Intelligent Recognition of Tool Wear with Artificial Intelligence Agent

Abstract

Tool wear, closely linked to operational efficiency and economic viability, must be detected and managed promptly to prevent significant losses. Traditional methods for tool wear detection, though somewhat effective, often lack precision and require extensive manual effort. Advancements in artificial intelligence (AI), especially through deep learning, have significantly progressed, providing enhanced performance when combined with tool wear management systems. Recent developments have seen a notable increase in the use of AI agents that utilise large language models (LLMs) for specific tasks, indicating a shift towards their integration into manufacturing processes. This paper provides a comprehensive review of the latest advancements in AI-driven tool wear recognition and explores the integration of AI agents in manufacturing. It highlights the LLMS and the various types of AI agents that enhance AI’s autonomous capabilities, discusses the potential benefits, and examines the challenges of this integrative approach. Finally, it outlines future research directions in this rapidly evolving field.

1. Introduction

The application of deep learning has witnessed a transformative evolution across various fields, fundamentally reshaping problem-solving approaches and solution development. According to the 23rd Shanghai CNC Machine Tool and Metal Processing Expo 2023 report, machine vision is progressively supplanting manual vision, with China’s machine vision replacement rate anticipated to reach 50% [1]. This paradigm shift has particularly reverberated in the domain of tool wear recognition in manufacturing and within the broader scope of brain-inspired artificial intelligence. AI-driven methodologies, spanning from traditional machine learning algorithms to sophisticated deep learning architectures, have not only facilitated advanced, real-time diagnostics but have also paved the way for the realisation of autonomous manufacturing systems. Large language models (LLMs), a notable advancement in this arena, have revolutionised language understanding and generation tasks, outperforming conventional rule-based approaches. LLMs exhibit proficiency in translation, summarisation, question answering, and creative content generation [2,3]. In the context of smart manufacturing, LLMs hold the potential to elevate human–computer interactions, simplify communication, and facilitate advanced decision-making processes [4,5]. First, by embedding LLMs into the production framework, the LLM’s analytical capabilities can predict mechanical failures, optimise operating parameters, and personalise machine behaviour like never before. These models serve as a middle layer that transforms complex machine data into actionable insights, effectively bridging the gap between human operators and mechanical processes. In addition, LLM integration can facilitate a more intuitive and interactive platform for factory personnel to use. Through natural language processing capabilities, LLMs can understand and execute commands based on common language input, making advanced computing technology accessible to people of all skill levels. This technology not only improves operational efficiency but also enhances the potential for collaboration between humans and machines, promoting a workspace where technology complements human skills. Tool wear is a key factor in machining operations, directly affecting product quality, production efficiency, and equipment life. Effective tool wear identification is critical to preventing unexpected machine failures, reducing downtime, and ensuring cost-effective manufacturing [6]. This study comprehensively reviews and analyses the history, status, and future prospects of brain-inspired artificial intelligence and intelligent identification methods, with a special focus on tool wear identification. By emphasising the critical role of AI in tool wear, this article emphasises the importance of accurate and timely identification. The conclusion summarises the current achievements and significant potential demonstrated by AI agents in fields such as mechanical design optimisation and tool wear recognition.

2. Brain-Inspired Artificial Intelligence

2.1. The Origins of Brain-Inspired Artificial Intelligence

The brain has been an important source of inspiration in the development of artificial intelligence (AI) [7]. From the Turing test to the modern renaissance of neural networks and deep learning, many of the key ideas in AI have been deeply influenced by neuroscience [8]. As early as the 1930s and 1940s, Alan Turing proposed the concept of a machine that could mimic “any human intelligence”, which undoubtedly inspired later researchers [9]. In the 1950s, Frank Rosenblatt invented the Perceptron, one of the earliest models of neural networks, whose design was clearly inspired by the workings of biological neurons [10]. The human brain consists of about 86 billion neurons that are interconnected through a complex network [11]. This observation led to the development of early neural networks in which “artificial neurons” mimicked the basic functions of biological neurons [12]. However, due to the limitations of computing power at the time, these early models were relatively simple. In the 1980s, the introduction of the backpropagation algorithm paved the way for the advancement of multi-layer neural networks [13]. While not directly inspired by biological neural networks, it proved to be a highly effective technique for training neural network models. This development eventually led to the emergence of deep learning, which is still undoubtedly inspired by the brain, even though the structure and training algorithms of deep learning models do not exactly mimic the biological brain [14]. With the remarkable success of deep learning models on a variety of tasks, such as image recognition, natural language processing, autonomous driving, and medical image analysis, the focus of research has gradually shifted to how to explain the mechanisms by which these models work [15,16,17,18,19]. This question has in turn inspired new research into the working mechanisms of the brain, as it is hoped that understanding the inner workings of neural networks will lead to a better understanding of the brain.

2.2. Brain-Inspired Artificial Intelligence in Visual Recognition

In recent years of artificial intelligence research, brain-inspired models have had notable applications in visual recognition [20]. The human visual system is a hierarchical information processing system. Raw pixel information received from the eye is first processed in the primary areas of the visual cortex, and then gradually passed to higher-level areas for more complex feature extraction and integration [21]. This hierarchical processing provided the inspiration for the design of convolutional neural networks [22]. A convolutional neural network (CNN) is a special type of deep neural network that mimics the visual processing mechanism of the human brain.

In CNNs, different layers are designed to extract features of varying complexity from the original image [23]. For example, lower layers may focus on extracting edges and textures, while higher layers may identify various parts of an object and the overall shape [24]. This hierarchical approach to feature extraction bears a striking resemblance to the visual processing mechanisms of the human brain.

A model that integrates an encoder and a decoder was developed for analysing tool wear, as illustrated in Figure 1 [25]. In this model, the encoder combines transformers and convolutional neural networks (CNNs) to analyse temporal data, while the decoder leverages a fully connected network to process these features. The model is designed to predict tool wear and variance, and only the labels of tool wear are available. In this case, the variance is obtained in an unsupervised manner. Thus, the model is trained using semi-supervised learning techniques and employs a negative log-likelihood loss function.

In addition to CNNs, recurrent neural networks (RNNs) have been introduced as brain-inspired models for visual recognition [20]. The most basic structure of an RNN contains a recurrent unit [26] that is responsible for receiving inputs and then generating outputs and new hidden states for the current time step based on the state (or hidden state) of the previous time step [27]. These hidden states are looped throughout the network, hence the name “recurrent neural network”. Unlike traditional feed-forward neural networks, RNNs have a “memory” mechanism [28] that allows them to save information from previous time steps and use it for computation at later time steps [29]. This makes RNNs especially suitable for processing time-dependent or sequence-dependent data [30], such as time series analysis [31], natural language processing [32], speech recognition [33], etc. By combining RNNs with CNNs, a higher level of visual recognition can be achieved in terms of understanding object dynamics and scene transitions [34]. As illustrated in Figure 2, the framework flowchart of the novel DLM architecture with a fully automated framework is presented. In the first stage, researchers developed an N-RBF-DRNN model, comprising three critical steps. Initially, vibration data and tool wear values were collected during CNC machine operations; vibration data were captured using triaxial accelerometers, and tool wear was recorded via imaging techniques. Subsequently, 217 data fields were extracted from the collected vibration data, encompassing time-domain, frequency-domain, machine parameters, and power spectral density (PSD) information. Finally, based on the extracted data fields, an N-RBF-DRNN model was constructed. This model consists of an input layer, an RBF layer, a recurrent layer, and several fully connected layers. It is capable of recursively processing wear predictions related to the previous cutting process while considering variations in initial conditions to effectively predict tool wear [35].

The success of brain-inspired models in visual tasks is well established. With the advent of increased computational power and the availability of large datasets, methods using CNNs and RNNs have achieved unprecedented success across various visual tasks. These include image classification, object detection, semantic segmentation, and 3D reconstruction. In many instances, systems employing CNNs have even surpassed human performance, further underscoring the effectiveness of these brain-mimicking visual processing techniques [20,36,37,38].

2.3. Brain-Inspired Artificial Intelligence in Manufacturing

In manufacturing, brain-inspired AI has demonstrated great potential, especially in quality control and predictive maintenance [39]. For example, by using brain-inspired CNNs for image analysis, factories can automatically identify and classify minute defects in products, ranging from cracks to irregular shapes [40]. This automated inspection not only greatly improves the efficiency of quality control, but also enhances product quality, thereby reducing rework and scrap rates [41]. Similar research involves analysing machine operating data to predict potential equipment failures [42], enabling factories to perform proactive maintenance. This approach significantly reduces unplanned downtime and associated costs [43].

Intelligent monitoring systems (IMSs) also play an important role in modern manufacturing. An IMS achieves real-time monitoring of production processes and equipment performance by integrating Internet of Things (IoT), AI, and big data technologies. These systems analyse large amounts of data to optimise production processes and improve efficiency and product quality while reducing downtime and maintenance costs through predictive maintenance. Looking to the future, with the development of edge computing, robotics, and augmented reality technology, IMSs are expected to further revolutionise the manufacturing industry, enhance automation and real-time decision-making capabilities, and promote industry efficiency and innovation [44].

In the context of Industry 4.0, the collaboration between AI and robotics has also had a transformative impact on the manufacturing industry. The synergy of AI and robotics increases automation, implements predictive maintenance, enhances quality control, and optimises supply chain management. AI algorithms enable machines to learn and adapt, improving manufacturing efficiency and adaptability. At the same time, integrating AI and robots introduces new methods for human–machine collaboration and process improvement, despite challenges like labour adaptation and standardising communication. As technology progresses, addressing these challenges effectively will be crucial for advancing Industry 4.0 and showcasing the potential of AI and robots working in tandem [45].

These use cases not only enhance the efficiency and quality of the manufacturing industry but also pave the way for broader adoption of brain-inspired AI in the future. Although these studies are still in their initial stages, they present numerous challenges and opportunities. With continued research and development, brain-inspired AI is expected to further revolutionise the manufacturing industry.

3. Intelligent Recognition of Tool Wear

3.1. Effects of Tool Wear

Tool wear is the process by which the surface of a cutting tool gradually loses its original shape and dimensions during machining. It is a natural phenomenon in which the tool surface experiences wear and fatigue due to friction between the material and the workpiece, high temperatures, and cutting forces, ultimately leading to a decline in tool performance [6]. Firstly, tool wear reduces cutting efficiency and extends machining time, while affecting the quality of the workpiece surface, which may lead to surface defects. Secondly, wear increases cutting forces and friction, raising machining energy consumption and increasing production costs. In addition, a shorter tool life means more frequent tool changes, further increasing maintenance costs. Wear can also cause tools to lose their original geometry, negatively impacting machining accuracy and dimensional control [46]. In manufacturing, reducing tool wear is crucial for enhancing productivity, lowering costs, and ensuring product quality. To address these challenges, solutions include using wear-resistant materials for tools, optimising cutting parameters, implementing effective lubrication and cooling systems, and employing advanced monitoring technologies for the real-time tracking of tool conditions [47].

3.2. Traditional Methods

In traditional manufacturing environments, tool wear identification typically relies on a range of tools, including direct visual observation [48], acoustic emission (AE) monitoring [49], cutting force measurement [50], vibration analysis [51], and infrared and temperature monitoring [52]. Direct observation is intuitive but imprecise and time-consuming [53], while acoustic emission and cutting force measurements require specialised equipment and expertise [54,55]. Vibration analysis requires specific sensors and data analysis software [56], while infrared and temperature monitoring are often used to infer cutting temperature changes due to tool wear [52]. Additionally, some manufacturers use rough estimates of wear by considering the life cycle and cost of the tool [57], which is used more for planned maintenance than real-time monitoring. Traditional methods, though effective in certain scenarios, often lack precision or necessitate extensive manual effort [58]. Consequently, contemporary manufacturing is moving towards advanced data-driven and algorithmic identification approaches.

3.3. Intelligent Identification Methods

In the early days of smart manufacturing and automation, tool wear identification relied primarily on traditional physical measurement methods. With advancements in computing power and data analytics, machine learning techniques began to be integrated into this field. Machine learning, a subset of artificial intelligence, builds AI-driven applications that enable computers to learn from data and make decisions or predictions. This technology has proven particularly effective in tool wear recognition, offering a more efficient and accurate solution to this longstanding challenge [59,60]. Initially, machine learning algorithms like support vector machines (SVMs), decision trees, random forests (RFs), and K-Nearest Neighbours (K-NN) were employed for classifying and predicting tool wear [61,62,63].

3.3.1. Support Vector Machines

Support vector machines (SVMs) have significant applications and effectiveness in the field of tool wear recognition. An SVM is a supervised learning algorithm mainly used for classification and regression tasks [64]. In the application of tool wear recognition, an SVM is primarily used to classify tool wear states based on data collected by various sensors (e.g., cutting force sensors, vibration sensors, or acoustic emission sensors). These states are usually classified into several categories such as “new tool”, “slightly worn”, and “severely worn”. One method to monitor micromilling tool wear uses an SVM model by combining vibration and sound signals [65]. This method selects parameters through the recursive feature elimination method and effectively identifies wear conditions with an accuracy as high as 97.54%. Another SVM-based machine learning model for online monitoring of metal cutting processes predicts cutting phenomena by analysing sound signals during machining [66]. This system demonstrates high prediction accuracy and proves effective as a monitoring tool. The model is not only efficient and economical but also features a stable sensing position that significantly reduces downtime and scrap rates.

A significant advantage of SVM is its ability to classify efficiently in high-dimensional spaces [67], which is particularly important for tool wear recognition, as this often involves multiple features and sensor readings. By finding an optimal hyperplane, SVMs can accurately separate tools with different wear states. In addition, the SVM’s kernel function allows for the algorithm to classify in a higher dimensional space, which makes it capable of capturing more complex wear patterns [68].

Another advantage of an SVM is that it suffers relatively little from overfitting, meaning that it maintains a high level of accuracy even in the face of unknown or new data [69]. This is particularly important for tool wear identification, as the conditions of use and wear patterns of tools may change frequently in real production environments.

However, a major challenge of SVMs is the need to manually select and tune model parameters (e.g., kernel functions and regularisation parameters), which can require significant experimentation and expertise [70]. Nevertheless, through parameter optimisation and feature selection, SVMs have achieved impressive results in tool wear identification and have become one of the important tools in this field.

3.3.2. Decision Trees and Random Forests

Decision trees and random forests have a wide range of applications in tool wear recognition. These two machine learning algorithms are mainly used for classification problems and are favoured for their high degree of interpretability [71].

The decision tree is a tree structure model that makes decisions based on a tree structure. As shown in Figure 3, each node represents a feature, each branch represents a decision rule, and each leaf node represents an output [72]. In the scenario of tool wear recognition, decision trees can be used to determine the wear state of a tool (e.g., “new tool”, “slightly worn”, or “severely worn”). One of the main advantages of decision trees is the simplicity of the model structure, which is easy to understand and interpret [73], and very valuable for engineers and operators. A study compared various methods and models to predict surface roughness, depth of cut, and tool temperature [74]. The classification and regression tree (CART) model performed best on all evaluation metrics, with a prediction error of only 1% to 3%, significantly better than the 17% to 28% error of the two-dimensional finite element model (2D FEM). Additionally, the analysis time for the CART model was shorter.

The random forest is a collection of multiple decision trees, each of which classifies or regresses the data and then takes a majority vote to determine the final output [76]. In tool wear recognition, the random forest algorithm is commonly used to improve the accuracy and robustness of the model. Since random forest contains multiple decision trees, it can handle noise and outliers more efficiently, thus providing more accurate wear state identification in complex and variable production environments. An evaluation of multiple machine learning algorithms on a dataset of 315 milling tests showed that, despite a longer training time, random forest achieved higher prediction accuracy compared to feed-forward back propagation ANNs and SVR with a single hidden layer [63].

Both algorithms can be effectively combined with other machine learning techniques to enhance recognition accuracy. However, each has its limitations; decision trees are susceptible to overfitting, while random forests require substantial computational resources [77]. Despite these challenges, decision trees and random forests are highly valuable in tool wear recognition due to their interpretability and accuracy. With careful feature selection and parameter tuning, these algorithms can accurately identify wear states in real production settings, thereby improving productivity and extending tool life.

3.3.3. K-Nearest Neighbours (K-NNs)

The K-Nearest Neighbours (K-NNs) methodology has been applied in the domain of smart identification of tool wear [61]. This algorithm is proficient in handling data of high dimensionality sourced from an array of sensors, including but not limited to, vibration sensors, acoustic emission detectors, and cutting force measurement devices. The data collected are then projected into a multi-dimensional feature space where the K-NN algorithm conducts a real-time assessment of the tool wear condition. Such real-time evaluations facilitate prompt maintenance actions or tool replacements, thereby minimising production downtime [61,78]. Moreover, the inherent flexibility of the K-NN algorithm enables its integration with other machine learning methods, such as decision tree classifiers or artificial neural networks [79]. This amalgamation can increase the accuracy of the tool wear state and improve the predictability of future wear for a wide range of tool types and machining materials.

For example, a proposed multilevel fusion algorithm integrates multiple BP neural network classifiers based on K-NN clustering analysis [80]. The algorithm generates the final recognition result by adaptively capturing the weights of individual BP neural network classifiers and using a linearly weighted fusion technique. Another approach involves an initial step of constructing a comprehensive model using least squares regression to capture the general behavior of the system, followed by the superimposition of a local model that utilizes the fuzzy K-NN smoothing algorithm to account for localized nonlinearities. Comparative studies have demonstrated that this hybrid incremental model offers superior error-based performance metrics compared to transductive and inductive neurofuzzy models [61]. Additionally, as shown in Figure 4, a multi-sensor fusion method utilises time–frequency statistical features and psychoacoustic parameters for WOA (feature weighting and selection) to support the classification of drill bit wear degrees [81]. By using the K-NN classification algorithm, accurate classification of five different wear levels was successfully achieved with 100% classification accuracy.

The primary strength of the K-NN algorithm lies in its straightforwardness and effortless implementation. Given that the algorithm fundamentally requires no training phase, it is exceptionally adaptable to new datasets, a characteristic highly beneficial in industrial settings where the conditions affecting tool wear can alter dynamically. Nonetheless, the K-NN algorithm is not without its limitations. One significant drawback of this method is its ‘greedy’ nature, which requires storing the entire training dataset for future predictions. This can be problematic in scenarios involving large-scale data [82]. Experimental comparisons of several common AI methods for tool wear recognition have shown that while K-NN demonstrates fast training times, its accuracy can be lower compared to other methods [83]. Additionally, the algorithm’s performance is highly contingent upon the scaling of features, requiring careful normalisation or standardisation procedures, and the selection of an optimal value for k and the appropriate distance metric poses a challenge that could significantly impact the model’s effectiveness [84].

Based on the applications of the three machine learning algorithms discussed above, a table summarising the advantages and disadvantages of each algorithm was created (see

Table 1. Summary of three machine learning algorithms in tool wear analysis.

Algorithm	Advantages	Disadvantages
SVMs (Support Vector Machines)	- In tool wear identification, it can effectively process multiple features and sensor data - It can capture complex wear patterns through kernel function technology - Significant classification efficiency, suitable for processing high-dimensional data	- Parameter tuning is complex and requires extensive expertise to ensure optimal performance - Training on large-scale datasets can be slow
Decision Trees and Random Forests	- The model is highly interpretable and easy for engineers and operators to understand and apply - Random forest improves the recognition accuracy and robustness in complex production environments by integrating multiple decision trees	- Single decision trees may be sensitive to noise and data anomalies and are prone to overfitting - Random forest requires a large amount of calculation and long training time in a big data environment
K-NN (K-Nearest Neighbours)	- Real-time tool wear assessment supports immediate maintenance decisions - Simple and suitable for multi-sensor data integration - Can flexibly integrate other algorithms to enhance predictability	- High storage requirements, especially inconvenient to implement when processing large-scale industrial data - Sensitive to feature scaling, requires precise distance metric selection

).

3.4. Deep Learning

Deep learning techniques have demonstrated considerable promise and applicability in the domain of automated tool wear detection. Unlike traditional machine learning approaches such as SVMs and decision trees, where manual feature extraction is often necessary, deep learning algorithms are capable of automated feature learning. This obviates the need for domain-specific expertise for manual feature selection, thereby reducing human intervention [85]. Additionally, the self-taught feature extraction mechanisms in deep learning algorithms enhance the model’s adaptability [86], enabling it to generalise across various tool types and degrees of wear. Furthermore, the expansive parameter space and layered architecture inherent to deep learning algorithms typically result in superior accuracy rates.

3.4.1. Convolutional Neural Networks (CNNs)

CNNs constitute a specialised category of deep learning architectures that excel in manipulating data arranged in grid-like formats, predominantly in the domain of image analytics [87]. These networks autonomously discern salient features from visual inputs through a sequential pipeline of convolutional layers, non-linear activation functions, and pooling operations. Subsequently, these networks employ fully connected layers for categorisation or various computational objectives [88,89]. This architectural design renders CNNs highly effective and precise for tasks such as image identification and other visual computational challenges.

For tool monitoring, high-definition images or scanned imprints of a tool can be analysed using a CNN model. The network is proficient in autonomously isolating and categorising critical attributes of the tool’s surface (for example, fissures, abrasions, areas of attrition, etc.) [90], thereby facilitating the assessment of the tool’s condition, the extent of its wear, and so forth. One proposed method for milling tool wear estimation uses infrared laser vision and deep learning algorithms, where tool wear images are processed by a multi-angle camera and an exposure fusion technique, followed by estimation using a multi-view convolutional neural network, which can improve machining efficiency [91]. Another proposed a CTNN (CNN–transformer neural network) model for the parallel processing of condition monitoring data, combining a transformer model and a convolutional neural network for accurate tool wear estimation while taking into account the stochastic uncertainty caused by data noise [25]. In milling experiments, the model hyperparameters are optimised and the results show that the model has excellent generalisation performance and high accuracy, reducing the need for complex feature engineering.

CNNs present notable merits in the domain of tool wear identification. Firstly, these networks inherently facilitate the extraction of high-level attributes from visual data, thereby diminishing the requirement for manual feature engineering [92]. Secondly, the hierarchical architecture of CNNs, coupled with their extensive parameterisation, contributes to their exceptional performance in recognition tasks. A proposed MFB-DCNN (Multi-Frequency-Band Deep Convolutional Neural Network) model for processing machining big data and monitoring tool conditions extracts features and predicts tool wear through three-layer wavelet packet decomposition and deep convolutional neural networks [93]. Moreover, the robust generalisation capabilities of a trained CNN model make it feasible to extend its application to diverse categories or manufacturers of cutting instruments [94]. Another proposed a CNN-based tool-wear-state monitoring method for aerospace manufacturing processes, dividing the tool wear level into different state modes and establishing a CNN state recognition model with sensitivity considerations to validate the applicability of the method in tool-state monitoring practices [95]. Additionally, given adequate computational infrastructure, real-time or quasi-real-time identification of tool wear is achievable, which is crucial for immediate quality assurance in manufacturing processes.

Nonetheless, there are constraints associated with the utilisation of CNNs. Primarily, these networks often necessitate substantial volumes of annotated data for effective training, which may pose limitations in settings where data are sparse [96]. Secondly, deeper CNN architectures increase computational costs, potentially raising the financial burden of implementing these models. Additionally, the inherent complexity of CNNs makes it difficult to understand how decisions are made, presenting challenges in situations where clarity is crucial [97]. Lastly, the potential for overfitting remains a concern if the training dataset lacks diversity or if the model’s architecture is overly intricate, thereby compromising its ability to generalise.

3.4.2. Recurrent Neural Networks (RNNs)

RNNs constitute a specialised architecture in the realm of neural computation, explicitly tailored for the analysis of sequential or temporal data. This feature renders them highly pertinent for tasks such as tool degradation detection. Nevertheless, the adoption of RNN architectures is not devoid of challenges. They are computationally intensive, prone to issues like gradient dissipation or gradient amplification, and often lack transparent model interpretability [98].

In order to solve the problems of traditional RNNs, LSTM was introduced which is a special type of RNN architecture with additional internal state and gating mechanisms for more efficiently capturing and managing long-term dependencies in sequential data. One study analysed cutting force signals in the unidirectional carbon-fibre-reinforced polymer (UD CFRP) using multi-channel 1D CNN and LSTM deep learning models, considering the effect of CFRP anisotropy on the force [99]. Experiments conducted under different milling conditions predicted tool wear. The effect of CFRP inhomogeneity was reduced by moving window and distribution analysis. The model predicted tool wear by the coefficient of determination (R2) with a mean average error (MAE) of 2.94 μm, improving accuracy by more than 2%. Additionally, a stacked LSTM encoder–decoder model for tool wear anomaly detection was developed [100]. It was trained using only normal data and was able to generate anomaly scores in real time. The model is suitable for production lines under rare anomalies and is an end-to-end learning architecture without preprocessing features.

3.4.3. Generative Adversarial Networks (GANs)

Generative Adversarial Networks (GANs) offer an innovative paradigm for the identification of tool wear conditions. GANs consist of two mutually antagonistic sub-networks: the generator and the discriminator, which function in a competitive manner. The primary objective of the generator is to simulate the various stages of tool degradation, synthesising data that closely mirrors the actual state of wear. This data can then be used as supplementary training material for other machine learning algorithms. Conversely, the discriminator serves a dual role: it validates the authenticity of the data generated by the generator and performs anomaly detection to identify instances that deviate from standard wear patterns. This iterative rivalry between the generator and the discriminator eventually propels the former to create increasingly precise approximations of real-world data [101,102].

In the context of benefits, GANs possess the capability for augmenting datasets, an asset that is especially pertinent when the volume of available data is limited. A study investigated multi-sensor data fusion in the frequency domain, where extracted features were used for tool wear classification [103]. To increase the amount of training data, a GAN was designed for data augmentation, while an early stopping strategy was employed to improve the effectiveness of the GAN. The experimental results showed that the proposed GAN significantly improved the performance of the tool wear classification model and reduced the amount of real industrial data required. Additionally, the versatility of GANs in processing multiple data modalities—ranging from visual imagery to auditory signals—renders them highly applicable for the multifaceted recognition of tool wear. Classifiers such as support vector machines (SVMs), random forests (RFs), decision trees (DTs), and Generalised Regression Neural Networks (GRNNs) trained on these augmented datasets have demonstrated classification accuracies approaching or equating to 100%, thereby outperforming methods based on the Synthetic Minority Over-sampling Technique (SMOTE) [104]. Another study researched and developed a Conditional Generative Adversarial Network (CGAN) for generating power signals associated with different process parameters. The experimental results showed that using CGAN data augmentation effectively reduced the surface roughness prediction error from 58% (the error when the CNN was trained without synthetic data) to 1.25% (with nine synthetic samples) [105].

3.4.4. Deep Reinforcement Learning (DRL)

Reinforcement learning (RL) is a machine learning algorithm in which a model (often called an “intelligent body”) learns how to perform a specific task by interacting with its environment. The intelligences are rewarded or punished for performing the task, and the goal is to find a strategy that maximises the cumulative reward [106]. In recent years, the combination of deep learning (specifically neural networks) has significantly advanced the development of reinforcement learning, giving rise to the subfield of deep reinforcement learning (DRL) [107]. Deep reinforcement learning uses deep neural networks to represent and optimise the strategies of an intelligent body, which is particularly effective in dealing with high-dimensional data and complex environments, most notably AlphaGo [108].

In tool wear recognition, the core principle of reinforcement learning is to optimise for a specific goal, such as extending the life of a tool or reducing the cost of production by interacting with the production environment. In this case, the intelligent system will make decisions based on collected data such as tool usage frequency, load, and other sensor readings to determine when to change tools or adjust machine parameters [109]. One proposed framework is the MLIL (Multi-Label Imitation Learning) framework for monitoring tool wear and bearing failures under different operating conditions [110]. MLIL transforms multi-labelled samples into imitation objects and employs a discriminator and deep reinforcement learning (DRL) to mimic the features of these imitation objects. DRL enhances a deep neural network by augmenting the deep neural network without the need for a reward function while using a discriminator to distinguish between the performance of DRL and that of the mimicking objects. To address the problem of difficulty in obtaining sufficient failure datasets when studying new types of tools or machining high-value parts, a DTRL network based on LSTM has been developed for extracting tool-state features from sensor data and dynamically adjusting the size of the training network [109]. LSTM is employed within the DTRL network to construct the value function. It smooths temporal information and captures long-term dependencies. A Q-function update and migration strategy is proposed to migrate the historical DRL network to the RUL prediction of new tools. As shown in Figure 5, the framework of another proposed method integrates tool wear prediction and cutting parameter optimization based on reinforcement learning to minimize energy consumption and production time [111]. The characteristics of processing data are extracted through a multi-source heterogeneous data fusion framework, and time-varying bias is introduced to eliminate the impact of machine aging, making the model prediction accuracy reach 96.09%. Additionally, a multi-constraint, multi-objective actor–critic reinforcement learning framework for dynamic process control has been proposed [112]. This method enables the RL agent to adaptively select cutting parameters by observing workpiece characteristics, processing requirements, and tool wear status. The test results show that the energy efficiency of this method is more than 20% higher than methods that ignore tool wear, the learning efficiency is three times that of the metaheuristic method, the online sampling time is less than 0.1 milliseconds, and it is suitable for real-time control.

3.4.5. Autoencoders

Autoencoders are often seen as part of deep learning, especially when they use deep neural networks as their encoder and decoder components. The main purpose of an auto-encoder is to reconstruct the input data by learning a compressed representation, which is usually achieved through a neural network [113].

Self-encoders are mainly used for feature extraction and dimensionality reduction in tool wear recognition. During the manufacturing process, the data collected are usually multidimensional and include various physical parameters of the tool, machine tool status, and production environment variables. These high-dimensional data are not only difficult to manage but also contain a lot of redundant information. The self-encoder learns a lower dimensional representation of the data through its encoder part that captures the most important characteristics of the raw data, which helps in the subsequent tool wear identification [114]. One study addressed multicore GPR (Gaussian Process Regression) to overcome the limitations of the original GPR, namely the difficulty of capturing non-isomorphic data structures from multiple data sources or different data types. The method was developed by using fusion features learned from stacked multilayer denoising autoencoders, which highlighted the more useful structures in the force signal characterisation data, thus improving the prediction accuracy and maintaining the stability of the confidence intervals [115].

In an autoencoder, the encoder component compresses high-dimensional input data into a lower-dimensional latent representation. The decoder then attempts to reconstruct the original data from this compressed form. The primary goal during training is for the model to learn a latent representation that allows for efficient and accurate reconstruction of the input data. Once trained, this latent representation captures the essential features of the data and can be used as input for other machine learning algorithms, such as support vector machines or decision trees. This is useful in various applications, including the monitoring of tool wear in manufacturing, where the model helps identify the wear state of a tool based on its operational data. Another study utilised autoencoders and GRUs combined with multi-sensor signals for tool wear estimation and compared their performance with a conventional feed-forward neural network, taking into account the effect of memory-based hyperparameters [116]. Additionally, a deep learning method based on stacked sparse self-encoder (SSAE) and multi-sensor feature fusion for milling cutter wear prediction was proposed. The method extracts signal features from different domains, selects the best sensor features through correlation analysis, and inputs them into SSAE for deep feature learning. Finally, a wear prediction model is built using BPNN, and its performance is verified by multiple sets of experimental data [117].

Table 2. Summary of five deep learning algorithms in tool wear analysis.

Algorithm	Advantages	Disadvantages
CNNs (Convolutional Neural Networks)	- Efficiently identify and classify tool wear features in images - Excellent performance in high-resolution tool image analysis, accurately identifying cracks, wear, etc. - Good generalisation ability to different types of tools	- Deep networks have high computational requirements and real-time applications may be limited - Requires large amounts of annotated image data for effective training - The black-box nature of the model makes the decision-making process difficult to explain
RNNs (Recurrent Neural Networks)	- Suitable for processing time-dependent wear signals such as cutting force and vibration data - Can capture long-term dependencies in sequence data, helping to predict the future wear status of tools	- Easily affected by gradient disappearance or explosion, making training challenging - The model has low transparency, and it is difficult to explain its specific behaviour in predicting wear
GANs (Generative Adversarial Networks)	- Solve the problem of scarcity of real wear data by generating supplementary training data - The generated data can be used to train other models, improving overall recognition accuracy - Can achieve high classification accuracy	- Training is challenging, as it requires fine control to adjust the balance between the generator and discriminator - There is a risk of generating spurious wear patterns that may affect the accuracy of the final identification
DRL (Deep Reinforcement Learning)	- Can optimise specific goals such as extending tool life or reducing production costs - Dynamically adjust policies to changing tool usage by interacting with the production environment - Suitable for learning and decision-making from complex high-dimensional data	- The training process is complex and requires a large amount of interactive data and time - The model has low transparency, and the specific decision-making process of the strategy is difficult to explain - Real-time applications have high resource requirements, which may affect the feasibility of deployment
Autoencoders	- Efficient dimensionality reduction and feature extraction of high-dimensional wear data - For complex multi-sensor data, key wear features can be automatically extracted, reducing the need for manual feature engineering	- Sensitive to data preprocessing and network parameter selection - The quality of autoencoder reconstruction depends on the network’s capabilities and may not be sufficient to capture all critical wear characteristics

summarises the advantages and disadvantages of each deep learning architecture in tool wear identification and provides a clear overview that can help select the appropriate method for specific applications (Table 2).

4. AI Agent

Artificial intelligence agents (AI agents) are software systems designed to simulate aspects of human intelligence, autonomously performing tasks, processing data, and making decisions without direct human intervention. Capable of interacting with their environment, they are typically used to automate complex, repetitive tasks or those requiring rapid decision-making, employing various AI techniques to enhance efficiency and effectiveness across applications.

AI agents typically consist of several key components that work together to simulate human decision-making processes and behaviours. These components include sensory units, processing units, actuator units, and learning mechanisms. Sensory units serve as the agent’s means of receiving information from the external world, with a variety that includes but is not limited to visual cameras, microphones, temperature sensors, and internet data streams. The data collected by these sensors provide the AI agent with the necessary raw information for further processing. The processing unit, the “brain” of the AI agent, is responsible for analysing and processing data received from the sensory units. This processing may involve data cleaning, feature extraction, and running machine learning models for prediction. The complexity of the processing unit can range from simple algorithms to advanced deep learning models. Actuator units are how the AI agent interacts with the external world and can include the arms of a physical robot, tools on an automated production line, or automatic execution modules within software. These units carry out physical or virtual actions based on instructions from the processing unit. The learning mechanism allows the AI agent to learn from experience and optimise its behaviour, often involving machine learning or deep learning technologies, enabling the agent to adapt to new environments and improve over time.

Currently, the field of AI agents has undergone a remarkable trajectory of progress, characterised by an unprecedented convergence of theoretical rigour and practical applicability. Leveraging advances in machine learning algorithms, large language models, and computer vision, AI agents have been transformed from mere rule-based systems to increasingly complex context-aware entities capable of autonomous decision-making [118]. The implementation of reinforcement learning and neural networks has given these intelligences the ability to learn from their environment, enabling a more dynamic and adaptive interaction paradigm. This has far-reaching implications across a variety of sectors, including healthcare, finance, and transport [119], where AI agents are increasingly being deployed to optimise processes, reduce operational inefficiencies, and improve the user experience. In addition, ethical considerations such as fairness, transparency, and accountability have begun to be systematically integrated into the design of AI agents, reflecting a mature understanding of their social and moral implications. However, it is important to note that this rapid development also calls for rigorous scrutiny to address challenges related to security, data privacy, and potential bias. As we stand at this juncture, the development of AI agents reflects not only a paradigm shift in technological capabilities but also in our conceptual understanding of agency, autonomy, and intelligence.

4.1. AI Agent with LLMs

Recently, there has been a burgeoning interest in large-scale language models (LLMs) like ChatGPT [120], marking a pivotal moment in the field of artificial intelligence. LLMs are a technology in the field of artificial intelligence that understands and generates language by learning from large amounts of textual data. LLMs learn the complexities and patterns of language by analysing a large number of linguistic examples and are thus able to generate text that looks like it was written by a human. As the training data increases, LLMs can better understand context and subtleties in language. However, the emerging concept of “LLM agents” signifies a paradigm shift, extending beyond the limitations of conventional large language models. While large language models primarily function as interactive platforms for text-based human–computer dialogues, LLM agents take on a more dynamic role [121].

LLM agents not only offer guidance but also actively engage in task execution. These agents serve as an intermediary layer, automating the repetitive interactions humans typically have with large language models such as GPT. Leveraging the advanced natural language understanding of LLMs, along with their goal-oriented capabilities, these agents can independently generate tasks, restructure task hierarchies, and methodically achieve objectives until the goal is attained.

In contrast to traditional AI models, AI agents aim to mimic the comprehensiveness and versatility of human intelligence and are theoretically capable of performing any cognitive task in any domain, whereas traditional AI models focus on performing specific, trained tasks, often lacking self-awareness and cross-domain adaptability [122]. When integrated with application programming interfaces (APIs), these agents acquire the capacity to perform a wide array of operations, such as internet navigation, application utilisation, file manipulation, and even automating government processes. To illustrate, governments can automate their processes by implementing API-based modern architectures, such as online payment systems, license applications, and citizen feedback mechanisms, to enhance service efficiency, transparency, and accessibility, while also addressing related security and privacy concerns [123].

AI agents guided by LLMs are intrinsically linked with elements of planning, memory, and tool utilisation. Planning is conceptualised as the decomposition of intricate tasks into more digestible sub-objectives. This entails iterative reflection, evaluation, and optimisation of previous actions to enhance the agent’s adaptability and the quality of the resulting outcomes. Memory functionality can be categorised into short-term memory for context-dependent learning and long-term memory which facilitates the indefinite retention and retrieval of information, often facilitated by external storage mediums. The utility of tools is manifested in the agent’s capability to employ external application programming interfaces (APIs) to supplement gaps in its pre-trained models.

When these three components are cohesively integrated, AI agents are capable of not only cognitive reasoning akin to human thought processes but also of executing actions that resemble human behaviour. In this regard, the ReAct (Reasoning and Acting) subsystem plays a pivotal role. It fuses the reasoning capabilities of a comprehensive model with behavioural determinants, thereby enabling the agent to make logically coherent plans based on its extant knowledge base [124,125].

The “Reflexition” architectural framework equips AI agents with the capacity for dynamic recollection and introspection. Instead of merely updating model weights, these agents are fine-tuned through linguistic feedback, thereby facilitating continuous improvement by learning from past actions and rectifying previous errors. In terms of memory management, AI agents aim to replicate the structure of human cognitive memory by implementing an efficient memory system that encompasses sensory memory, short-term memory, and long-term memory. These are represented as raw input learning (e.g., textual, visual data), context-aware learning, and external vector storage, respectively. Information stored in these memory modules is made available during user interactions, thereby enriching the conversational context [126]. A structured framework designed for LLM-based AI agents includes two different types of agents (one-step agents and sequential agents) to perform reasoning within this framework [127]. By instantiating the framework and evaluating the Task Planning and Tool Usage (TPTU) capabilities of different LLMs on multiple tasks, this study reveals the great potential of these models in handling complex tasks, while also pointing out areas for further research and improvement. One of the most defining human traits is the development and application of tools. AI agents emulate this by interfacing with external tools and APIs, thus enhancing their capacity to undertake more complex operations. Despite advancements, there remain unresolved challenges, including data governance and the optimisation of long-term memory systems.

4.2. Features and Capabilities of AI Agents

Artificial intelligence agents have multiple functions that can be combined with each other to form powerful and flexible artificial intelligence agents suitable for various tasks and application scenarios.

4.2.1. Learning and Adapting

Fully autonomous artificial agents need to have the ability to learn and adapt to themselves. The learning function mainly interacts with and analyses rich data and continuously accumulates knowledge and experience through supervised, unsupervised, or reinforcement learning. This enables the agent to recognise patterns, optimise strategies, and continuously improve its performance levels when faced with complex tasks. At the same time, adaptive capabilities serve as a supplement to learning capabilities and focus on automatically adjusting strategies in a changing environment to achieve real-time adaptability. This adaptability enables the agent to flexibly face new situations and unknown information, allowing it to quickly adjust decisions and behaviours to deal with new challenges. Learning and adaptive capabilities jointly build an artificial intelligence system with continuous learning and flexible adaptation capabilities. Through a survey of 103 participants, it was found that the autonomy of an AI agent may need to be reduced if the task requires a more formalised process and predictability to ensure stable task execution. Conversely, if the task is more dynamic and unpredictable, the AI agent may be granted a higher level of autonomy to enhance flexibility and adaptability [121]. A proposed framework called SOLA can continuously learn and adapt as changes occur in an open world and has been implemented on conversational robots [128].

4.2.2. Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) capabilities seamlessly integrate the benefits of retrieval and generation capabilities, leveraging the power of large language models (LLMs) to provide a comprehensive approach to user queries. Search functionality leverages the capabilities of LLMs to understand user queries and retrieve relevant information from large datasets or specific knowledge bases. The agent combines text understanding and information retrieval capabilities to precisely extract the pieces of information most relevant to the user’s request. In operation, the AI agent first analyses the semantic content of the query and then searches connected databases or internet resources to provide answers or relevant information [129]. Such capabilities are particularly valuable in areas such as law, healthcare, and scientific research, where rapid access to large amounts of information is critical. Retrieval functionality typically requires a framework that enables LLM-based agents to enhance their answers through retrieval operations. When the agent receives a question, it analyses the text and extracts relevant information from a vast embedded vector database [130]. The LLM then integrates this information to generate detailed and informative responses, enhancing the user’s understanding. For example, when looking for treatments that match the symptoms of a disease, an AI agent can use search capabilities to quickly browse medical databases to find relevant treatment options and provide them to the inquiring doctor or patient [131]. The core advantage of the retrieval function lies in its efficiency and accuracy, as well as the ability to process and understand large amounts of information in a short period of time and extract the most valuable data for the user.

On the other hand, generative functionality, another key component of RAG, specialises in content creation using large-scale language models. They use large-scale language models to generate text, ranging from writing plans to composing emails and even creating code. These agents connect to LLMs that enable them to generate coherent, contextual, and often novel content based on a given prompt or data input. Such systems are used for creative writing, generating reports, automating customer service responses, and more, demonstrating their versatility in a variety of areas requiring content creation [132]. A new approach to small-sample planning using LLM called LLM-Planner, enables embodied intelligences to perform complex tasks in a visual environment through natural language commands [133]. This method generates complete and high-quality plans, significantly improves sample efficiency and task versatility, demonstrates competitive performance on the ALFRED dataset with minimal training data, and provides a promising direction for the creation of intelligence capable of rapidly learning a variety of tasks. Another study used natural language to model complex human behaviour in a simulated environment, implementing generative agents through three memory types: observation, planning, and reflection [134]. The Phaser game framework is used to create interactive virtual environments where agents can perform tasks, move, and interact with objects. The server uses JSON to maintain information about each agent, including the location, activity, and objects with which it interacts. Agents are initialised with natural language descriptions, which are then parsed into actionable mnemonics. The environment and objects are represented in a tree structure that can be converted to natural language. Memory modules include observation memory (perception of the environment), reflection memory (thinking about past actions), and planning memory (future intentions). Although these memory modules make agents relatively plausible and believable in dynamic environments, there are limitations, such as incomplete or inaccurate memory retrieval and over-reasoning or reinforcement of memory.

4.2.3. Interaction and Communication

Interaction and communication functions are indispensable key features of AI agents, and their impact in terms of technology, user experience, and application scenarios has attracted much attention. Through the continuous advancement of speech recognition and natural language processing technologies, agents are able to achieve an accurate understanding of the user’s linguistic input, which provides an efficient and intuitive way for human–computer interactions [135]. This technological improvement not only makes agents more intelligent in performing tasks but also provides users with a more natural and intuitive experience. This is particularly evident in scenarios such as virtual assistants and voice assistants [136].

Interaction and communication features are crucial to enable personalised services. By interacting with users, agents can gradually understand users’ personal preferences and usage habits, thus providing services that are more personalised and meet users’ expectations [137]. This personalised service not only meets the user’s demand for individuality and customisation, but also enhances the user’s acceptance of the AI agent. Research in this area has deep applications in areas such as personalised recommendation systems [138].

In terms of technological innovation, the development of interactive and communicative functions has driven continuous innovation in areas such as speech recognition, affective computing, and natural language generation. By introducing affective computing, agents can better understand and respond to users’ emotional tendencies, making communication more human and emotional [139]. These technological innovations not only improve the accuracy and realism of agents in communication but also provide new possibilities for a wider range of future application scenarios and fields. Related studies have shown that affective computing plays an important role in human–computer interactions [140].

Finally, interaction and communication functions expand the application areas of AI agents. Agents adapt to different scenarios and needs by interacting with users, which makes the application of AI more extensive and the services more pervasive. For example, in the smart home, the agent achieves control of home devices through voice interaction, which enhances the convenience and comfort of life [141,142].

4.3. Application of AI Agents in Industry

The integration of LLMs into industrial applications demonstrates significant potential [143,144,145]. An innovative approach in mechanical design involves combining pre-trained LLMs with finite element method (FEM) modules, significantly enhancing the efficiency of structural optimisation [143]. Traditional mechanical design often relies on experts iteratively refining designs through experience and finite element analysis (FEA). While reliable, this method is time-consuming and heavily dependent on extensive prior knowledge and experience. The new approach leverages LLMs to generate initial design proposals, evaluates these designs using FEM, and provides feedback scores to guide the LLM in continuous learning and optimisation, without the need for domain-specific training data. This framework has shown remarkable efficacy in truss structure optimisation, where LLM agents can produce designs adhering to natural language specifications and iteratively improve these designs through prompt optimisation techniques. Experimental results indicate that LLM agents exhibit high adaptability in design generation and optimisation, autonomously developing and implementing effective design strategies, thereby showcasing significant potential in the field of engineering design automation.

Similarly, MechGPT is a fine-tuned LLM designed to bridge different scales, disciplines, and modalities, connecting knowledge in the domains of mechanical and materials modelling [144]. By extracting question–answer pairs from raw data and fine-tuning for specific knowledge areas such as multiscale material failure, MechGPT demonstrates exceptional capabilities in knowledge retrieval, language tasks, hypothesis generation, and interdisciplinary knowledge integration. The model can generate ontological knowledge graphs and execute multi-LLM interactive agent modelling, providing an explanatory framework and a visual representation of knowledge.

Furthermore, a multi-agent collaboration system named MechAgents has been proposed, combining LLMs with the FEM to tackle complex mechanical problems [145]. As shown in Figure 6, through dynamic interactions among multiple LLM agents, this system can generate new data, integrate interdisciplinary knowledge, and effectively solve classical mechanics problems, including elasticity and hyperelasticity, in mechanical design. Studies indicate that through self-correction within two-agent teams and the division of labour in multi-agent teams, this framework can automate the writing, execution, and correction of code to handle a variety of complex mechanical tasks.

Looking ahead, with further development of these technologies, AI agents are expected to play a crucial role in the field of tool wear recognition. Future AI agents, especially those based on LLMs, may possess stronger data parsing capabilities, capable of handling complex data from various sensors, including vibration signals, acoustic emissions, and temperature variations. Through multimodal data fusion, these agents can provide more accurate wear recognition results. AI agents can analyse sensor data in real time, provide instant wear state feedback, and perform predictive analysis to preemptively warn of maintenance needs, optimising maintenance schedules and reducing downtime. Furthermore, AI agents can continuously learn and optimise by acquiring new data, reducing the need for frequent human intervention and improving recognition accuracy and efficiency. This potential future approach could not only process and parse vast amounts of complex data to extract key wear information but also provide specific maintenance recommendations, enhancing production line efficiency and reliability. Additionally, AI agents can work in tandem with other automation systems to achieve fully automated tool wear monitoring and maintenance processes, boosting production efficiency and product quality. While traditional methods and conventional AI approaches still have their applications in tool wear recognition, with continuous technological advancements, LLM-based AI agents are poised to become a pivotal technology in this field, driving the development of intelligent manufacturing and automation.

In summary, these advancements collectively highlight the increasing importance of LLM-based agents in industrial applications, showcasing their potential in automating and optimising complex engineering tasks while also pointing to areas for further research and improvement. Through ongoing research and development, AI agents are expected to play a key role in more industrial domains, advancing intelligent manufacturing and automation technologies. Particularly in the field of mechanical design, the application of LLM-based agents has achieved notable success. LLMs not only excel in natural language processing but also address key challenges in mechanical engineering, for example, engineers can more efficiently generate design proposals, optimise parameters, and conduct complex simulations and analyses, significantly shortening development cycles and reducing costs. Multi-agent systems can coordinate multiple AI agents to collaboratively complete complex tasks, enhancing the overall system’s reliability and flexibility. These advancements not only introduce new methods and tools for mechanical design but also provide robust technical support for the intelligent transformation of the industrial sector.

4.4. Possible Future Ethical Issues

When artificial intelligence agents are used in conjunction with large-scale language models such as GPT-4, a few ethical issues may arise. Firstly, in terms of data privacy and confidentiality, models may collect sensitive information and face data security risks [146]. Second, these models may inherit social biases from the training data, which in turn may reflect these biases in outputs and decisions [147]. In addition, large-scale language models often operate as “black boxes” with a lack of transparency and interpretability in their decision-making processes, which makes the attribution of responsibility unclear [148]. In terms of the labour market, highly automated AI agents may replace human jobs and may affect people’s mental health [149]. From an autonomy and manipulation perspective, models may filter and present information in a biased way and even risk potential manipulation [150]. Finally, legal and regulatory issues cannot be ignored, especially when it comes to copyright, defamation, and geographical differences [151]. Therefore, addressing these ethical issues requires a combination of knowledge and methods from multiple disciplines, including technology, law, philosophy, and social sciences.

5. Conclusions

This paper underscores the revolutionary role of artificial intelligence (AI) in enhancing tool wear detection in manufacturing settings. These advanced algorithm technologies enhance predictive capabilities, minimise the need for extensive manual intervention, optimise production processes, and reduce downtime. Additionally, the incorporation of AI agents, synergised with large language models (LLMs), further augments the functionality of tool wear recognition systems. These AI agents autonomously process complex data, provide real-time feedback, and adapt to new information, substantially enhancing the efficiency and reliability of manufacturing operations. The potential applications of these technologies extend beyond tool wear recognition to various facets of intelligent manufacturing, indicating a promising future for fully automated and optimised production systems.

However, the implementation of these advanced AI technologies also presents challenges, particularly in terms of data privacy, ethical considerations, and the requirement for substantial computational resources. Addressing these challenges necessitates ongoing research and development, as well as a multidisciplinary approach encompassing technology, law, and social sciences. Ensuring transparency, fairness, and accountability in AI systems is crucial for their widespread acceptance and successful integration into industrial applications.

In conclusion, the integration of AI and LLM-based agents into tool wear recognition and other industrial processes represents a transformative development in the industry. As these technologies continue to evolve, they promise substantial improvements in productivity, efficiency, and overall manufacturing excellence. The broad applicability of AI agents across different sectors underscores the critical role of AI in shaping the future of industrial operations. Future research should focus on refining these technologies, addressing ethical and practical challenges, and exploring new applications to fully harness the power of AI in industry.