The Intelligent Monitoring Technology for Machining Thin-Walled Components: A Review

Abstract

Thin-walled components are extensively utilized in the aviation, aerospace, shipping, and nuclear energy industries due to their advantages of being lightweight and easily integrated. With an increased design quality and complexity of structures, thin-walled components have rendered traditional offline machining state prediction techniques inadequate for meeting the rising demands for machining quality. In recent years, advancements in intelligent manufacturing have led to the emergence of intelligent monitoring technologies that offer new solutions for enhancing the machining quality. This review categorizes technologies into online signal collection, state recognition, and intelligent decision-making, based on the implementation processes of intelligent monitoring. It summarizes the roles and current development status of various technologies within intelligent monitoring and outlines the existing challenges associated with each technology. Finally, the review discusses the challenges and future development trends of intelligent monitoring technology.

1. Introduction

Thin-walled components are widely used in aviation, aerospace, shipbuilding, nuclear energy, and other industrial fields due to their significant advantages for integration and light weight [1,2,3]. As the main process of thin-walled component formation, the machining quality directly affects the reliability and overall performance of products [4,5,6]. However, in the machining process, due to the characteristics of a large material removal, weak rigidity, and a complex force/thermal coupling effect between the tool and workpiece, the machining accuracy is significantly affected. As shown in Figure 1, the machining of thin-walled components is prone to deformation [7], chatter [8], tool wear [9], poor surface integrity [10,11], etc. The influencing factors include the machine tools [12], processing parameters [13], and initial stress field [14]. The traditional offline quality prediction and control method can improve the machining quality to a certain extent, but it cannot capture the dynamic change in the machining process online. It is difficult to meet the processing needs of more and more complex thin-walled components.

With the development of sensor and computer technology, such as big data and cloud computing, intelligent monitoring technology provides a new strategy to solve the problem of the quality of intelligent control with its functions of real-time monitoring, analysis, and control [15,16,17]. At the same time, the technology greatly improves the adjustability of the component processing process, and is an important part of the intelligent development of the manufacturing industry. This review summarizes the research literature on intelligent monitoring technology for machining processes in the last 15 years, as shown in Figure 2. The Web of Science database was utilized to retrieve the relevant literature on intelligent monitoring technology. The retrieval keywords were ‘machining’ and ‘intelligent monitoring’. Then, we counted the number of SCI papers published by the top-ranked countries based on the quantity of publications. From 2017 to now, the number of annual research publications on intelligent surveillance technology has increased from 344 to 2169, indicating that it has gained the attention of researchers in recent years. At the same time, comparing the number of intelligent monitoring technology papers between China and developed manufacturing countries (such as the United States, Germany, the United Kingdom, etc.), it can be found that the number of articles in China in 2023 accounted for 50% of the total literature. Chinese research attention to intelligent monitoring technology continues to rise, and it has become the development hotspot and research frontier of intelligent monitoring technology.

The implementation process of intelligent monitoring technology includes data acquisition, data processing, feature extraction, state recognition, and intelligent decision-making, as shown in Figure 3. Intelligent monitoring technology has shown great potential and practical application value in the field of thin-walled component processing. It measures various physical parameters (such as the cutting force, vibration, temperature, etc.) in the machining process through different sensors [18]. Then, efficient data processing algorithms are used to process these data [19], such as time–frequency domain transformation and other methods to extract key features reflecting the processing state [20]. On this basis, combined with machine learning, deep learning, and other methods, the abnormal state in the machining process can be automatically identified, such as tool wear [21], workpiece deformation [22], chatter, and surface integrity [23]. Finally, intelligent algorithms are used to make intelligent decisions on processing schemes, adjust processing parameters in a timely manner, or take preventive measures to achieve the accurate control of the machining process [24]. Lacalle et al. [25,26] prepared a data acquisition system that simultaneously obtained the cutting force and tool position. They established a relationship model between the geometric shape of the machined surface and the cutting force value. Artetxe et al. [27] proposed a cutting force prediction model that incorporated tool runout and radial meshing. Then, the surface roughness and morphology were obtained through a force wall restoration equilibrium.

This review aims to clarify and analyze the latest progress in intelligent monitoring technology in recent years, especially the key technologies and methods for the intelligent monitoring of thin-walled component processing. This paper summarizes the latest technologies and algorithms of intelligent monitoring technology in terms of data acquisition, processing, feature extraction, state recognition, and intelligent decision-making, and discusses in depth the specific contributions of these technologies to improving machining accuracy. In addition, this paper also points out the shortcomings of the current research and a possible development direction in the future, and provides a reference for researchers in related fields, aiming at promoting the further development and application of intelligent monitoring technology in the field of thin-walled component processing.

2. Online Signal Acquisition Technology

2.1. Signal Acquisition Technology

Signal acquisition technology is the basis of intelligent monitoring technology to perceive the physical state of the machining process. With the development of sensor equipment, the online measurement of a variety of physical signals in the machining process has been realized. The cutting force, cutting heat, vibration, and other physical qualities produced in this process will affect the workpiece machining quality and tool life. In essence, as long as the sensor is advanced enough, the physical qualities during the process can be measured. However, due to the generation mechanism of physical quality signals being different and the complex environment in the cutting machining area, most physical qualities are difficult to measure directly. At present, the physical qualities commonly obtained are the cutting force [28], cutting heat [29], acceleration [30], sound, internal power and current of the machine tool [31], etc.

The above physical qualities can be obtained by different sensors. Cutting force data are commonly collected by a dynamometer. Cutting heat data are commonly collected by a thermocouple or infrared thermometer. Acceleration is usually obtained by a triaxial acceleration sensor. Sound is usually obtained by a microphone. Zhang et al. [32] designed a force-sensing unit integrated into the tool holder to provide a solution for cutting tool condition monitoring during machining. Liu et al. [28] developed a tool wear monitoring system by establishing an online measurement system for the cutting force and cutting temperature. Han et al. [33] proposed a near-infrared fiber multispectral in situ online temperature measurement method to verify the real-time response of the temperature measurement system to the cutting state, as shown in Figure 4. Liu et al. [34] realized the online detection of flutter during machining by collecting the current signal of the feed motor online and conducting empirical mode decomposition processing. Rahimi et al. [35] collected vibration data during the machining process through a triaxial acceleration sensor and converted it into a short-time spectrum to realize the online detection of flutter and its frequency.

According to the statistics of the literature on the intelligent monitoring of machining processes in the past 5 years, as shown in Figure 5, the research literature on monitoring the tool wear and remaining life accounted for 36% of the total, the research literature on monitoring process sustainability accounted for 19%, the research literature on monitoring chatter and machining deformation accounted for 22%, the research literature on monitoring surface quality accounted for 9%, and the literature on monitoring other machining conditions accounted for 13%. Among them, most researchers believe that the cutting force is one of the best monitoring signals of the tool state and workpiece state during machining. Because of the wear and failure of the cutting tool, the flutter, deformation, and damage of the workpiece will be directly reflected in the change in the cutting force. However, high-precision force sensors are expensive, require special installation methods, and will affect the stiffness of the machine tool itself, and they are difficult to use in large quantities at industrial machining sites.

Acceleration sensors, which can monitor the vibration signal, have advantages in terms of the installation and cost. The vibration signal can accurately reflect the dynamic response via the change in the processing state. However, acceleration sensors are usually positioned in non-contact areas and take indirect measurements. At the same time, because the vibration signal is easily affected by the low-frequency signal generated by the machining path, machine tool, tool clamping structure, etc., they present higher requirements for a signal noise reduction processing method. Acoustic emission (AE) sensors can supersede the vibration sensors because the acoustic emission signals reflected have a higher frequency, which can eliminate the interference of low-frequency vibration. How-ever, the attenuation of acoustic emission signals is serious, which affects the integrity and reliability of the collected data.

Through the above analysis, it was found that each sensor has its own advantages and disadvantages. With the increasing requirements for online monitoring accuracy and robustness, the simultaneous acquisition and processing of multiple signals using multiple sensor fusion methods can improve the accuracy and reliability of processing state monitoring [36]. The multi-sensor fusion method can collect the two or more different sensors’ data simultaneously, which reduces the detection uncertainties that may arise from the single-sensor method. Ghosh et al. [37] combined the cutting force, spindle vibration, spindle current, and sound pressure to extract features to characterize the average flank wear of the main cutting edge of a tool. Sun et al. [38] extracted features from the cutting force, vibration, and AE signals to form a feature matrix, and realized the online evaluation of the remaining tool life. Zhu et al. [39] extracted multiple machining state features by collecting acceleration signals and fusing cutting force signals to realize the online monitoring of the tool state, as shown in Figure 6. Wang et al. [40] extracted multi-sensor signal features from sound, acceleration, and cutting moment signals, and simultaneously identified flutter and tool wear. Li et al. [41] proposed a decision-level fusion method across multi-domain features by collecting vibration, acoustic emission, and spindle power data, and realized the online monitoring of the remaining tool life. Hao et al. [42] proposed an adaptive denoising method based on the fusion of sound and acceleration signals to realize flutter recognition.

The sensor signal (especially the multi-source sensor signal) data of the machining process cover the parameters of force, temperature, acceleration, sound, and power, which constitute a large and complex dataset. This surge in the data volume undoubtedly brings unprecedented opportunities for the accurate monitoring and status recognition of the machining process, but at the same time, it also presents more stringent requirements for subsequent signal processing technology.

In addition, various non-signal-based monitoring technologies have been integrated into the manufacturing process, driven by advancements in sensor technology. The temperature significantly influences the machining quality. Bagavatiappan et al. [43] utilized infrared thermal imaging technology to monitor the temperature changes in cutting tools in real time during micro milling. Dit et al. [44] examined the effect of calamine on the temperature field as measured by infrared cameras and proposed a temperature monitoring method for processing based on this technology. Wang et al. [45] introduced an online temperature monitoring approach for thermocouples that employs a gradient boosting decision tree (GBDT) algorithm, enhancing the accuracy of cold end compensation and reducing nonlinear errors in the online dimensional monitoring process of thermocouples. Zhou et al. developed a monitoring model leveraging deep learning and computer vision, which successfully achieved the online monitoring of the tool wear status.

2.2. Signal Processing Technology

Faced with the huge amount of multi-source sensor signal data, the traditional signal processing methods have shortcomings in efficiency, feature extraction accuracy, and recognition ability. Therefore, efficient data processing technology is urgently needed to realize the rapid processing and deep mining of massive data. In order to accurately capture the key information in the signal and effectively eliminate the noise interference, it is also necessary to have a strong generalization ability to adapt to various complex conditions and abnormal states that may occur in the machining process. Simultaneously, the raw data of physical qualities collected by sensors during the processing may be affected by factors such as noise, errors, and missing values, resulting in a decline in signal quality. Signal processing technology can perform data cleaning and pre-processing, remove outliers, fill in missing values, and correct erroneous data to ensure data accuracy and reliability.

Commonly used signal processing methods include filter denoising, correlation analysis, the Fourier frequency transform (FFT) [46] and short-time Fourier frequency transform (STFT) [47], wavelet analysis [48], the Hilbert–Huang transform (HHT) [49], the Wigner distribution function (WDF) [50], etc. For example, Zheng et al. [51], Chen et al. [52], and Ding et al. [53] used FFT analysis to process acceleration signals and extract flutter features, as shown in Figure 7. Joseph et al. [54] used a discrete wavelet transform and the HHT to process non-stationary sound signals to highlight the applicability of tool sound signals in the turning process. Matthew et al. [55] used wavelet synchronous compression transform (WSST) and HHT methods to process the force and acceleration signals, which enhanced the robustness of the flutter detection algorithm.

At present, multi-source heterogeneous data acquired by multi-sensors require data fusion technology to integrate data and ensure their reliability. Data fusion technology aims to integrate large-scale, heterogeneous, imprecise, inconsistent, and even conflicting data into accurate, reliable, consistent, and valuable data. Probabilistic fusion (such as Bayesian fusion [56]), belief reasoning fusion (such as Dempster–Shafer theory [57]), and rough set-based fusion [58] are traditional data fusion methods. Data fusion methods are necessary to ensure data quality. They also should reduce data uncertainty and enhance data manageability. From the perspective of data processing, data fusion is one of the key technologies to promote intelligent monitoring technology. It is a bridge that transmits perceptual data to a state recognition model and decision algorithm. For example, Macias et al. [59] designed a multi-sensor data fusion framework and established a data fusion evaluation system, as shown in Figure 8.

Current data fusion techniques are capable of processing heterogeneous data collected from multi-source sensors, such as physical quality data including the cutting force, cutting temperature, acceleration, and acoustic emission. After the fusion of these physical quality data, they prompt recognition models to map the current physical state. However, with the continually rising demands for processing accuracy, the current data fusion technologies still require further enhancement in terms of the precision and efficiency of processing multi-source data. Simple methods of normalization and weight coefficient assignment are insufficient to meet these requirements. Signal processing technology requires the unified characterization of multi-disciplinary and multi-dimensional data. In addition to relevant physical quality data, it is also necessary to integrate data derived from the physical environment, commonly used industrial software data, human–computer interaction data, etc., to ensure that the merged data have good scalability and compatibility.

3. Online Recognition Technology for Machining Status

3.1. Signal Feature Extraction

The machining signals collected through signal acquisition technology include machining status information such as the deformation, chatter, surface integrity, tool wear, and interference. Therefore, it is necessary to extract features corresponding to each processing state from the model data, as their effectiveness directly influences the reliability of the state recognition results. The feature extraction methods can be divided into time domain, frequency domain, and time–frequency domain methods.

A method incorporating both dimensionless and dimensional characteristic indicators is one of the time domain methods, which are based on time domain signals. For instance, Lamraoui et al. [60] extracted the peak value, mean, variance, and root mean square data, which are dimensional indicators. Meanwhile, the kurtosis, skewness, and waveform factor were extracted. They are dimensionless indicators. Ye et al. [61] proposed the milling chatter monitoring method, which transforms the acceleration signal into a square sequence and calculates the mean of this sequence. Additionally, Wan et al. [62] introduced an adaptive filter to process the original signal and proposed using the variance ratio of the filtered signal to the original signal as a characteristic indicator for monitoring chatter, as illustrated in Figure 9. A second method involves extracting data features after applying mathematical transformations to the time domain signal. It includes singular spectrum analysis [63], principal component analysis [64], and phase space reconstruction [65]. For instance, Wang et al. [66] established an auto-regressive moving average (ARMA) model utilizing the cutting force time domain signal from the milling process, employing the Q factor as a chatter index. Rafal et al. [67] elucidated the recursiveness of a dynamic system through phase space reconstruction. In summary, time domain signals can effectively extract features for processing state identification, offering advantages. They have the characteristics of simplicity, ease of understanding, and high computational efficiency. However, this method’s results could be less robust, as they are affected by interference from external signals. Additionally, given that cutting is a complex physical process, the relevant signals exhibit nonlinear characteristics. Consequently, accurately describing the signal with limited parameters in time domain analysis proves challenging.

The frequency domain method extracts features by conducting a spectrum analysis on time domain signals. Common spectrum features include the power density, power, phase, and amplitude [68]. In comparison to time domain analysis, the frequency domain method can filter out interference signals through time–frequency domain transformation and is frequently employed for the feature extraction of vibration and sound signals. A classic frequency analysis technique is the fast Fourier transform (FFT). The FFT is widely used in signal pre-processing, and the power spectral density (PSD) is also the same. This is particularly relevant as chatter is characterized by variations in the frequency content and energy distribution. It heightens the uncertainty of the signal [69]. Furthermore, spectra are often utilized to analyze signals, which ascertain the acquired data pertaining to stable cutting. Recent research has predominantly focused on enhancing frequency analysis through the development of new metrics. For instance, Jo et al. [70] extracted statistical features to identify chatter by calculating the sum of frequency components within the high-frequency band. Chang et al. [71] identified the occurrence of chatter without employing a threshold by examining vibration frequencies under varying cutting conditions. However, the basis function can resemble the global function when utilized in the FFT. Consequently, when processing nonlinear and non-stationary signals, it can only determine the presence of certain frequencies without providing insight into the timing of their occurrence.

The time–frequency domain processing method is designed to address the issue of the non-coexistence of time and frequency information in frequency domain methods. It can more accurately extract the non-stationary signals. Empirical mode decomposition [72], wavelet analysis [73], and singularity decomposition [74] are time–frequency domain feature extraction methods. Kuljanic et al. [75] extracted energy-indicative features using the wavelet transform for cutting force and vibration signals. This method solves the single-sensor feature’s stability and robustness issues. However, the short-time Fourier transform has a fixed size, which prevents the simultaneous achievement of high temporal resolution and high frequency resolution. The wavelet basis function can combine with the scale function to achieve the wavelet transform, which has a high frequency resolution at a low time resolution. It achieves an effective balance between time domain and frequency domain resolution. The low frequency is decomposed by a wavelet packet transform based on the wavelet transform. It can achieve the multi-scale and multi-band analysis of different signals [76]. Empirical mode decomposition (EMD) is an adaptive time–frequency analysis technique. The signal features are decomposed step by step into different scales based on the scale characteristics, which extracts sensitive signal features. Yao et al. [77] divided acceleration signals into three layers using wavelet packet transform technology. The variance and energy ratio were extracted to reflect the chatter frequency bands. Liu et al. [34] used a feature from EMD and a support vector machine, which achieved online vibration monitoring in turning machining. The empirical mode decomposition method has the disadvantage of an endpoint effect, due to the lack of a complete mathematical theoretical basis. Moreover, the instantaneous frequency of the Hilbert transform will be lost when a single component is not complete enough. Subsequently, integrated modal decomposition (IMD) added an appropriate amount of Gaussian white noise to the signal to stabilize the decomposition process and improve the aforementioned drawbacks. Fu et al. [78] proposed two flutter indicators using IMD and Hilbert transform methods. Then, the Gaussian mixture (GM) model was used to determine the threshold. Although time–frequency domain processing methods can obtain the frequency components of cutting signals, the biggest bottleneck lies in the inability to combine high frequency domain resolution and high time domain resolution. The computational complexity is relatively high compared to time domain methods.

3.2. Machining State Recognition Technology

Feature recognition technology is needed to determine the machining state corresponding to the features after collecting signal data and extracting features. The mapping relationship is established by combining signal characteristics and machining states, which is the core principle. The widely used feature mapping methods currently include filters [79], wrappers [80], and embedding methods. Optimal feature selection is effective in improving quality. However, cutting is a time-varying process with complex nonlinear characteristics. It is difficult to obtain more accurate mapping results for many signal features using traditional modeling methods. Therefore, it is necessary to introduce machine learning techniques to reduce the complexity of process analysis. The equivalent models are established to identify machining states through calculating system inputs and outputs.

In an intelligent manufacturing context, the mining and identification of characteristics from substantial amounts of physical quality data in machining have emerged as increasingly prominent research topics. With advancements in computer technology, machine learning techniques are regarded as effective methods for processing extensive nonlinear data. Researchers employ machine learning approaches to establish relational models between data features and processing states, thereby facilitating the recognition of processing states.

Machine learning methods are composed of deep learning and shallow learning. Common shallow machine learning methods include artificial neural networks (ANNs), support vector machines (SVMs), random forests (RFs), and naive Bayes classification (NBC). Common deep learning algorithm models include deep belief networks (DBNs), autoencoders (AEs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). Each algorithm can be divided into supervised learning and unsupervised learning based on the presence or absence of labels in the training data, as shown in Figure 10.

The several critical steps of status recognition via machine learning are as follows. First, the extracted signal features are sorted and labeled according to the machining status. Second, the training and testing samples are processed to compose the dataset. Finally, the machine learning model should be selected appropriately. The training data are used to train the model, and its accuracy is tested with the test samples. Research is conducted on various machining state recognitions based on different types of signals collected during the machining process and the application of diverse machine learning models. For instance, a machine learning method was combined with cutting force measurement signal features to predict residual stresses in machining online [81]. However, given the complexity of the multiple features present in the signals collected during machining, as well as the intricate processing states associated with these signal features, selecting an appropriate machine learning algorithm is crucial for enhancing the accuracy of processing state recognition. Li et al. [82] proposed an online monitoring method for the tool status by integrating deep learning networks with a cocktail solver library. They compared and validated the effectiveness of their method against traditional approaches, as illustrated in Figure 11. Urbikain et al. [83] developed a monitoring tool using Labview 2015 that allows for the simultaneous recording of cutting force, acceleration, and sound signals. It also has extension functions to collect signals such as the temperature, displacement, and speed. On this basis, they used a combination of mathematical and geometric models to predict the surface roughness during the machining process [84]. In addition, they achieved production sustainability by using specific cutting force energy or sustainable productivity gain (SPG) factors, which made different choices in the machining of grooves [85].

According to the majority of research results, the scope of machining state recognition dictates the type of recognition model employed. For instance, the deep autoencoder (DAE) model is an unsupervised learning approach characterized by strong generalization capabilities and a lack of requirements for labeled data. This model is commonly utilized for anomaly detection in machining processes, such as tool status and chatter monitoring. Conversely, the DBN model can capture a joint distribution of observed data and labels; however, it necessitates that the input data exhibit translational invariance. This model is frequently applied in identifying tool wear and chatter. The RNN model is capable of establishing a sequential content model, but it is susceptible to issues such as gradient vanishing or exploding and lacks feature learning capabilities. It is primarily used for predicting machining deformation and the remaining tool life. Each machine learning model presents its own set of advantages and disadvantages. Therefore, when selecting a model for machining state recognition and prediction, it is essential to consider its specific strengths.

4. Status of Intelligent Decision Technology

Intelligent monitoring technology commonly includes signal acquisition technology, state recognition technology, and intelligent decision technology. Intelligent decision technology is a key technology that reflects intelligent monitoring. The machining status could be identified using intelligent decision-making to predict results and optimize machining conditions [86]. After predicting the machining status through machine learning methods, it is possible to consider adjusting the machining process conditions to improve the machining status [87]. Due to the optimization process being a nonlinear, large-scale, and multi-constrained NP-hard problem, multi-objective decision-making methods have been designed by many researchers. The experimental methods (response surface methodology [88]) established experimental analysis models. Then, the multiple weight factors of different objectives were integrated using gray relational analysis, which found the optimal solution [89]. These methods cannot provide the best solutions to meet various preferences, which are limited by experimental settings. In order to solve the above problems, multi-objective metaheuristic algorithms are widely used due to their excellent global search ability, wide constraint inclusiveness, and strong operability for application problems, such as the genetic algorithm (GA) [90], gray wolf optimizer (GWO) [91], particle swarm optimization algorithm (PSO) [92], etc. Tian et al. [93] used the non-dominated sorting genetic algorithm (NSGA-II) to conduct multiple optimization experiments. The optimal parameters of each tool wear status were obtained. Duc et al. [94] used a multi-objective particle swarm optimization algorithm to optimize cutting parameters. Salem et al. [95] combined genetic programming with NSGA-II to establish an intelligent optimization model, which can provide an intelligent learning ability for machine tools to intelligently adjust cutting conditions. However, the search of these optimization algorithms has a certain degree of blindness and randomness [96] because they generate a fixed dimensional solution and perform search operations by finding acceptable adjacent solutions. Therefore, heuristic algorithms are not as effective as adaptive optimization methods. Meanwhile, the adaptive optimization method has higher efficiency.

Reinforcement learning is an upgraded version of heuristic optimization algorithms. It can solve complex dynamic optimization problems. This approach enables agents to autonomously acquire processing knowledge through a trial-and-error search process [97]. The intelligent agent receives feedback that guides it in learning the optimal decision-making strategy following the execution of each decision. Furthermore, deep learning aligns with the dynamic optimization of cutting parameters, wherein the objective function is optimized by selecting appropriate cutting parameters within a dynamic machining environment [98]. Reinforcement learning facilitates the acquisition of processing knowledge by exploring the processing environment. Reinforcement learning can swiftly determine suitable cutting parameters based on the acquired machining knowledge when the status of tools changes, which can enable the intelligent monitoring of the machining status. Xiao et al. [99] established a reinforcement learning framework to monitor tool wear. The input data (workpiece features, machining requirements, and the tool wear status) were selected to adaptively generate control parameters. Li et al. [100] developed a tool wear prediction model. They subsequently designed a Markov decision process based on a reinforcement learning architecture to model the cutting parameter optimization process, which successfully achieved the adaptive optimization of cutting parameters. Zhang et al. [101] proposed an intelligent generation method for machining process routes based on reinforcement learning, as shown in Figure 12. They demonstrated its effectiveness in part machining process planning and its capacity to overcome the limitations of traditional methods.

Intelligent monitoring requires a carrier to achieve machining state perception, machining quality control decision-making, and decision instruction execution. In recent years, Cyber Physical Systems (CPSs) have been recognized as an important method of achieving intelligent manufacturing processes which reflects the integration of information space and physical space. Digital twin technology provides an effective approach for achieving the integration of CPSs. CPSs can be seen more as a concept, while the digital twin is a technological implementation. The core of the digital twin is building a digital virtual model equivalent to physical entities, the online perception of the manufacturing system status, the real-time analysis and prediction of the system status, making timely control decisions based on optimization goals, the execution of decision instructions by physical environment terminal devices, and then forming a closed loop with the perception module sensing the regulated system status.

The digital twin system has the abilities of autonomous perception, autonomous decision-making, precise execution, and self-improvement. The research on unit-level digital twins includes the virtual modeling of hardware such as processing equipment, tool holders, and fixtures, which are the carriers “reflecting reality with virtuality”. Modeling includes geometric information, physical information, and process information. On the basis of “reflecting reality with virtuality”, “controlling reality with virtuality” is a critical technology in realizing digital twins. It mainly includes the analysis, prediction, visualization, and control of the processing process. Cao et al. [102] proposed a digital twin system for CNC machining, which developed a synchronization algorithm based on the communication data of the computer numerical control system. They also developed a machining simulation algorithm based on three coordinates, which can efficiently track the actual machining process. Wang et al. [103] proposed a digital twin clamping force control method. They established a full factor information model of the clamping system, integrated the dynamic information of the clamping process, clarified the workflow of clamping force control and the interoperability method between digital twin models, and achieved the improvement of the machining accuracy of thin-walled components. Ward et al. [104] proposed a digital twin system capable of the real-time adaptive control of intelligent machining operations, which combined real-time simulation with online feedback to achieve closed-loop residual stress control, as well as vibration prediction and control and adaptive feed rate control. Afazov et al. [105] presented a model for predicting chatter by measuring cutting forces, which can be integrated into digital twins for process monitoring and control. Zhou et al. [106] established a digital twin model to use in cutting parameter optimization and chatter detection, and implemented a monitoring window for cutting scene visualization and a database for recording manufacturing data in the digital twin model. Zhu et al. [107] proposed a digital twin manufacturing framework that includes preparation, processing, and measurement, achieving the intelligent monitoring of the manufacturing process of thin-walled components, as shown in Figure 13.

In recent years, emerging technologies such as Cyber–Physical Systems (CPSs) and augmented reality (AR) have made significant advancements in online machining monitoring, providing intelligent frameworks for the implementation of smart monitoring technologies. A CPS integrates information technology with physical processes to achieve comprehensive oversight and intelligent management. This integration helps to dismantle information silos and promotes interconnectivity among devices. Furthermore, it enhances production efficiency and product quality through real-time data analysis and optimization. While a CPS can be viewed as a conceptual framework, the digital twin serves as its technological implementation. The essence of the digital twin lies in creating a digital virtual model that mirrors physical entities. This model facilitates the online perception of the manufacturing system’s status, enabling the real-time analysis and predictive assessments of system conditions. Consequently, control decisions are made based on optimization objectives, with the precise execution of decision instructions carried out by terminal devices in the physical environment. Ultimately, the perception module senses the regulated status, forming a closed-loop system.

Meanwhile, AR has significantly contributed to the real-time monitoring of machining processes due to its unique capability to merge virtual and real elements. Through AR technology, real-time data, the equipment status, and potential risks associated with the machining process can be intuitively accessed, facilitating faster and more accurate decision-making. Liu et al. [108] integrated AR technology into the online monitoring of the manufacturing process, assisting workers in maintaining product quality. Ceruti et al. [109] employed AR technology to superimpose virtual models of 3D printed objects onto their physical counterparts, enabling the monitoring of various printing stages. Becher et al. proposed a data visualization method that leverages handheld touch devices and AR for the real-time oversight of the manufacturing process.

5. Challenges and Prospects

Currently, the application of intelligent monitoring technology in digital twin systems for thin-walled milling is expected to become increasingly widespread. It is mainly reflected in status perception, decision-making, execution, and integration. The existing research was analyzed to summarize the challenges in the development of intelligent monitoring technology. Then, the proposed future development prospects for each challenge were proposed.

In signal acquisition technology, the processing technology becomes increasingly complex and the requirements for processing quality become higher. Collecting physical data related to the processing process is a complex task. Firstly, the hardware devices of sensors face significant challenges which require a higher accuracy and sampling frequency. The ability to transmit data online is also crucial. Secondly, the processing techniques for multi-source heterogeneous data collected by multiple sensors could experience challenges in terms of generalization and overfitting, requiring more efficient data fusion processing techniques. In addition, residual stress inside components during machining is the main factor leading to machining deformation. There is a lack of online equipment capable of collecting residual stress and machining deformation data that is suitable for complex machining environments. Consequently, it is challenging to verify the machining deformation conditions inferred from intermediate physical quality signals, such as the cutting force and cutting temperature.

In future, multi-source sensors will be integrated for complex machining conditions. They will be employed to collect physical signals online, including direct states such as machining deformation, residual stress, and the surface roughness. An artificial intelligence data source model that combines signal- and non-signal-based data sources could be developed. Meanwhile, cloud computing and edge computing technologies will be combined to develop data fusion models with higher accuracy and efficiency. They will realize fast fusion and remote storage.

In feature extraction technology, with the machining process and environment becoming increasingly complex, signal characteristics are not only reflected in the time domain, frequency domain, and time–frequency domain. They are also reflected in nonlinearity, non-stationarity, and other aspects. Each feature extraction algorithm has its own advantages and disadvantages. It is necessary to improve the adaptability and robustness of the algorithm to cope with complex machining conditions. Meanwhile, the technology for the recognition of the machining status is mainly implemented through machine learning models. The various machine learning models have their unique advantages; for example, some models are suitable for classification problems, some models are suitable for regression problems, and some states even require multiple machine learning models to be mixed. This puts higher demands on machine learning model selection and hyperparameter optimization.

In future, deep learning techniques can be combined with traditional feature extraction methods to form more advanced and adaptive feature extraction frameworks that can adapt to different processing and monitoring needs. In addition, with the continuous development of unsupervised learning and self-supervised learning techniques, feature extraction will no longer rely solely on a large amount of annotated data. By designing reasonable internal prediction tasks, the model can self-learn and extract useful features.

In intelligent decision-making technology, the relationship between state characteristics and processing conditions is predominantly nonlinear, lacking a one-to-one correspondence. Cutting is a complex physical process, and the state features extracted through signal collection are intricate and varied. Furthermore, intelligent decision-making technology typically relies on sophisticated machine learning models, which necessitate substantial amounts of annotated data and computational resources for training, making the training process time-consuming. Given the complexity and variability of the processing environment, models may experience performance degradation or failure in practical applications. Consequently, a key challenge lies in how to rapidly and accurately train intelligent decision models that are suitable for machining status monitoring while ensuring their stability and reliability in real-world applications. Additionally, although intelligent decision-making technology can achieve high levels of automation and intelligence, manual intervention and decision-making are still required in certain scenarios. Therefore, exploring how to facilitate human–machine collaboration to fully harness the strengths of both intelligent technology and human intelligence is a crucial direction for future research in intelligent decision-making technology.

In the future, intelligent decision-making algorithms must be tailored to specific processing techniques to enhance accuracy in state analysis and regulation. It is essential to develop models with robust adaptive capabilities that can autonomously adjust and optimize decision-making strategies in response to changes in the processing environment and user requirements. Furthermore, intelligent decision-making technology should be integrated with smart devices to facilitate the execution of these strategies. For instance, an intelligent machine tool could be developed that adjusts machining parameters online during the machining process. Additionally, an intelligent fixture could be designed to modify the clamping force in real time during machining, thereby mitigating the effects of residual stress field distribution on machining deformation.

Finally, intelligent monitoring technology operates as a closed-loop system. The intelligent algorithms must be executed by terminal devices following decision-making to ensure effective machining quality control. Furthermore, the essence of intelligent monitoring technology lies in the integration of perception, decision-making, and execution, necessitating an intelligent platform as a carrier to fulfill its functions. The digital twin platform is recognized as an effective means of facilitating intelligent monitoring. However, digital twin technology is still in the nascent stages of research and requires further refinement in its definition and architecture. The technology of the ‘virtual control of reality’ based on the principle of the ‘virtual reflection of reality’ necessitates in-depth investigation.

In future, a digital twin framework for the physical geometric characteristics of the machining process can be established to integrate the critical technologies for the intelligent monitoring of processing, which will achieve better monitoring capabilities. The integration of augmented reality technology within the digital twin system will enhance human–computer interaction, providing researchers with more intuitive and comprehensive insights into the processing status. Concurrently, in alignment with the sustainability, resilience, and people-oriented principles of Industry 5.0, the aim is to reduce system risks and uncertainties, conserve energy, and improve manufacturing resource productivity through digital twin systems. This approach seeks to enhance the adjustable autonomy and collaborative control capabilities between human and machine systems, ensuring superior performance in collaborative endeavors. Researchers will be able to simulate manufacturing performance, predict faults, and investigate issues within the manufacturing process by mapping resources and products in a virtual environment through the digital twin system, thereby establishing a platform conducive to a people-oriented approach.

6. Conclusions

In recent years, the degree of the integration and lightweighting of thin-walled components has been increasing, and the complexity has also increased accordingly. With the conventional offline machining status recognition and prediction technology, it is difficult to meet the increasing demand for quality. With the significant improvement in software and hardware technology capabilities such as sensor devices, computer technology, and artificial intelligence technology, machining intelligent monitoring technology has developed rapidly and been widely applied. The implementation of intelligent monitoring technology requires multiple technical supports, including signal acquisition technology, processing state recognition technology, and intelligent decision-making technology. This paper summarizes the current research status of intelligent monitoring technology for the machining status and proposes prospects for future development.

Firstly, existing research collected data on the cutting force, cutting temperature, acceleration, sound, internal physical quantities of machine tools, etc. These data were processed using techniques such as denoising, Fourier transform analysis, and wavelet analysis. The signal features were extracted using time domain, frequency domain, and time–frequency domain methods. Conventional filters, packers, embedding methods, and machine learning models were employed to identify the current processing status. Multi-objective metaheuristic algorithms, including genetic algorithms (GA), the gray wolf optimizer (GWO), and particle swarm optimization (PSO), were adopted to make decisions regarding the processing status. Additionally, reinforcement learning is an upgraded version of heuristic optimization algorithms. It can solve complex dynamic optimization problems. It can further enhance the intelligence level of state decision-making. Finally, terminal intelligent execution devices were relied upon to achieve the intelligent control of the processing status. However, challenges remain in signal acquisition technology, optimizing model structures in online recognition technology, and ensuring stability and reliability in intelligent decision-making technology. In the future, multi-source sensor fusion systems should be developed. Machine learning hybrid models should be established. Customized intelligent algorithms for machining states should be proposed. Furthermore, an intelligent monitoring platform for the processing status based on digital twin technology should be developed to enhance human–computer interaction capabilities, ultimately providing technical support for the intelligent transformation of the manufacturing industry.