gaussian process regression were examined to predict tool wear. However, the reliability of the Gaussian regression model is distorted by the underlying disturbance in the manufacturing environment. and lowers the prediction algorithm effectiveness. Rajeev et al. [9] suggested an approach by applying an artificial neural network to estimate wear on the tool for hard machining of AISI 4140 steel. Using multiple sensor data, including tool vibration, force, and surface roughness data in the machining of hard steel, Kene and Choudhury [10] proposed a unique cutting insert condition monitoring method with approximately 8% wear estimation error. Mohanraj et al. [11] summarized indirect tool wear monitoring methods in milling operations. The different cutting insert life estimation techniques are described. The monitoring accuracy of the system can be improved by using various sensors. Bagga et al. [12] suggested a tool wear assessment method using indirect techniques. The data that had been gathered was processed using an artificial neural network. With a maximum inaccuracy of 5.38%, a strong connection between measured and anticipated tool wear values was discovered. Convolution neural networks, reinforcement learning, and recurrent neural networks were reviewed by Serin et al. [13], who also found a new use for deep learning in tool damage detection. Bergs et al. [14] employed the deep learning method to process images in a two-step approach. Convolutional neural networks were employed in the first stage to classify, and in the next stage, convolutional neural network was used for segmentation. For turning operations in CNC machines, Sun and Yeh [15] created a tool condition monitoring method. The four conditions found in turning tool inserts by the suggested method are built-up edge, fracturing, flank wear, and chipping. Sun et al. [16] proposed a method for anticipating cutting insert conditions by applying deep learning techniques. The performance of the forecasting model was enhanced by median and mean-based modifications. Chetan et al. [17] chose the ideal cutting conditions using noise data and machine vision signals. Automation is increasingly being employed by manufacturing organizations as a competitive advantage worldwide. Due to the higher worker costs recorded in several countries, automation systems realization is required in the sector to cope with low-cost countries [18].

Direct techniques for measuring tool wear are far more accurate than indirect approaches. Furthermore, a sizable data set is needed to find a relation between the measured parameters with cutting insert wear by applying indirect approaches. On the other hand, the direct approaches are quick and precise when determining the wear of the tool during machining. Additionally, much research has been done on estimating the remaining life of the tool, wear forecasting methods, and several methods (indirect and direct) for predicting cutting insert life [19]. In order to anticipate tool life in turning process, Mikoajczyk et al. [20] explored several machine vision approaches and applied the neural network toolkit. Ruitao et al. [21] introduced a novel method for monitoring tool wear and changing tools at the appropriate time using computer vision techniques to decrease tool costs and increase tool usage. The relative wear measurement inaccuracy of 7.5% was found between values recorded using a microscope and the proposed system. Bhat et al. [22] proposed a novel wear classification technique based on support vector regression of machined surface images. In-cycle tool wear measurements are possible when machining using a computer vision technique that has been put out by Bagga et al. [23]. To capture pictures of tool wear, the authors employed a camera with CCD (charge-coupled device) sensor. They then used various computer vision techniques to measure cutting insert wear. With advances in computer vision techniques, the suggested methodology offers greater accuracy and versatility. In order to identify tool wear, several writers [24] [25] [26] [27] offered various ways to evaluate the images. Bagga et al. [28] developed an intermittent wear measurement system built on a vision system. Image processing was used to calculate the measurement error, which was 4.98% compared to manual wear measurement. The researchers found that a powerful illumination system is required to evaluate the wear area more precisely. Using texture, form, geometry, and grayscale characteristics in the surface image, Hu et al. [29] utilized a genetic algorithm to categorize surface abnormalities. When performing micro-milling operations, tool insert wear was measured using an image processing technology by Dai and Zhu [30]. A CCD camera and an LED light were employed to achieve homogenous illumination.

The current tendency is to use SVM and boosted trees to address prediction issues that fall under either classification or regression issues [31] [32]. These techniques are less prone to overfitting than other prediction models like neural networks or K-nearest neighbor. This algorithm is more effective for generalization. SVM-based algorithm development was first carried out for classification issues and then regression issues [33]. Directed graphs representation of trees is easy to comprehend, which makes it easier to understand the process being researched [34]. When compared to a single tree, ensemble approaches benefit from the collective knowledge of several trees. Prediction accuracy is increased by acquiring the array of weak learners’ abilities. On continually partitioning the information based on inquiries, a gradient-boosted tree is created [31]. Overfitting is possible for a single tree. Therefore, the overfitting issue can be resolved, and the accuracy of these decision models increased by merging multiple decision trees [35]. Unlike random tree forests, which combine trees concurrently to generate an ensemble, boosted decision trees join each tree in sequence, where a tree is created onto its progressively connected weak learners. Each tree makes an effort to minimize the errors of the one before it. There are various trees in the random tree forest model. It makes predictions by taking into account the average output of multiple trees. As opposed to the random tree forest algorithm, boosted trees algorithm is built on each tree separately, building each one on the one before. Each tree is therefore created using the residuals of the previous tree. As a result, the algorithm’s accuracy steadily improves.

Industry 4.0 has been adapted due to the growing need for high-quality items at competitive production costs. The necessity of automated tool condition monitoring is growing with advancements in hardware, computation, and computer vision methods. If the tool is dull or damaged, it will significantly reduce the workpiece quality and the productivity of the machining processes. By developing robust computer vision systems, optimum use of tools can be achieved without compromising part quality. The creation of machine learning models for tool life prediction is the goal of this study. The proposed approach uses a direct method to capture and process images of the wear zone to monitor the tool wear. With an automated process, time-consuming and tedious manual intervention for tool-wear monitoring is no longer required.

The literature study revealed limited research on developing prediction methods for turning operations by utilizing machine vision technologies. Additionally, prior research has concentrated on using indirect methods to forecast tool life, such as measuring temperature, acoustic emission, vibrations, and forces. No prior research had looked into employing gradient-boosted trees and SVM to forecast tool wear directly using computer vision techniques. Considering these issues, the proposed research aims to fill the gap between creating a machine learning-based and machine vision-based model for predicting the cutting insert tool. In order to forecast tool wear during turning operations, this work introduces unique SVM and gradient-boosted decision trees learning systems. Consequently, the following are some of the goals of the research work:

  • Installation of a vision system mounted on the machine and illumination source to take clear pictures of a cutting tool during turning operations;

  • Creation of a process that enables computer vision methods to identify tool wear directly; and

  • Applying image processing techniques along with machine learning systems, particularly gradient-boosted decision model and SVM to estimate tool wear.

The proposed method uses captured image of the tool wear zone to calculate wear area in terms of white pixel counts. The output of the image processing is utilized to predict the remaining tool life using proposed prediction models. The machining trials generated a source dataset with one output and four inputs (white pixel counts, cutting speed, feed, and depth of cut) (remaining tool life). The use of a small sample size to train artificial intelligence systems is a current research topic. Many prediction models are merged to increase the strength of various artificial intelligence techniques, such as random forests and boosted decision trees, to improve the predictive accuracy of models with small sample sizes. These models can be used in tool condition monitoring systems with little experimental data. However, most of the previous research work is focused on the use of indirect method to collect input data for prediction model from sensors like force, current, vibration, and noise. To avoid the additional expenses in sensors and make the prediction system more economical which is suitable for its practical application in industrial environment, one input parameter from imaging sensor and three machining parameters (speed, feed, and depth of cut) are used to train the proposed model. Compared to direct manual measurement, the suggested model’s error is comparable, falls within the same range, and applies to all industrial settings.

The outcomes demonstrated how well SVM and decision trees predicted results. Figure 1 shows the experimental methodology followed. The developed system offers a broad array of industrial benefits. If the cutting-edge wear of the tool is identified during machining, the time-consuming and expensive manual measurement of the tool wear can be eliminated, lowering the chance of workpiece surface distortion caused by higher tool wear. It offers a straightforward yet efficient method for estimating the remaining tool life.

Fig. 1
figure 1

Procedure developed for the proposed scheme

Full size image

2 Materials and techniques

2.1 Experimental design and methodology

The CNC lathe machine of HMT Starturn makes used to carry out the experiments. The raw material used is AISI 4140 steel. The workpiece used in conducting experiments are cylindrical in shape of 50 mm in diameter and 250-mm length. The cutting tool material used is coated carbide insert with ISO designation SNMG120408 of Widia make, and a tool holder with ISO designation PCLNR1616H12 is used for experimental work. Table 1 displays the level of cutting parameters chosen for the studies. The cutting parameters are selected based on the review of literature, preliminary experiments performed, and recommendations of the tool manufacturer. Furthermore, the mid-point of the given parameter is considered based on operating conditions in the industry. The tool flank wear value of 0.4 mm is decided according to ISO 3685 standard, and the machining is stopped when the decided tool wear (in mm) is reached [20] [36].

Table 1 Turning process parameters
Full size table

Compared with traditional monitoring systems, the proposed direct tool wear monitoring using machine vision is more robust and accurate. The tool condition monitoring system consists of imaging sensor, lens, and lighting system. The imaging sensor used is an industrial camera with rugged construction to operate in harsh industrial environments. It is provided with a mounting facility to mount it on the machine tool. It has a Gigabit Ethernet communication interface standard for high-speed data transfer. A telecentric lens is used which has special characteristics of orthographic projection used in measurement tasks [37]. It has a constant field of view that eliminates magnification errors. Other benefits of a telecentric lens include the absence of perspective deformation, and negligible image distortions. In order to keep the illumination constant during image capture, ring light is employed, which improves the consistency of the image capturing system. A ring light that is considerably bigger than the cutting tool is selected for a uniform light dispersion. Two groups of systems are used in a machine vision system for insert wear monitoring: software and hardware. A camera with a complementary metal oxide semiconductor sensor, a lens, an LED light (shape of a ring) for illuminating the wear area, workpiece, and a cutting tool make up the hardware component.

Images of a cutting tool insert are taken using a Baumer® EXG50 monochrome industrial camera, which features a CMOS sensor, a frame rate of 14 frames per second, and a resolution of 2592 × 1944 pixels (Fig. 1). It was chosen above other cameras because of its smaller size, easier installation, shorter shutter duration, higher frame rate capacity, and higher shutter speed. It uses the Computar® TEC-V10110MPW, a bi-telecentric lens with a 1× magnification, a 110.2-mm focal length, and 0.015% distortion. Because object light rays are parallel to enteraxis of the image sensor, a bi-telecentric lens is chosen independent of the field of direction. The light rays can pierce without relying on the depth field. The flank wear region of cutting tool inserts is illuminated by the white LED ring light, model number “Tolifo-R-160S.” The imaging system is connected to a computer, which also functions as the machine vision system’s storage device, frame grabber, and information exchange system. The computer runs Windows® and has an Intel® core 5 duo processor clocked at 2.4 GHz and 8 GB of Random Access Memory. Additionally, it features the camera-triggering software Baumer® camera Explorer (Fig. 2).

Fig. 2
figure 2

A test setup using a machine vision to assess cutting insert wear

Full size image

The imaging hardware is mounted on a CNC lathe machine. At the start of machining, turning operation for the given length of cut is performed. Following the turning cycle, the tool is moved to a predetermined spot in front of the image sensor, where a photo of the tool wear zone is then captured. To ensure that the image of tool wear region is consistently recorded in the same location, the tool’s position is provided in the CNC part program. So, the image of the tool is taken when the tool is in a standstill position. As a result, monitoring the tool’s condition is automatic and essentially applicable in industrial settings. Therefore, the tool image capturing is online but intermittent. Thus, the image is obtained when the tool is standstill and frame rate and shutter speed are less important parameters for capturing images.

3 Machine vision algorithms

The tool wear is represented as an area in terms of number of pixels. Independent variables namely, speed, feed, depth of cut, and wear area are fed as input to prediction models. Because of the presence of noise in the image, it is more difficult to precisely measure flank wear than wear area using computer vision methods. Therefore, to increase the reliability of prediction algorithms, wear in terms of number of pixels is considered in place of flank wear length. After the worn tool image has been obtained, image processing is used to determine the tool’s level of wear. In Fig. 3, the flowchart is displayed. It covers pre-processing of images, Otsu segmentation, extraction of data from images, and calculations of pixel values of the images.

Fig. 3
figure 3

Flowchart for cutting insert wear monitoring system based on machine vision

Full size image

3.1 Preprocessing of images

Cropping will take place later when the captured image is given to the computer vision algorithm, and it must be done carefully because it might reduce the system’s overall accuracy. Cropping should be followed by a noise reduction procedure. While protecting the boundaries of the image, noise reduction eliminates isolated noise points. Edge-preserving denoising filters use a bilateral filter. The gray values of each pixel point in the image are evenly arranged, and the gray value of each point is calculated by computing the sorting bi-lateral value. Figure 4(a) shows the image captured by the imaging sensor and Fig. 4(b) shows the filtered image of the tool. Image contrast is enhanced following the gray level histogram-stretching process (Fig. 4c).

Fig. 4
figure 4

Steps of processing an image. (a) Captured image. (b) Filtering and cropping. (c) Histogram stretching. (d) Segmented image

Full size image

3.2 Binary segmentation of image

Otsu’s method for binarizing pictures in image processing uses an adaptive thresholding mechanism. The algorithm will find the proper threshold level by iterating over all feasible threshold values for an input image. (0 to 255). The most appropriate threshold value for the image is chosen by Otsu segmentation. The tool’s segmented image is shown in Fig. 4(d).

3.3 Extracting image data

Data extraction employs pixel counting and color percentage techniques. Several pretreatment and processing methods are utilized, including several filters and various auto thresholding methods like Otsu. A binary image is created from the data, and an algorithm then counts the pixels in the binary image.

4 Tool life prediction model using support vector regression and boosted decision trees

If the usable life of tool inserts is to be estimated, measuring tool wear is crucial [38]. This study proposes a machine vision method to measure flank wear and predict it by applying machine learning. In particular, boosted trees and SVM were employed. Artificial neural networks are receiving more attention, although the random forest is still comparatively new in this sector [39]. Although the prediction accuracies of boosted tree and SVM approaches were comparable, boosted trees somewhat outperformed SVM. The accuracy of deep learning algorithms is on par with the machine learning models proposed in the present study. Deep neural networks still need to be taught despite having superior learning capacities. To train deep neural networks, many sensory inputs and extensive experimental data collection are needed, which increase the cost of the prediction system [40]. Deep neural networks also demand expensive graphics processing unit, which raises the cost of implementation [41]. Therefore, the purpose of selecting the proposed algorithms is to develop a low-cost and accurate prediction system for its adoption in metal-cutting industries compared to the currently used manual measurement approach. In machine learning, the often used algorithm for classification is SVM [42] [43]. SVM, a supervised machine learning technique, scrutinizes the data for regression and classification. SVM is frequently applied for classification, while they can also be utilized for regression. The ideal hyperplane, which separates every group, then emerges. These supporting vectors provide a synchronized reflection of each individual observation. A frontier approach is used to distinguish the two groups. [15].

The regression tree algorithm is part of the supervised learning branch. It can address issues like classification and regression [18] [44]. A tree representation is used to solve the issue; each leaf node on the tree represents a class mark, and the internal node of the tree is where characteristics are expressed. The regression tree can represent every boolean expression using discrete attributes. The categorization guidelines are described as a journey from the root to the leaf. The leaf node, the root node, and the branch node are three of its nodes. The goal is to use a decision tree to learn decision rules from prior results (training data) to develop a training model for predicting the category or value of a response variable. Boosting may be considered a sequential mechanism whereby each consecutive model makes an effort to fix the flaws of the preceding model. One kind of ensemble learning is called bootstrap aggregation. Regression trees can perform better when using the random forest method. The regression tree can be boosted using these six stages.

  • First, a subset is created from the original dataset. Initially, the weights of all the data points are equal. A foundation model is constructed using this subset.

  • In order to produce predictions, this model is applied to the complete dataset.

  • Finding this model’s inaccurate predictions is the next step.

  • These erroneous predictions are given more weight.

  • A new model is created using this dataset, and forecasts are made. This new model tries to address the shortcomings of the old approach.

  • As indicated in Fig. 5, all steps should be continued until the result is appropriately anticipated.

Fig. 5
figure 5

Flowchart of boosted tree

Full size image

Gradient-boosted decision trees are constructed one tree at a time, each attempting to fix the errors of the one before it. The gradient descent method computes the weights of the trees, which minimizes the loss function by adjusting the weights of the trees [45]. The gradient descent algorithm modifies the weights as it moves against the gradient of the loss function relative to the weights. This procedure is performed until the loss function converges or for a predetermined iteration. The final set of weights is used to make predictions.

5 Results and discussion

On a CNC lathe, turning experiments have been conducted, and photographs of the tool insert have been recorded. Table 2 presents the outcomes.

Table 2 Training set of cutting tool life measured and forecasted
Full size table

To predict tool life, this study employs machining parameters and a characteristic that was identified from cutting insert wear images; for training the models and determine the accuracy of the models, the entire dataset has been split into two groups in a 2:1 ratio. the next step was to create and simulate two predictive models, the boosted tree and SVM using MATLAB®. Pictures of cutting insert wear are taken at several points throughout the turning operation. A total of 32 data have been gathered following image processing, as shown in Table 2. Once tool wear exceeds a maximum value of 0.4 mm, tool life is deemed to have ended. Equations 1 through 3 can be used to state the performance measures of MAPE (mean absolute percent error), mean error, and RMSE (root mean square error) that have been used to evaluate the accuracy of prediction in estimating tool life for both boosted tree and SVM models.

$${E}_{m} = \frac{Pi-Ti}{Ti}$$
(1)
$$\mathrm{MAPE }= \frac{{\sum }_{i=1}^{n}\left|{E}_{m}\right| }{n}$$
(2)
$$\mathrm{RMSE }= \sqrt{\frac{{\sum }_{i=1}^{n} {(Pi-Ti)}^{2 }}{n}}$$
(3)

where T is tool life measured manually, p is tool life predicted by boosted trees and SVM, and n is sample size. Boosted tree and SVM models have minimal absolute percentage errors of 0.22 and 0.30%, respectively. Similarly, maximum absolute percentage errors of 3.77 and 4.86%, respectively, as shown in Table 3. As a result, the model’s output can be judged appropriate for use in industrial settings.

Table 3 Values of errors
Full size table

Figures 6 and 7 show that the estimated tool life value using boosted tree and SVM strongly correlates to experimental results since all the values are closely spaced around the line. In these plots, the line is referred to as the best-fit line, while all the blue spots are referred to as residuals. The difference between the residuals and the line is the error at that location. Hyperparameters are re-evaluated in order to improve the boosted tree model even more. Kernel scale, signal standard deviation, and kernel function are optimized hyperparameters. To reduce the error in the boosted tree model, the method focuses all these combinations of hyperparameters during this optimization phase. When these hyperparameters are optimized, the boosted tree model’s prediction accuracy improves. The hyperparameters are optimized throughout 300 iterations, with the optimal point hyperparameters found after 50 iterations. The RMSE value of a boosted tree has been enhanced through optimization to 0.219, which is regarded as good. The regression value (R2), 0.96 and 0.97 for boosted tree and SVM, respectively, provide additional evidence of the prediction model’s effectiveness. Boosted trees and SVM regression findings show that these predictive models do not account for about 3 to 4% of the changes in tool life. According to Table 4, the gradient-boosted decision tree and SVM models have calculated RMSE values of 0.219 and 0.332, respectively, indicating that both models perform well. Compared to SVM, the boosted trees method achieved lower prediction error. Comparing the results achieved to those published by Alajmi et al. [46], there has been a significant improvement.

Fig. 6
figure 6

True response vs predicted (SVM)

Full size image
Fig. 7
figure 7

True response vs predicted (boosted tree)

Full size image
Table 4 The forecasting model’s accuracy
Full size table

In this case, SVM and gradient-boosted decision tree algorithms are employed for supervised learning. The gradient-boosted trees performed fairly well when the results are compared to earlier studies. In the study Hu et al. [29] reported, the SVM model built on a genetic algorithm was applied and achieved an accuracy of 95%. Another study that was presented used the wavelet transform with an accuracy of 87.9% [47]. The performance of the boosted tree method is enhanced because it uses the knowledge of multiple classifiers against one for predicting tool wear in comparison to the research of Ong et al [7]. D’Addona et al. [48] proposed the methodology wherein a single classifier was applied. So compared to a single classifier, the algorithm proposed in the current study, which is based on building trees from prior weak learners that together form a robust classifier, is more accurate.

The remaining tool life was projected using constructed SVM and boosted tree models using a validation data set and the outcomes are shown in Table 5 and visually as shown in Fig. 8. Using SVM and boosted trees, respectively, the highest relative errors in predicted tool life are 7.56% and 5.89% as shown in Table 6. For SVM and boosted trees, the RMSE values are 0.356 min and 0.579 min, respectively. Similarly, SVM and boosted tree have MAPE values of 4.0% and 2.3%, respectively. The values are depicted in Table 7. Results demonstrate that the boosted tree model outperforms the SVM model. The effectiveness of the data shows that a system for monitoring tool wear can be helpful in an industrial setting.

Table 5 Using training data to validate the measurement of flank wear
Full size table
Fig. 8
figure 8

Results of SVM and boosted tree algorithm along with experimental data for tool life prediction for (a) tool insert edge 1, (b) tool insert edge 2, (c) tool insert edge 3, and (d) tool insert edge 4

Full size image
Table 6 Percentage of error for validation set
Full size table
Table 7 Error rate for the validation set
Full size table

Figure 9 compares the boosted tree and SVM tool life prediction errors. With a maximum MAPE value of 4% and RSME of 0.579 min, the prediction error for SVM is larger than that for the boosted decision tree. The comparison shows how the suggested system, which tracks tool wear during turning, might be used in various industries. The machining variables and total pixels in the wear region are the inputs for the prediction method. Compared to the method mentioned by Mikoajczyk et al. [49], which used the same cutting parameters for the prediction of tool life using neural networks, the proposed prediction models can be implemented for various cutting situations.

Fig. 9
figure 9

Boosted tree and SVM output error comparison

Full size image

Utilizing cutting tools to their fullest potential is one method an industry can use to lower its production costs. With an online tool wear monitoring system, the laborious and costly tool-wear measurement will not be required. The experimental findings demonstrate that applying the proposed tool wear monitoring technique, which is reliable under a wide range of cutting parameters, can predict the in-process remaining tool life in machining. Additionally, this method will substantially automate and simplify the determining of cutting-edge life. The groundwork for tool wear monitoring in actual manufacturing environments is laid forth in this study. The prediction accuracy rates with minimum error showcased by the proposed method are significant in tool condition monitoring, hence resulting in optimal tool replacement, improved part quality, and production rates.

6 Conclusions

Important insights into the necessity for tool wear prediction in machining employing image processing to examine the measurement of tool wear and prediction of the tool wear are offered by the review of the literature. This paper presents a two-step approach to predict tool wear during machining operations: gradient-boosted decision trees and support vector regression, and boosted tree techniques are used to predict the tool life. Both predicting models are created using inputs such as machining factors and total pixels in the wear region from images. The projected values of the life of the cutting tool insert derived using the gradient-boosted decision tree and support vector machine methods are contrasted with the actual tool life values. The research results are described as below:

  1. 1.

    In an experimental condition, a computer coupled with a camera is deployed to record tool wear photographs and analyze them to obtain helpful image elements. These relevant image properties are then sent to the boosted decision tree and support vector machine algorithms to predict the life of the cutting insert.

  2. 2.

    An optimized gradient-boosted decision tree was developed to enhance its effectiveness by adjusting hyperparameters. The data analysis revealed that the enhanced gradient-boosted decision tree produced the best MAPE, RMSE, and R-squared values.

  3. 3.

    The outcomes demonstrate that boosted trees outperform SVMs by being more efficient in terms of MAPE and RMSE.

  4. 4.

    The SVM and enhanced boosted decision tree methods result in 97% and 96% prediction accuracy rates, demonstrating the efficacy of prediction algorithms. By pitting the results of several classifiers against one another and building a model using earlier weak learners, which, when combined, produces a robust classifier, the performance of the boosted trees technique is improved.

  5. 5.

    The validation of the suggested models revealed that the SVM and boosted tree produced maximum absolute errors of 5.89% and 7.56%, respectively, and can be applied for predicting the life of the tool in an industrial environment.