Machine-Learning- and Internet-of-Things-Driven Techniques for Monitoring Tool Wear in Machining Process: A Comprehensive Review

Abstract

Tool condition monitoring (TCM) systems have evolved into an essential requirement for contemporary manufacturing sectors of Industry 4.0. These systems employ sensors and diverse monitoring techniques to swiftly identify and diagnose tool wear, defects, and malfunctions of computer numerical control (CNC) machines. Their pivotal role lies in augmenting tool lifespan, minimizing machine downtime, and elevating productivity, thereby contributing to industry growth. However, the efficacy of CNC machine TCM hinges upon multiple factors, encompassing system type, data precision, reliability, and adeptness in data analysis. Globally, extensive research is underway to enhance real-time TCM system efficiency. This review focuses on the significance and attributes of proficient real-time TCM systems of CNC turning centers. It underscores TCM’s paramount role in manufacturing and outlines the challenges linked to TCM data processing and analysis. Moreover, the review elucidates various TCM system variants, including cutting force, acoustic emission, vibration, and temperature monitoring systems. Furthermore, the integration of industrial Internet of things (IIoT) and machine learning (ML) into CNC machine TCM systems are also explored. This article concludes by underscoring the ongoing necessity for research and development in TCM technology to empower modern intelligent industries to operate at peak efficiency.

1. Introduction

Future Industry represents a paradigm shift in manufacturing, blending advanced sensor technologies, embedded systems, machine learning, and cloud computing technologies with human ingenuity to create smarter, more efficient, and more sustainable production systems. This evolution not only enhances operational performance but also introduces new opportunities for innovation and growth in the manufacturing sector [1]. These technologies can streamline the obtaining and processing of real-time data of a computer numerical control (CNC) machine, which is critical for achieving real-time monitoring and optimizing the metal-cutting process to meet the demands of smart manufacturing.

CNC machines are vital equipment in the modern manufacturing industry, and the integrity of cutting tools is of utmost importance in the metal cutting industry. The damages caused by these tools can have a significant impact on the overall efficiency and productivity of the manufacturing process. It is also identified that up to 20% of machine downtime is due to tool wear alone [2]. On continuous running, the CNC turning center machine cutting tool will become damaged due to improper machining parameters, material selections, and environmental conditions, which leads to a poor surface finish on the machined component. Thus, monitoring the cutting tools continuously and changing them when needed is a must. Changing cutting tools too late can result in poor workpiece quality and replacing them too early can result in less utilization and an increase in overall processing costs. Previous studies have shown that cutting tools are typically used within a range of 50–80% of their effective life. Therefore, it is important to monitor the condition of the cutting tools during use by implementing a TCM system to determine when they need to be replaced [3].

This monitoring can take place in two ways, which are the direct and indirect methods. Direct monitoring methods often involve gathering visual data from optical sensors and probes such as cameras [4]. The optical sensors and probes used in direct monitoring methods are expensive and may require specialized installation and maintenance. As a result, many tool condition monitoring systems prefer indirect monitoring methods that use sensors to capture signals such as cutting force, vibration signature, acoustic emission (AE), and tool temperature [5]. In the indirect approach, previous studies have commonly used the dynamometer, thermal imaging camera, accelerometer, and AE sensors to determine the degree of tool wear and surface roughness of the component being machined. However, the cost and inconsistency of sensor measurements can be a limitation, particularly under different operating conditions and during the real-time monitoring of tool conditions [6]. The TCM system undergoes different stages for monitoring the cutting tool, namely, sensor integration (contact type and non-contact type), signal collection, signal processing, feature classification, tool wear state and fault classification, and decision-making [7], as shown in Figure 1.

The modern TCM systems that use sensors to continuously monitor the cutting tool’s condition are much more effective at predicting and preventing tool wear. Using IoT technology and machine-learning algorithms to analyze the data can provide operators with real-time information about the machining process, which allows them to act quickly and avoid unexpected downtime or damage to the machine or tool [8]. The signal processing stage of TCM is crucial and needs more importance as it helps to analyze and interpret the input signals to identify tool wear. The use of signal-processing techniques such as fast Fourier transform (FFT), wavelet packet decomposition (WPD), ensemble EMD (EEMD), and Hilbert–Huang transform (HHT), along with frequency spectrum analysis can provide valuable insights into the cutting process [9,10,11,12]. The time-domain analysis is limited in its ability to display changes in a signal over time, as it only provides information about the signal amplitude and time duration.

By combining multiple sensors and analyzing the data using multi-sensor fusion techniques, it is possible to collect more information and patterns related to cutting tool wear. Multi-sensor fusion techniques can improve the accuracy and reliability of TCM systems [13]. Selecting the most important features and variables that are correlated with tool wear is essential for developing effective multi-sensor TCM systems. This process involves identifying the most relevant sensor data that can be used to accurately predict tool wear and other problems. A flexible multi-sensor approach for TCM systems has been developed that uses vibration, current, and force signals to monitor tool wear. This approach involves analyzing the total power obtained from the cutting process, consistent with the tool wear curve [14].

The investigation was carried out by Deo et al. [15] on the white-box support vector machine (SVM) and swarm-based optimization with meta-heuristic algorithms for the TCM. This approach can help optimize the monitoring process and improve the prediction accuracy. The performance comparison of five different meta-heuristic algorithms was involved in his research. The research of Venkatesh et al. [16] demonstrated the potential for the transfer learning (TL) model to be an effective approach for TCM. The testing and comparison of different models and the effects of hyperparameters optimized the process and improved the accuracy of manufacturing processes. The use of Bayesian optimization search to fine-tune hyperparameters is a promising approach that can enhance the accuracy of predictions and classification in TCM. The research conducted by Bajaj et al. [17] on the fine-tuned hyperparameters of Bayesian optimization search demonstrated an impressive accuracy rate of 93.3% in TCM indicating that his approach is industry-ready. Moreover, using a custom-made data acquisition (DAQ) system highlights the potential for customized hardware solutions to improve TCM in manufacturing processes.

Stuhr et al. [18] proposed a flexible and minimal TCM system that can effectively detect and diagnose tool wear and failure during repetitive machining operations, using similarity analysis techniques, by comparing the collected sensor data during machining to data of known tool conditions. You et al. [19] developed a system that utilizes a lightweight image-processing technique. Additionally, the system uses a neural network with multiple activation functions to improve the accuracy and addresses several challenges that are commonly encountered in TCM systems, such as issues with image quality, model parameters, dataset scale, cloud migration, and the use of embedded devices.

The review by Patil et al. [20] focused on the use of deep l(DL) based methods for TCM in CNC milling operations. Their framework was designed to be versatile and replicable, allowing it to be applied across different milling operations and tool types. Their article evaluated the recent attempts to enhance TCM systems, including the use of advanced signal-processing techniques, feature extraction methods, and novel sensor technologies. Schwendemann et al. [21] conducted a survey specifically focused on TCM for the bearings in grinding machines. The purpose of this survey was to investigate the current state of TCM in grinding machines and identify the challenges and opportunities associated with this technology. Pimenov et al. [22] conducted a review and comparison of direct and indirect sensing methods for TCM in milling operations. The purpose of this review was to provide insights into the advantages, limitations, and prospects of different measurement methods for TCM in milling. Mohamed et al.’s [7] article centers around the advancement in wirelessly embedded sensors and the utilization of techniques for reducing dimensionality in preparation for the implementation of industrial Internet of things (IIoT) TCM systems.

Butler et al. [23] presented a comprehensive review of contemporary techniques and methodologies used in feed drive TCM and prognostics health management (PHM), classified into four primary categories: sensorless, sensor-based, signal-based, and machine-learning-based. The authors also emphasized the significance of monitoring the mechanical constituents of feed drives and provided further details on these constituents. Serin et al. [24] provided a comprehensive overview of TCM and the latest DL techniques, convolutional neural networks (CNNs), and recurrent neural networks (RNNs) were discussed along with their potential applications in the manufacturing industry. Nasir et al. [25] forecasted the opportunities for data-driven smart manufacturing, including feature engineering automation, high-dimensional data processing, sensor fusion, and hybrid intelligence. This article provided a comprehensive overview of intelligent monitoring and the different ML and DL models used for TCM, along with their potential applications.

These recent works of literature have identified a gap in exploring interconnected smart factory networks, which offer significant advantages. These networks enable the sharing of digital knowledge related to tools and machining processes through a cloud database. This database serves as a reference point for local computers, allowing them to predict and detect issues using standardized and frequently updated TCM models. This smart factory network enhances system performance, facilitates the widespread adoption of TCM, and reduces the reliance on local computing hardware and custom installations. Machines within this network can access information about tool faults that may not have occurred locally but have been recorded elsewhere. The database can also automatically clean up extraneous data, optimizing storage, reducing stress, and ensuring network stability. Such interconnected factories can range from multiple plants within a local manufacturing area to a global network of worldwide factories. To support this network, a stable cloud service provider with a suitable database capacity, robust cloud computing power, and efficient TCM system optimization is essential.

Therefore, this article highlights the need for advancements in the emerging TCM technology to realize the vision of smart factory networks. It focuses on several key research areas, including the integration of IIoT in TCM; this review emphasizes the pivotal role of IIoT in advancing TCM systems, especially within the context of Industry 4.0. It explores how IIoT enables real-time monitoring, data fusion from multiple sensors, and low-latency transmission, which are critical for effective TCM and emerging ML techniques, particularly DL models such as long short-term memory (LSTM) networks, and their application in TCM. It provides insights into how these models overcome challenges like small datasets and noise in sensor data, offering improved accuracy and reliability in tool wear prediction.

This article also discusses virtual machining techniques, presenting them as a promising approach for simulating and optimizing machining processes. It highlights the potential benefits of virtual machining in TCM, such as reducing the need for physical experiments and enabling the better prediction of tool wear. The generalized model for TCM based on the literature is shown in Figure 2. Based on the TCM model, each section of this paper provides a detailed analysis.

Section 2 details the sensor signal propagation and data acquisition methods, Section 3 explains the data-processing methods, Section 4 explores the advancements in ML and DL for TCM, Section 5 explains the IIoT and its usage in modern TCM, and Section 6 shows the current challenges and prospects for the future, followed by Section 7, which concludes the review with its main goal.

1.1. Evaluation of Modern TCM

• Early Beginnings: 1950s–1970s
  • 1950s: The concept of monitoring tool wear and its impact on machining quality began to gain attention. Researchers focused on understanding the basic wear mechanisms and their effects on the cutting process.
  • 1960s: The first experimental studies were conducted to measure tool wear using direct methods, such as optical and microscopy techniques. These studies primarily focused on wear patterns in traditional machining processes like turning and milling.
  • 1970s: The emergence of numerical control (NC) machines highlighted the need for more systematic approaches to monitor and manage tool wear. The initial efforts in this era involved simple manual inspections and scheduled maintenance, but the limitations of these approaches led to an increased interest in automated monitoring systems.

2.

Introduction of Sensors and Signal Processing: 1980s–1990s

• 1980s: The development of sensors capable of measuring physical quantities such as force, vibration, and acoustic emissions led to the first generation of automated TCM systems. These sensors were integrated into machining systems to capture data that could be analyzed for signs of tool wear.
• Late 1980s–Early 1990s: Researchers began exploring the use of signal-processing techniques, such as fast Fourier transform (FFT), to analyze sensor data in real time. This marked the beginning of indirect monitoring methods, where the condition of the tool was inferred from patterns in the sensor signals.
• 1990s: The use of artificial intelligence (AI) techniques, including fuzzy logic and rule-based systems, was introduced to enhance decision-making processes in TCM. These early AI applications focused on interpreting sensor data and providing automated alerts for tool replacement.

3.

Advancements in Machine Learning and Early IoT Concepts: 2000s

• Early 2000s: The integration of machine-learning algorithms, such as neural networks and SVMs, into TCM systems became more common. These algorithms were employed to predict tool wear based on historical data and real-time sensor inputs, offering more accurate and reliable monitoring.
• Mid-2000s: The concept of predictive maintenance started gaining traction, with machine-learning models being used to forecast tool failures before they occurred, thus reducing downtime and improving productivity. This era also saw the development of more sophisticated feature extraction techniques, both in time and frequency domains.
• Late 2000s: The idea of connecting machining systems to the Internet for remote monitoring began to surface, laying the groundwork for the IIoT. Early implementations focused on data collection and remote access, with limited real-time processing capabilities.

4.

The Rise of IIoT and Advanced AI Techniques: 2010s

• 2010–2015: The rapid advancement in IoT technologies led to the widespread adoption of smart sensors and connected devices in manufacturing environments. IIoT platforms were developed to enable real-time data collection, processing, and analytics, paving the way for smart manufacturing.
• 2015–2018: Deep-learning techniques, particularly long short-term memory (LSTM) networks, were introduced to handle the complexity of time-series data generated by sensors. These techniques improved the accuracy of tool wear predictions and allowed for a more nuanced analysis of machining processes.
• Late 2010s: The integration of cloud computing with IIoT enabled large-scale data storage and processing, allowing manufacturers to leverage big data analytics for predictive maintenance and tool condition monitoring. This period also saw the emergence of edge computing, where data processing was moved closer to the source (e.g., sensors) to reduce latency and improve real-time decision-making.

5.

Current Trends and Future Directions: 2020s–Present

• 2020–Present: Research has focused on enhancing the interoperability and scalability of TCM systems in Industry 4.0 and 5.0 environments. The combination of edge/fog computing, advanced AI models, and cybersecurity measures is being explored to create more robust and flexible monitoring systems.
• Emerging Focus Areas: With the increasing complexity of manufacturing systems, researchers are now looking into hybrid AI models that combine traditional machine learning with deep learning to handle diverse data sources. The application of transfer learning, where models trained in one domain are adapted to another, is also gaining interest.
• Future Directions: As Industry 5.0 emphasizes human–machine collaboration, future research may explore more intuitive TCM systems that integrate human expertise with AI-driven insights. The development of more sophisticated, self-adapting monitoring systems that can dynamically respond to changing machining conditions is another key area of interest.

The research and technology development in TCM has evolved from basic manual inspections to sophisticated AI-driven systems integrated with IIoT. The above timeline reflects a gradual shift from purely reactive maintenance strategies to predictive and real-time monitoring approaches, driven by advances in sensor technology, machine learning, and the advent of smart manufacturing paradigms. As the field continues to develop, it is expected that future innovations will focus on enhancing system interoperability, scalability, and human–machine collaboration, shaping the next generation of TCM systems [26,27].

2. Data Acquisition

Figure 3 illustrates the monitoring methods used in TCM, highlighting both the direct and indirect approaches. Direct monitoring methods involve measuring the tool’s wear or damage directly through techniques such as optical measurements or using dedicated sensors that can capture the wear state of the tool. Indirect monitoring methods, on the other hand, rely on the analysis of data such as force, vibration, acoustic emission, and temperature, which are correlated with the tool’s condition.

The evolution of modern TCM systems has allowed for the more accurate and efficient monitoring of tool conditions during the CNC machining process. In many previous studies, sensors such as dynamometers, accelerometers, and AE sensors have been used to monitor the degree of tool wear indirectly. These sensors capture signals related to the cutting process, which are then processed using a DAQ system [28]. The results obtained from different sensors can be inconsistent when used in different operating conditions, which can make it challenging to compare and interpret the data. Despite these challenges, sensors remain an important tool for monitoring tool wear. Cutting force sensors, for example, are particularly effective in monitoring tool wear, as there is a strong relationship between cutting force and tool wear. In particular, the magnitudes of the static cutting forces in the frequency domain are correlated with tool wear in previous research [29].

The cutting force in the feed direction is the most used signal [30]. Vibration signature analysis is an effective and widely used technique for monitoring tool wear and can provide valuable information about the condition of the tool [31]. AE sensors are also an effective and widely used tool for monitoring tool wear and detecting changes in the machining operation. The high-frequency range of the AE signals allows for the more sensitive and accurate detection of changes in the tool condition [32]. The use of root mean square (RMS) voltage is a simple and effective way to monitor tool wear using AE signals. By monitoring changes in the RMS voltage over time, it is possible to detect when the tool is approaching failure and needs to be replaced or repaired [33]. By correlating RMS values with tool wear levels, Xiqing et al. [34] have demonstrated the potential of AE-based monitoring for improving machining efficiency and reducing tooling costs. In recent years audible sound signal analysis has become a popular technique for tool wear monitoring. Unlike other monitoring techniques that require the use of additional sensors or equipment, sound signal analysis can be performed using only a microphone or piezoelectric sensor, which makes it easy to implement in practice [35]. It has been found that the maximum values of the wavelet packet decomposition coefficients have the most important correlation with flank wear [36].

Using multi-sensor signals and developing multisensory fusion techniques can improve the accuracy and reliability of the TCM system, which can help optimize machining operations [37]. The study conducted by Zhou et al. [13] on the fusion of cutting force, vibration, and AE signals states that feature selection is indeed a significant challenge in multi-sensor signals, especially in the context of tool wear. Identifying the most relevant features and variables related to tool wear from a vast amount of data collected by multiple sensors is a complex task that requires careful consideration and analysis. The development of this flexible multi-sensor approach is a significant advancement in the field of TCM systems and has the potential to improve manufacturing processes by reducing downtime and improving product quality [38]. The application and development process of different TCM-wear-sensing methods and techniques are shown in

Table 1. Application and development of TCM-sensing techniques.

Process Technology	Method	Description	Development
Visual Inspection	Direct	Use of cameras and imaging systems to directly observe and measure tool wear	Early methods relied on manual inspection. Recent advancements include automated optical systems using AI for analysis.
Thermal imaging	Direct	Infrared cameras measure the temperature distribution on the tool surface to assess wear	Initially used in aerospace industries, now integrated with real-time monitoring systems in various machining processes.
Optical microscopy	Direct	Direct examination of tool surface under magnification to measure wear	Traditionally used in labs, with modern applications including automated image analysis using machine learning.
Laser displacement sensor	Direct	Measure tool surface wear by detecting small changes in position using laser technology	Widely used in micro-machining industries, with integration into IoT systems for real-time monitoring.
Spectroscopic analysis	Direct	Analyzes material composition and change in the tool using spectroscopic techniques	Advanced methods used in specific industries like aerospace and automotive, now enhanced with AI for detailed analysis.
Capacitive and inductive sensor	Indirect	Detects proximity changes caused by tool wear or deformation	Commonly used in automated systems, advancements include integration with machine-learning algorithms for analysis.
Ultrasonic sensor	Indirect	Uses ultrasonic waves to detect changes in material properties due to wear or cracks	Applied in high-precision machining, with AI-based analysis introduced in recent years for better accuracy.
Electric current monitoring	Indirect	Monitors the electric current consumed by the machine, with variations indicating tool wear	Gained popularity in the 1980s, now often combined with AI for predictive maintenance in smart manufacturing setups.
Force monitoring	Indirect	Measures cutting forces during machining, and the variations indicate tool wear or failure	Common since the 1970s with continuous enhancements in multi-axis dynamometers and real-time data processing.
Vibration analysis	Indirect	Uses accelerometers to measure vibration which is correlated with the tool’s condition	Developed significantly in the 1990s with improvements in sensor accuracy and signal-processing techniques.
Acoustic emission	Indirect	Detects high-frequency acoustic waves generated by tool wear and other machining processes	This method has been widely adopted since the 1980s with advancements in sensor technology and data analysis using AI.

.

3. TCM Signal and Data Processing

3.1. Signal Processing

Pre-processing and processing are crucial steps in the development of a data-driven TCM system. They ensure that the sensor data used to train the tool wear prediction model are accurate, reliable, and informative, leading to a more effective and efficient system [7]. The amplification and sampling rate are important considerations for sensor data acquisition. In common, amplification is typically used to increase the low-level output signal of sensors to a level that can be accurately and reliably digitized. For a more accurate representation of the signal, the author recommended using a sampling rate that is five to ten times the highest frequency of interest [39], for the effective filtering of the raw sensor signals. The filtering techniques are used to remove the noise in the obtained signals [40].

Filtering techniques such as high-pass, low-pass, or band-pass filters can be used to obtain the desired cutting signals by removing noise or undesired signal frequency components. In addition to filtering, segments of the signal can also be generated when the cutter engages with the workpiece. These segments provide valuable information about the tool condition, as they capture the behavior of the tool during the cutting process [41]. Applying segmentation to every tool rotation will allow for the creation of repetitive segments. An overlapping time frame is then utilized to preserve data continuity [42].

The continuous collection of data can result in a large amount of high-dimensional data that require significant storage and computing power, which makes it a challenge when trying to extract meaningful information from the data for decision-making [43]. Therefore, this pre-determined representation serves as a lightweight description of the relevant variables in the cutting signals [44]. Feature processing can create new features that can lead to increased dimensionality and unnecessary computational burden [45]. The dimensionality reduction techniques are used to reduce the number of features or variables in the data while retaining the most relevant information. This is particularly useful in cases where the data have a high number of features or variables, as it can improve the efficiency and accuracy of the classification models.

3.2. Feature Extraction

Effective feature extraction and dimensionality reduction are key components of successful TCM systems [46]. Feature representation is a powerful tool for extracting hidden patterns of tool wear from cutting signals, from both the frequency and time–frequency domains [47]. Analyzing time as a feature domain is common and straightforward, and it can provide useful information about the signal amplitude corresponding to a certain time. Time-domain features can be derived from the raw signal by calculating statistical values such as the average, RMS, maximum, minimum, and peak-to-peak amplitude. Other time-domain statistical features that can be useful for signal analysis include the variance, skewness, and kurtosis of the signal. These statistical values provide information about the distribution of the signal amplitudes and can be used to identify patterns or trends in the signal over time. A compression of feature extraction from different domains [32] is listed in

Table 2. Comparison of feature extraction.

Features	Pros	Cons
Time-domain feature	Display different signals immediately and takes less time to process	Excessive noise
Frequency-domain feature	Suitable for steady-state systems	Not easy to identify the relevant frequency band
Time–frequency-domain feature	Suitable for non-steady-state systems	It does not have a standard procedure to select importantly features

.

The frequency-domain features provide a complementary perspective, especially when dealing with vibrational signals in TCM [48]. The concept of transforming time responses into frequency domains is a fundamental principle in signal processing. By doing so, it is possible to analyze the frequency content of a signal, which can provide information about its characteristics and behavior. One widely used method for transforming signals into the frequency domain is the Fourier transform. The fast Fourier transform (FFT) [49] is a widely used algorithm that computes the Fourier transform efficiently, making it a practical tool for signal analysis. By performing an FFT, the signal is decomposed into its constituent frequencies. This allows the identification of dominant frequencies, harmonics, and sub-harmonics that might correspond to specific machine vibrations, tool wear patterns, or faults [50]. Power spectral density (PSD) analysis is one of the frequency-domain features used to estimate the power distribution of a signal across different frequencies. Peaks in the PSD can indicate resonant frequencies or periodic disturbances in the machining process, often linked to specific wear or failure modes [51].

Other methods for transforming signals into the frequency domain include the discrete Fourier transform (DFT) [52], which is similar to the FFT but computes the transform differently, and improved versions of the FFT, which can provide a higher accuracy and speed for certain applications. In addition to frequency-domain analysis, time–frequency analysis is another important tool for signal processing. Time–frequency analysis is used to examine how the frequency content of a signal changes over time. The short-time Fourier transform (STFT) [53] is a widely used method for time–frequency analysis that provides information about the signal’s frequency content at different time intervals. Other methods for time–frequency analysis include the continuous wavelet transform [36], discrete wavelet transform [54], and empirical mode decomposition algorithms [55]. These methods can be used to extract features from signals such as the wavelet coefficient average energy and statistical values, which can provide valuable insights into the signal’s behavior [56]. Cepstrum analysis is another advanced technique where the inverse Fourier transform of the logarithm of the signal’s spectrum is calculated. It is particularly useful for detecting periodic components in the frequency domain, such as repetitive impacts or bearing defects, which can indirectly indicate tool wear [57].

3.3. Dimensionality Reduction

The performance of sensors is highly dependent on the specific application and environment in which they are being used [28]. The degradation of a tool can lead to changes in its physical properties, such as its geometry or material properties, which can, in turn, affect the accuracy and precision of the features it constructs [58]. Feature selection can reduce the number of features that need to be processed and analyzed, which can, in turn, reduce the computational complexity. This leads to a lack of certainty and limited generalization in the CNC machine TCM systems [59]. This is avoided by eliminating unwanted or redundant features [60]. To do this, feature normalization is applied, which is an important technique for pre-processing data in TCM systems and it links to the tool’s health directly [61]. Other methods such as the analysis of variance (ANOVA) and F-test are valuable tools for identifying features that are more related to cutting states and less related to tool conditions [45].

High-dimensional input data can be challenging for TCM systems, especially when the features are from different domains and may contain irrelevant or noisy elements [62]. In such cases, the classifier’s accuracy may be reduced, which can lead to overfitting, poor generalization, and low performance [63]. Applying dimensionality reduction techniques helps in remapping high-dimensional input data to a lower-dimensional space. By filtering out unique and dominating features, the resulting feature set can improve the accuracy and power efficiency of a local machine model for an online TCM system [64]. In multi-sensor TCM, dimensionality reduction techniques can help address issues by reducing the number of features in the dataset while retaining the most important information. By reducing the number of features, the epoch time and data space required for processing can be significantly optimized, leading to faster and more efficient data processing [65]. Additionally, removing noisy and irrelevant features can improve the accuracy of AI classifiers, as they can focus on the most relevant information [66].

Dimensionality reduction techniques are generally divided into two main categories: feature selection and feature transformation. Feature selection is the process of selecting a subset of the most relevant features from a larger set of features. The objective of feature selection is to reduce the dimensionality of the dataset while still retaining the most important information. By selecting only the most important features, feature selection can help to optimize computational power and improve the accuracy of the model. Feature selection typically sorts the features based on how sensitive they are to the target variable, in this case, the tool’s health state.

The most significant features are then selected for use in the model, while less important features are discarded. On the other hand, feature transformation techniques involve transforming the original features into a new set of features using mathematical transformations. The objective of feature transformation is also to reduce the dimensionality of the dataset while retaining the most important information. However, unlike feature selection, feature transformation techniques do not discard any features. Instead, they transform the original features into a lower-dimensional representation while preserving important information. The transformed features can be used as inputs for a model, but the original features cannot be directly interpreted from the transformed features [67]. The selection of a suitable feature transformation algorithm depends on various factors, such as the size, characteristics, and quality of the dataset. Different feature transformation algorithms may work better for different types of data, and it is important to select the most appropriate algorithm for a particular dataset [68].

The techniques used in feature extraction and dimensionality reduction are closely related to those used in data mining and ML in general. The challenges faced in feature extraction and dimensionality reduction in TCM systems represent important research topics in both ML and data science and provide opportunities for developing new algorithms and methods to enhance the efficiency of big data handling [67].

4. Advancements in ML and DL for TCM System

4.1. Machine Learning for TCM

Machine learning for TCM uses advanced statistical and computational techniques to predict equipment failure, optimize maintenance schedules, and improve overall production efficiency. The approaches for the ML-based TCM system structure are shown in Figure 4. The data collected from the cutting tool are, first, labeled and then normalized, with the noise removed and prepared for testing and training. By applying different ML models, the best classifier is identified. ML models for TCM have the potential to significantly improve machining efficiency and reduce workpiece damage by accurately categorizing tool conditions, estimating the remaining tool life, and predicting future tool failures [68]. The decision-making stage in ML can indeed be handled by classifier-based algorithms, particularly in situations where tool wear needs to be monitored progressively [7]. However, it is important to note that other types of ML algorithms can also be used for decision-making in this context, namely, artificial neural networks (ANNs), support vector machines (SVMs), Bayesian neural networks, k-nearest neighbor (KNN) models, decision trees (DTs), fuzzy logic (FL), and Gaussian process regression (GPR) [69,70,71,72,73,74,75].

Mainly, tree-based algorithms have been effectively applied to classifying cutting tool failures and improving tool life prediction. They can help identify important features and relationships between the features and the target variable, and they can be used to build interpretable models [76]. Bagging and boosting are two popular techniques used in ensemble learning, which involve combining multiple models to improve prediction accuracy. Bagging, or bootstrap aggregating, is used when the base algorithm is unstable or has a high variance. Boosting is used to improve the accuracy of a weak learner by iteratively building a sequence of models that correct the errors of the previous model [77]. Patange et al. [78] conducted a study comparing the performance of five different tree-based ML algorithms for predicting the wear of cutting tools in a machining process. The results of the study showed that the best-first tree algorithm achieved the highest accuracy among the five algorithms, with an accuracy of up to 90%. Pantage et al. [15] demonstrated that white-box tree-based models, such as random forest (RF) classifiers, can provide a superior performance for classification tasks in turning. These models offer greater transparency and interpretability compared to black-box models.

In situations where there is a large dataset for tool fault classification but only a small portion of it is labeled, semi-supervised learning (SSL) methods can be useful for enhancing the overall performance of the classification task. Peikarim et al. [79] proposed the use of an SSL method called self-training in conjunction with a pre-labeled dataset to classify a larger dataset of unlabeled data for tool fault classification. The study showed that the self-training SSL method was effective in improving the accuracy of the tool fault classification task compared to traditional supervised learning methods that use only labeled data. The use of SSL methods, such as clustering-based SSL, can provide an effective approach to dealing with large datasets with limited labeled samples [80].

The study by Tejas et al. provides evidence that swarm-based algorithms, and, in particular, the Harris hawks optimization algorithm, can be effective tools for solving complex optimization problems. However, it is important to note that the performance of these algorithms can depend on various factors, including the specific problem being solved, the parameter settings used, and the characteristics of the optimization landscape [15]. Both meta-heuristic optimization algorithms and neural network models have their strengths and weaknesses, and the choice of approach depends on the specific problem being solved and the available data and resources [79].

The ANN models have gained popularity as one of the most used ML approaches. These models are frequently applied to build prediction models for detecting stages of tool wear progression [32]. The concept of neural networks is inspired by the structure and function of the human brain. The human brain is composed of a large number of interconnected neurons that work together to process and analyze information, and neural networks attempt to simulate this process using mathematical models. In a typical neural network model, there are multiple layers of interconnected nodes, or neurons, that process input data and generate output predictions [31]. Neural networks are commonly used to analyze uncertainty, reveal hidden patterns, and handle big data in various domains, including TCM [81]. However, they come with drawbacks such as a slower epoch time, the presence of local minima, and the need for tuning multiple weights and biases. Addressing these challenges requires the careful consideration of computational resources, optimization techniques, and hyperparameter-tuning strategies to ensure effective model training and performance [82]. Data-driven models can be used to identify potential weaknesses or vulnerabilities in a system by analyzing data from different modes [83]. The decision selection conclusion was made from various classification outputs by referring to the different ML approaches which are listed in

Table 3. ML-based prediction model for TCM.

Authors	Signal	Feature Domain	Data-Processing Methods	Data Prediction Model	Prediction Accuracy
Salgado et al. [84]	Vibration	Frequency	Singular spectrum analysis (SSA)	ANN	RMSE: 15.11
Kilundu et al. [85]	Vibration	Frequency	SSA	ANN	67.4% accuracy
Miao et al. [86]	Vibration	Frequency	CNN	CNN	99.92% accuracy
Segreto et al. [87]	Force, AE, vibration	Frequency	Linear predictive analysis	ANN	98.9% accuracy
Seemuang et al. [88]	Sound	Time–frequency	STFT	Tested spindle noise at various feeds	_
Liu et al. [36]	Sound	Time–frequency	WPD	ANN	8.59% error
Salgado et al. [84]	Motor current, Sound FR	Time–frequency	SSA	LS-SVM	4.94–8.72% error
Tran et al. [53]	Cutting force	Time–frequency	Continuous WT	CNN	99.67% accuracy
Kothuru et al. [35]	Sound	Frequency	FFT	SVM	95.92% accuracy
Yao et al. [61]	Vibrations	Time, frequency, time–frequency	FFT	ANN based on FL	0.0003% MSE
Lu et al. [89]	Sound	Frequency	FFT	Hidden Markov model	91.8% accuracy

.

The probability of misclassification, which includes false negatives and false positives, is indeed an important aspect to consider when evaluating the accuracy and reliability of an ML model. The probability of misclassification can increase with the complexity of the machining process and the uncertainty of the working environment. This is where the role of an experienced human supervisor becomes crucial. In particular, when a TCM system falls into the failure zone, the experienced human operator can review the data on the dashboard and take appropriate action to resolve the problem. This can help ensure that the system continues to operate reliably and efficiently, even in the face of unexpected events or challenges.

4.2. Deep-Learning Models for TCM

Deep learning is a subset of ML that involves the use of ANN with multiple layers to model complex patterns in data. Compared with traditional ML approaches, DL models have a simple effective structure with multiple hidden features, as shown in Figure 5. With the recent advancement in computational power, DL has become an approachable and efficient way to solve classification problems. Unlike traditional ML approaches that require the selection and assignment of features, DL models can automatically map the relationship between input signals and output conditions by learning representative features within the original dataset in TCM systems; previous studies have shown that there is a strong relationship between force signals and tool wear. By collecting cutting signals from several sensors and using a DL network model with multiple layers, the model can learn patterns and generate representative features automatically within the original dataset [90].

This can outperform traditional ML approaches that require deep expertise and knowledge to analyze signals, and design and select important features within the dataset [24]. One of the significant challenges in applying DL models to TCM is the availability of large, labeled datasets that capture various tool states. Small datasets can limit the model’s ability to generalize and accurately predict tool conditions. However, several studies have proposed innovative approaches such as data augmentation and synthetic data generation, TL, few-shot learning, and semi-supervised and unsupervised learning to overcome this limitation [91].

Zhonghao Chang et al. proposed a novel weakly supervised anomaly detection network, feature map conversion and hypersphere transformation (FMC-HT), to address the challenges in fault prediction and health management in photovoltaic (PV) systems to overcome the traditional DL methods’ limitations due to the high cost of labeled data, and unsupervised methods fail to effectively utilize prior knowledge of anomalies. The model was trained through separate backpropagation processes to distinguish between normal and anomalous samples effectively. Testing on real PV EL datasets shows that FMC-HT performs impressively and maintains stability across varying levels of prior knowledge and anomaly rates. This approach offers a new solution for anomaly detection and demonstrates the potential of deep learning in scenarios with limited labeled data [92].

Research conducted by Yunhao Chen et al. has shown that data augmentation techniques can be effective in addressing the issue of small datasets. By artificially increasing the size of the dataset through transformations such as rotation, scaling, and adding noise, the model can learn more robust features [93]. Few-shot learning is designed to work well with small datasets by learning from just a few examples. Techniques like Siamese networks and prototypical networks have been applied to TCM with promising results. Work conducted by Peng Yang et al. employed a few-shot learning approach with Siamese networks to achieve the accurate pixel-wise target tracking of real-time objects, and his network with a two-level U-structure encode–decode model achieved 80 fps accuracy [94].

4.3. Transfer Learning Models for TCM

Transfer learning can reduce the computational cost of training a new model from scratch since the pre-trained model has already learned some of the underlying features and patterns of the data [95]. This model can leverage the knowledge learned from the pre-training, and the model can learn faster and more accurately on the new task, leading to more efficient and effective optimization [96]. Venkatesh et al. conducted a study to investigate the effectiveness of pre-trained DL neural networks for TCM. Specifically, he examined the performance of several popular DL architectures, including AlexNet, GoogLeNet, and ResNet-50, which had been pre-trained on large datasets [16]. The study presented by Haiyue Yu et al. [97] uses a novel method for predicting tool wear states and enhances signal visualization and DL feature extraction, proving effective in predicting various tool wear states. The proposed approach utilizes the Gramian angular field–Markov transition field–AlexNet (GAF-MTF-AlexNet) architecture. The method involves transforming the vibration signal’s envelope spectrum into a two-dimensional image using GAF and MTF transformations, then horizontally stitching these images to create a comprehensive dataset. This dataset is analyzed using a customized AlexNet model, significantly improving prediction accuracy to 96.83%. This approach can be particularly useful when dealing with limited amounts of training data or when the task at hand is similar to a task that the pre-trained model was originally trained on [7].

Nimel Sworna Ross et al. [98] ’s research on finding the best optimal TL model which detects the flank wear of the tool shows the challenges in predicting tool wear due to complex mechanisms and limited datasets under different operating conditions. The study evaluates several pre-trained networks, to determine the cutting tool’s state using the images of the cutting tool, where AlexNet showed a 93.6% accuracy in finding the cutting tool’s flank wear.

While TL offers several advantages, such as a reduced training time and enhanced neural network performance, it also has certain limitations including:

• Domain mismatch: Transfer learning assumes that the source and target domains are similar. If there is a significant difference in data distribution or feature space between the two domains, the model may not perform well in the target domain.
• Negative transfer: In some cases, knowledge transferred from the source domain can harm the model’s performance in the target domain. This occurs when the source task is not sufficiently related to the target task, leading to poor generalization.
• Limited data in target domain: While transfer learning can help with limited data in the target domain, it still requires a sufficient amount of labeled data to fine-tune the model effectively. If the target domain has too few labeled examples, the model may not learn the target task adequately.
• Overfitting to the source domain: If the pre-trained model is overly specialized to the source domain, it may overfit and fail to generalize to the target domain, especially if the target domain data are sparse or significantly different.
• Computational complexity: Fine-tuning a pre-trained model, especially large deep-learning models, can be computationally expensive and time-consuming, requiring significant resources for retraining.
• Interpretability: Transfer learning models, particularly deep neural networks, can be complex and difficult to interpret, making it challenging to understand why certain features are transferred and how they influence the target task.
• Dependence on source task quality: The effectiveness of transfer learning heavily depends on the quality and relevance of the source task. If the source model is not well-trained or the task is not closely related to the target task, the benefits of transfer learning may be minimal.
• Hyperparameter tuning: Fine-tuning a pre-trained model often requires the careful adjustment of hyperparameters, such as learning rate, batch size, and regularization methods. Poor tuning can lead to suboptimal performance or convergence issues [99].

There is significant potential for the development of more general and systematic TCM systems that can be applied across multiple smart factories. This will require ongoing research and development in areas such as ML, sensing technologies, and data analytics.

4.4. Long Short-Term Memory Networks

Long short-term memory (LSTM) networks are a type of RNNs that are particularly well-suited for tasks involving sequential data and time series, such as predicting tool wear in machining processes. LSTM networks are designed to overcome the limitations of traditional RNNs, such as the vanishing gradient problem, by incorporating memory cells that can maintain information over long periods. Due to their ability to model temporal dependencies and capture patterns in the data over time, LSTMs are widely employed in tool condition monitoring for predicting tool wear accurately and in real time.

In the research conducted by Niko Tursi c et al., a novel LSTM method is used for the analysis of the spindle current obtained during a manufacturing cycle and for assessing the tool wear range in real time with high accuracy [100]. The research conducted by Meng Liu et al. compared the ML, finite element analysis, and analytical model accuracy of cutting force during the machining process and found that the ML model possesses a high accuracy, in that the LSTM network possesses a 98% accuracy [101].

4.5. Scalability of ML Algorithms

Scalability is crucial for the successful deployment of ML algorithms in TCM systems, particularly in industrial environments where large volumes of data are generated from numerous devices. Distributed computing frameworks such as Apache Hadoop and Apache Spark can enhance scalability by enabling the parallel processing of large datasets. Cloud computing platforms like AWS, Azure, and Google Cloud offer a scalable infrastructure, dynamically allocating resources based on workload demands. Edge computing addresses latency issues by processing data locally at the edge devices, providing real-time insights, and reducing the need for constant data transmission to the cloud [102].

Integrating scalable ML algorithms into TCM systems is essential for maintaining efficiency and reliability in industrial settings. By leveraging distributed computing, cloud platforms, edge computing, model optimization techniques, and AutoML tools, TCM systems can handle increasing data volumes, provide real-time insights, and adapt to the evolving needs of industrial environments. This holistic approach ensures that TCM systems remain robust, scalable, and capable of delivering the accurate and timely status of TCM [103].

4.6. Comparative Analysis of Algorithms for TCM

Based on the previous research, the application of different ML algorithms used for TCM possesses unique performance accuracy. Depending on the performance metrics and requirements, these algorithms need to be chosen [104].

Table 4. Machine-learning algorithm performance metrics for tool condition monitoring.

Performance Metrics
Accuracy	The proportion of correctly predicted tool conditions to the total predictions
Computational efficiency	The time and resources required for training and inference
Robustness to noise	The algorithm’s ability to handle noisy or incomplete data
Real-time suitability	The feasibility of deploying the algorithm in real-time monitoring systems

and

Table 5. Comparative analysis of ML algorithm performance for TCM [105].

Algorithm	Accuracy	Computational Efficiency	Robustness to Noise	Real-Time Suitability
SVM	High (98%)	Moderate	High	Moderate
DT	Moderate (78%)	High	Moderate	High
RF	High (97%)	Moderate	High	Moderate
KNN	Moderate (73%)	Low	Low	Low
CNN	High (98%)	Low	High	Low
LSTM	High (99%)	Low	High	Low
Ensemble Learning	High (93%)	Moderate	High	Moderate
TL	High (94%)	High	Moderate	Moderate

show the performance metrics and comparative analysis of ML algorithm performance.

5. Industrial IoT and Its Application

5.1. IoT Structures for TCM

The rapid advancement in IoT technologies has unlocked tremendous potential within the industrial sector, leading to the emergence of smart factories and smart manufacturing environments. These innovations are transforming traditional manufacturing processes into highly automated, efficient, and interconnected systems. The development of IoT technologies has brought about significant opportunities for the industrial sector, including smart factories and smart manufacturing. Figure 6 shows the general four-layer IIoT architecture that forms the backbone of these smart industrial applications. This architecture is designed to ensure the seamless integration of various components, enabling real-time monitoring, data-driven decision-making, and automation. It consists of a sensory layer with sensing devices: this foundational layer is responsible for gathering data from the physical environment using a variety of sensing devices. These sensors can measure parameters such as temperature, pressure, vibration, and machine performance metrics. The data acquisition layer plays a critical role in providing accurate, real-time information about the operational status of the machinery and processes. Once the data are captured, they must be transmitted to a central location for processing. The network layer is responsible for securely conveying this information from the sensors to the cloud or local servers. This layer employs communication protocols such as Wi-Fi, Ethernet, 5G, or specialized industrial communication standards like message queuing telemetry transport (MQTT) to ensure reliable and low-latency data transfer.

Upon reaching the cloud or an edge server, the data enter the processing layer. This layer is tasked with storing, aggregating, and analyzing the data. Advanced analytics, including ML algorithms, are employed to extract meaningful insights from the raw data. The processing layer serves as the brain of the IIoT system, converting data into actionable intelligence that can predict maintenance needs, optimize operations, and detect anomalies. The final layer, the application layer, leverages the insights generated in the processing layer to drive decision-making and control actions. In a smart factory setting, this layer can execute real-time adjustments to machinery, optimize production schedules, or trigger maintenance alerts. By comparing the analyzed data with real-time sensor inputs, the application layer ensures that the factory operates within optimal parameters, creating a feedback loop that continuously improves efficiency and reduces downtime [106].

This interconnected structure, often referred to as a “cyber-physical system”, enables the factory to operate autonomously with minimal human intervention. The entire process, from data acquisition to application, is connected in a continuous loop, allowing the system to adapt dynamically to changing conditions. This capability not only enhances operational efficiency but also contributes to the creation of a truly automated and intelligent factory, where IoT technologies are at the core of innovation [107].

In recent years, researchers have made significant progress in implementing IoT-based online TCM systems using various approaches. The ML, DL, and IoT architecture have played critical roles in facilitating these advancements and have proven useful in monitoring machining stability and enhancing industrial processes [108]. These technologies enable real-time monitoring, predictive maintenance, and optimization, leading to improved productivity, efficiency, and cost-effectiveness in manufacturing and machining operations.

In recent times, new methods and ideas for IoT-based TCM systems have been proposed and developed by many researchers. Raja et al. [6] discussed the evolution of modern DAQ systems, highlighting the integration of IoT protocols, increased data channels, and improved industrial network capabilities. These advancements have led to the development of IoT-enabled DAQ systems that enable the real-time transfer of data in organized data streams between different components. Such systems play a crucial role in industrial applications, facilitating real-time monitoring, analysis, and control for enhanced operational efficiency and decision-making. Teti et al. accumulated accomplishments in testing machining conditions and introduced four essential key elements of a TCM system. They also presented a design for a cloud platform dedicated to observing cutting tools in turning processes [109]. The proposal of a cloud-based platform for the selection of an ideal machining state by Tapoglou et al. has the potential to improve the efficiency and effectiveness of machining processes [110].

The proposal by Chen et al., the energy proficiency TCM framework, has the potential to make a significant impact on the field of industrial automation [111]. The framework developed by Saif et al., a computer-vision-based IoT framework, has the potential to improve the efficiency and reliability of machining systems [112]. The online computing method for forecasting cutting tool conditions proposed by Li is designed to efficiently handle the large amounts of data collected during machining processes [113]. Wu et al. have developed a fuzzy computing method for data-driven ML and online monitoring in cyber-physical manufacturing. The approach involves the use of remote sensors installed on CNC machines to collect large amounts of data, which are then processed using fuzzy logic techniques to extract information [114].

Peng et al. have proposed a detailed cloud computing migration strategy to reduce transmission latency from the cloud. The strategy involves migrating some or all of the computing resources to a cloud infrastructure to reduce latency [115]. The latency issues faced by cloud computing solutions due to network transmission and migration requirements paved the way for You et al. to propose a lightweight network model (the rectified linear unit and exponential linear unit) for TCM that utilizes multiple activation functions, data augmentation, and cloud edge collaboration to reduce the size of the system while retaining data quality [19]. The strategies mentioned are concentrating on enabling smart factories to achieve a greater efficiency, flexibility, and productivity while meeting the demands of real-time data processing and decision-making.

5.2. Interoperability of IoT Devices

In an industrial environment, different devices will be employed to collect real-time tool state data. Therefore, interoperability is essential for creating a unified TCM system that can seamlessly integrate different sensors and devices. The standardization of data formats and communication protocols is crucial for ensuring compatibility between devices. Utilizing common communication protocols like the MQTT, Open Platform Communications Unified Architecture (OPC-UA), and Hypertext Transfer Protocol (HTTP) facilitates efficient data exchange and integration. Middleware solutions can act as intermediaries, handling data translation and communication to ensure different devices work together harmoniously. Achieving interoperability enables data fusion, combining information from multiple sensors to provide a comprehensive view of the tool condition, thereby improving the accuracy and reliability of tool wear predictions [116].

5.3. Edge and Fog Computing

Edge and fog computing is pivotal in modernizing TCM systems, particularly in the context of Future Industry. These computing paradigms bring computation and data storage closer to the data sources, such as sensors and machines, thereby reducing latency and bandwidth usage and also being applied in maintenance strategies. Hui Xiao et al. implemented edge computing devices in the cold rolling mill process and achieved more effective data, and his model showed a low latency in the sensor data [117].

Fog computing extends the cloud closer to the edge, creating a distributed computing infrastructure. In TCM, fog nodes can aggregate data from multiple edge devices, perform intermediate processing, and forward the refined data to the cloud. This approach balances the load between the local and centralized resources [118]. The study conducted by Ruoyu Liao et al. provides a novel maintenance strategy that goes beyond traditional physical-condition-based methods by incorporating operational risks into the maintenance decision-making process. This approach is shown to enhance the reliability of manufacturing systems while reducing the costs associated with equipment failures, quality issues, and production delays [119].

5.4. Possibilities of Industrial IoT Application

The purpose of this section is to utilize the cutting tool to its maximum potential and to prevent failures. The described system in the article [8] utilizes multi-sensor signals to predict tool wear during the cutting process. The signals are processed through hidden layers in an ML model. To visualize and control tool conditions, a web-based platform called “CONTACT Elements for IoT” is employed, leveraging a digital twin concept. This platform integrates with the MQTT protocol to present meaningful information on a graphical dashboard for further signal-processing tasks. The system can handle problems like cyber-attacks, sensor malfunctions, temperature, humidity, or noisy signals. It presents warning signs on the dashboard and automatically switches to a backup broker, ensuring the reliable and safe continuation of the processes.

The raw data collected from the sensors are pre-processed using signal-processing techniques. After that, significant features are generated from the pre-processed signals, such as statistical indices in the time–frequency domain. These features are used to recognize the current cutting condition. The sensor node is capable of transmitting data to the SQL server, which has a database system implemented for data management. The transmission of data from the sensor node to the SQL server takes place using a wireless connection. To enable communication between the sensor node and the cloud server, the MQTT protocol is used. This message protocol is particularly suitable for transferring telemetry data in areas with limited network capacity. By using MQTT, the architecture ensures that the data from the sensor nodes is transmitted efficiently and reliably to the cloud server for further processing and analysis. This allows for the real-time monitoring of the cutting condition and enables timely decision-making to ensure the integrity and longevity of the tools [120].

Once the data are processed and analyzed, they can be visualized through a graphical dashboard using CONTACT Elements with standard MQTT protocols. The dashboard provides a clear and intuitive representation of the data, enabling the user to easily monitor the tool wear and make informed decisions. However, accurate tool wear monitoring is a challenging task, and the system needs to be designed to handle various issues, such as cyberattacks, and inefficient sensor readings due to environmental factors like temperature, humidity, and noise signals. To ensure the system’s security, appropriate measures must be implemented, such as access controls, encryption, and authentication. Additionally, the system must be designed to handle sensor malfunctions and redundant sensors can be used to ensure accurate readings. In case of any issues, the system should be capable of alerting the user and taking appropriate actions, such as switching to a backup broker to ensure continuous operation [111].

Setiawan et al. developed a cutting tool condition monitoring system that aligns with the Industry 4.0 paradigm. The TCM system incorporates three sensors, namely, temperature, vibration, and cutting force sensors, which collectively gather data during the machining process. These data are then processed using ML algorithms to predict the condition of the cutting tool and estimate its remaining useful life. The system also includes a user interface (UI) that provides real-time information on the tool’s wear condition and a projection of its remaining life [121]. The data are then processed to reveal the relationship between cutting tool uncertainty and temperature variety, vibration strength, and power usage [120].

Cao et al. presented a method that goes beyond error detection and failure prediction in the manufacturing process. Their approach involved collecting data that could offer valuable insights into the characteristics of the monitored process [106]. The DAQ system proposed by Bajaj et al. is a low-cost and open-source solution for condition monitoring. It utilizes the ADXL335 accelerometer sensor to measure vibration signals, an Arduino Mega microcontroller to process and transmit the data, and Microsoft Excel for data analysis and visualization. The Parallax-DAQ IoT platform is used to transfer the data to the cloud for remote monitoring and analysis. This system can be easily replicated and customized for different monitoring applications [17]. The Parallax DAQ is an open-source platform designed to facilitate the transfer of data from the Arduino digital output to Microsoft Excel. With this setup, it is possible to continuously receive data from up to 26 microcontrollers and store it directly into an Excel datasheet. This platform provides researchers with a quick and simple setup process, utilizing commonly available hardware and software. Additionally, it offers customizable features that can be adapted to real-time conditions, allowing for greater flexibility in data monitoring and analysis.

In Figure 7, an intelligent TCM system is presented, utilizing a cyber-physical approach with three interconnected layers. In the first layer, sensors gather operational data from the machine, tool, and workpiece. This information is combined with additional data obtained from the manufacturing environment and the experience of machine manufacturers. The intermediate layer, known as the cyber-physical interface, stores and analyzes the collected machining dataset. Various techniques, including big data algorithms and machine-learning models, are employed to process the data, understand cutting processes, and learn from manufacturers’ experiences. The third layer called the cyberspace layer, is responsible for decision-making regarding machine failure prediction and maintenance. It generates machine failure models through the mining of sensor data. Ontologies form the basis for constructing knowledge related to machine health, while ontological reasoning approaches are utilized to estimate machine damage, deterioration, and future maintenance requirements. Finally, the results from the model are transferred back to the physical space, enabling prompt decision-making by the manufacturer. This cyber-physical TCM system provides valuable insights, aids in machine health assessment, and assists in making informed decisions to optimize maintenance and operations.

Recent research in the fields of cyber-physical systems and TCM is playing a significant role in shaping the future of smart manufacturing. However, the existing IoT architecture and network infrastructure for TCM are facing growing challenges in terms of connectivity, reliability, and systematic frameworks. This is primarily due to the increasing demand for more advanced ML techniques, a robust data pipeline, low-latency cloud computing, and expanding connectivity across multiple factories. These emerging trends indicate the need for further research and development efforts. One possible direction is to design a specifically tailored network infrastructure that caters to the requirements of TCM implementation. This would involve developing solutions that enhance connectivity, reliability, and real-time data processing capabilities. Alternatively, researchers may explore novel approaches to strengthen the existing infrastructure to meet the evolving needs of TCM. This could involve implementing new technologies, protocols, or frameworks to address the challenges related to connectivity, data management, and system integration.

5.5. Virtual Machining and Its Application

Virtual machining is a key aspect of modern manufacturing, providing a digital simulation of the machining process. This technology enables manufacturers to model and analyze the machining operations without physical trials, thereby saving time and resources. Virtual machining involves creating a digital twin of the machining environment, including the workpiece, tools, and machine tool dynamics [122]. Through simulation, manufacturers can predict outcomes such as tool wear, surface finish, and potential defects. This allows for the optimization of machining parameters, identification of potential issues, and enhancement of overall process efficiency. By integrating virtual machines with real-time data and advanced machine-learning models, it is possible to create a highly adaptive and efficient manufacturing process [123].

The advantages of virtual machining in TCM include:

• Simulation of tool wear: Virtual manufacturing models can simulate the wear and tear of cutting tools over time, helping to predict when a tool might fail or require maintenance. This helps in planning maintenance schedules and reducing unexpected downtimes.
• Process optimization: By using virtual simulations, manufacturers can optimize machining parameters (e.g., speed, feed rate, and depth of cut) to minimize tool wear and enhance tool performance. This reduces the need for costly and time-consuming physical trials.
• Testing new tool designs: Virtual manufacturing enables the testing of new tool designs in a simulated environment before they are produced and used in actual machining operations. This helps in refining the design for better performance and longevity.
• Cost reduction: Since virtual manufacturing relies on simulations, it reduces the need for expensive physical experiments and prototypes, saving time and resources.
• Predictive maintenance: By integrating virtual manufacturing with predictive maintenance strategies, manufacturers can use data from simulations to anticipate tool failures and schedule maintenance proactively, thereby extending tool life and improving efficiency.
• Real-time monitoring: Advanced virtual manufacturing systems can be integrated with real-time data from sensors and IoT devices to continuously monitor tool conditions and adjust machining processes dynamically to prevent excessive wear [124].

Based on the literature, we emphasize virtual machining as a promising approach due to its ability to provide insights into the machining process, optimize tool paths, and predict tool wear and failure. The use of virtual machining helps bridge the gap between theoretical models and practical applications, making it a critical component of TCM in Industry 4.0.

6. Challenges and Prospects

6.1. Challenges

Based on the literature, a comprehensive and efficient approach to tool condition monitoring systems integrated with data analytics and an IoT experimental setup was made and its accuracy and compatibility in acquiring real-time tool wear data was tested as shown in Figure 8, and found that achieving said vision requires advancements in emerging CNC machines and TCM technology, both current and future.

Additionally, the MQTT protocol setup is used to overcome the limited network capacity issue as shown in Figure 9, and its performance was more reliable than the traditional method. Creating industry-ready TCM systems for CNC machines presents several formidable challenges. These obstacles encompass a wide range of technological, operational, and economic factors that need careful consideration. One crucial challenge lies in the seamless integration of sensors into standard industrial CNC machines. This requires engineering solutions that do not disrupt the CNC’s core functionality while providing accurate and real-time data on tool conditions. Achieving this integration without compromising the CNC machine’s efficiency and safety is paramount.

Data accuracy and completeness are vital aspects of TCM, as any inaccuracies or gaps can lead to erroneous decisions and potential machine failures. Future research should focus more on enhancing data-processing methods to guarantee the dependability of TCM systems when deployed in actual manufacturing settings. Finding the right balance between efficiency and accuracy is another significant challenge. TCM systems should deliver timely alerts about tool wear or damage while avoiding false alarms. This necessitates the development of sophisticated ML models tailored to the specific needs of CNC turning centers. Generalizing ML models to different workspaces and conditions is essential for the widespread adoption of TCM for CNC machining.

Building online and offline infrastructure for shared databases is crucial for collaborative TCM efforts. This infrastructure should enable the efficient storage, retrieval, and analysis of TCM data, facilitating knowledge sharing among manufacturers. Latency issues in cloud computing must be addressed to enable real-time CNC machine TCM. Minimizing delays in data processing and decision-making is vital for preventing costly machine downtime. The rising expenses associated with TCM system development can hinder its widespread adoption. Therefore, further research should focus on cost-effective solutions and explore the economic aspects of TCM systems to ensure their practicality for mass implementation in smart factories.

In overcoming these challenges, future research should prioritize several key areas. This includes the development of custom-designed network infrastructure tailored to CNC turning center TCM systems, the exploration of diverse machine-learning models for TCM, and the enhancement of data handling and transfer capabilities for high-dimensional data. Additionally, constructing more versatile and general-purpose TCM systems, strengthening cybersecurity defenses, and incorporating knowledge of TCM models based on material science and mechanics are essential steps forward. Ultimately, a holistic approach that considers both technological advancements and economic viability will be crucial for the successful integration of TCM into the CNC machining industry, enabling smarter and more efficient manufacturing processes.

6.2. Industries and Their Products Contributing to TCM Technology

• Sandvik Coromant
  • CoroPlus Tool Guide and Tool Library: A digital solution that provides recommendations on cutting tools and tool assemblies. It assists in tool selection, improving efficiency and accuracy in the machining process [125].
  • CoroPlus ProcessControl: A real-time monitoring and control system that helps in optimizing machining processes. It collects and analyzes data from various sensors integrated into the machining tools, ensuring optimal performance and tool longevity [126].

2.

Kennametal

• NOVO: A cloud-based digital platform that provides tool recommendations and optimizations. It is designed to improve manufacturing efficiency by integrating tool data with CNC programming [127,128].
• ToolBOSS: An inventory management system that ensures the right tools are available when needed. It also tracks tool usage, helping to predict wear and schedule maintenance [129].

3.

IScar Metals and Tooling

• Tool Advisor: An intelligent tool selection system that provides recommendations based on the material, operation, and machine. It helps in reducing setup times and improving machining quality [130].
• Smart Factory Solutions: A suite of digital manufacturing tools that includes monitoring systems to track the performance and condition of cutting tools in real time [131].

4.

BIG KAISER Precision Tooling

• Electronic Wear Analyzer (EWA): A precision tool that monitors and adjusts cutting parameters automatically to maintain optimal tool conditions. It is used to reduce downtime and extend tool life [132].
• Digital Boring Heads: Equipped with digital readouts and connectivity features, these tools allow for precise adjustments and real-time monitoring, contributing to improved tool life and process stability [133].

5.

DMG MORI

• CELOS: An integrated platform that connects machines to the digital environment, enabling the real-time monitoring and control of machining processes. It provides data-driven insights to optimize tool performance and maintenance schedules [134].
• DMG MORI Tool Monitoring System (TMS): This system monitors tool wear and breakage in real time, allowing for immediate corrective actions. It is integrated with the machine’s control system to provide a seamless monitoring experience [135].

6.

Siemens

• MindSphere: An industrial IoT platform that connects products, plants, and systems to the digital world. It facilitates data-driven insights for tool condition monitoring and predictive maintenance [136].
• SINUMERIK Edge: A machine tool control system that integrates edge computing capabilities for the real-time monitoring and optimization of machining processes [137].

7.

Hexagon Manufacturing Intelligence

• SFx Asset Management: A cloud-based solution that provides the real-time monitoring of machine and tool conditions. It helps manufacturers optimize tool usage and reduce downtime [138].
• PC-DMIS: A software solution for dimensional measurement that can also be integrated into tool condition monitoring systems to ensure tools remain within tolerances [139].

8.

Marposs

• Tool Touch Verification (TTV): A system that monitors tool condition by verifying the tool’s geometry before and after machining. It helps in detecting wear and preventing tool failures.
• BLÚ: A modular monitoring system that collects data from multiple sensors in real time to provide insights into tool conditions and process stability [140].

9.

Zoller

• TMS Tool Management Solutions: A comprehensive software solution that tracks tool usage, wear, and inventory. It integrates with CNC machines to provide real-time monitoring and predictive maintenance capabilities.
• smartCheck: A tool inspection device that measures tool geometry and condition with a high precision, ensuring tools meet required specifications before use [141].

10.

Makino

• MPmax: A real-time machine and tool monitoring software that tracks tool performance, detects abnormalities, and provides predictive maintenance insights. It is designed to optimize tool life and reduce machine downtime [142].

These companies and their products represent the forefront of TCM technology, offering a wide range of solutions from real-time monitoring systems and predictive maintenance platforms to digital tools that integrate with smart manufacturing environments. The integration of the IoT, AI, and edge computing in these products reflects the industry’s shift towards more automated, data-driven approaches to tool condition monitoring and overall manufacturing efficiency.

6.3. Research Teams in the TCM Field

Research in tool condition monitoring (TCM) has been conducted globally, with significant contributions from research teams across various countries and regions. Here are some of the key countries and research teams in this field [143]:

• United States
  • Massachusetts Institute of Technology (MIT):
    Research Focus: MIT has been a leader in the development of advanced manufacturing technologies, including TCM. Their research includes the integration of the IoT and AI for real-time monitoring and predictive maintenance.
  • University of California, Berkeley:
    Research Focus: Berkeley’s research includes the development of machine-learning algorithms for predictive maintenance in manufacturing systems, including TCM.

2.

Germany

• Fraunhofer Institutes:
  Research Focus: Fraunhofer is a leading research organization in Germany that has made significant contributions to TCM, particularly in the development of sensor technologies and IoT-based monitoring systems.
• RWTH Aachen University:
  Research Focus: RWTH Aachen is known for its research in manufacturing technology, including advanced TCM systems that leverage AI and digital twins.

3.

Japan

• University of Tokyo:
  Research Focus: The University of Tokyo has been at the forefront of research in TCM, focusing on the integration of machine learning and IoT in manufacturing.
• Tokyo Institute of Technology:
  Research Focus: Tokyo Tech has made advancements in the application of AI for the real-time monitoring and control of machining processes.

4.

China

• Tsinghua University:
  Research Focus: Tsinghua has conducted extensive research on the integration of AI and big data in manufacturing, with a focus on predictive maintenance and TCM.
• Harbin Institute of Technology:
  Research Focus: This institution is known for its work on sensor fusion and machine-learning algorithms for TCM.

5.

South Korea

• Korea Advanced Institute of Science and Technology (KAIST):
  Research Focus: KAIST is a leader in smart manufacturing research, including the development of advanced TCM systems using the IoT and AI.

6.

United Kingdom

• University of Sheffield:
  Research Focus: The University of Sheffield is known for its research on advanced manufacturing technologies, including TCM and the use of AI for predictive maintenance.
• University of Nottingham:
  Research Focus: Nottingham has researched the application of machine learning for predictive maintenance in manufacturing, including TCM.

6.4. Future Trends

The evolution of TCM systems is driven by continuous advancements in technology and the ever-increasing demand for efficiency and precision in manufacturing. While significant progress has been made, several unresolved challenges and potential areas for innovation remain. The key future research directions that could shape the next generation of TCM systems are:

• The integration of edge/fog computing for real-time data processing which requires minimizing latency, optimizing bandwidth usage, and developing scalable edge computing solutions that can handle large volumes of data from numerous sensors.
• Obtaining small datasets in deep learning, because obtaining large labeled datasets for training deep-learning models remains a significant challenge in industrial settings.
• Interoperability and integration of IoT Devices with standardized protocols and interfaces for seamless integration in TCM systems has to be enhanced.
• The development of scalable machine-learning algorithms that can handle big data and complex sensor networks is essential.
• As an extension of the present review process, it is important to explore more on non-contact-type wear data acquisition methods like the acoustic emission and thermal imaging process
• The future industrial era will collaborate with humans to get the work done. Therefore, the product demand will also increase. To achieve this demand, an error-free production process is required. The onboard tool wear prediction system will help achieve zero-downtime production.
• The application of advanced sensor systems, such as AI-assisted sensors, is highly beneficial in meeting the requirements of Industry 4.0. Artificial intelligence methods represent cutting-edge technology and open new avenues in the context of Industry 5.0.
• Various AI methods, such as machine learning, deep learning, and artificial neural networks, will be incorporated into sensor design. These advancements will make CNC machine tool structures smarter compared to conventional machines.
• In the future, there is a need to develop low-cost, in-house sensor systems capable of smartly measuring machining responses at affordable prices.

7. Conclusions

In conclusion, this review brings out the necessary roles of TCM in contemporary manufacturing, which emphasizes its pivotal significance for CNC turning centers. TCM emerges not merely as a complementary aspect but as a crucial element in enhancing manufacturing performance by real-time monitoring, extending tool lifespan, minimizing downtime, and, ultimately, improving productivity. The integration of advanced technologies such as ML and IIoT into TCM systems is recognized for its transformative potential, offering intelligent online systems that learn and adapt to various manufacturing conditions. However, this review highlights the essential need for ongoing research to address challenges in sensor integration, data processing, cybersecurity, and economic feasibility for Industry 4.0 smart factories that are capable of autonomous decision-making based on real-time TCM data, fostering efficient and adaptive manufacturing processes.

Along with the above conclusions, the key findings of this review are:

• CNC Machining Operations: In this study, we found that CNC machines are critical in modern manufacturing, and the integrity of cutting tools is vital for the efficiency of these machines. Tool wear monitoring is essential to prevent machine downtime and ensure quality in production.
• Sensor Technologies: The review discusses various sensors like accelerometers, acoustic emission sensors, and cutting force sensors that are used to monitor the condition of tools. These sensors help in collecting real-time data that can be analyzed to predict tool wear and prevent failures.
• Signal-Processing Techniques: Advanced signal-processing techniques like fast Fourier transform, wavelet packet decomposition, and ensemble empirical mode decomposition are used to analyze signals from sensors. These techniques help in identifying patterns that correlate with tool wear.
• Machine Learning and IIoT Integration: The integration of industrial IoT and machine learning into TCM systems enable real-time monitoring and data analysis, providing operators with actionable insights to prevent tool failure and optimize machining processes.