As a core branch in the field of artificial intelligence, machine learning analyzes data through algorithms, discovers laws and patterns, and then makes predictions and decisions. In recent years, it has been widely used in the field of laser cladding. Various defects formed in the laser cladding process seriously affect the quality and performance of the cladding layer. The reliability and repeatability of the cladding quality are the biggest challenges facing laser cladding technology. Data-driven machine learning algorithms can be used for laser cladding process monitoring and defect detection, feedback control of cladding process, and optimization and suppression of cladding defects, which has become a research hotspot in this field. This paper reviews the types and formation mechanisms of defects generated in the laser cladding process, summarizes the signal characteristics generated in the laser cladding process and their monitoring principles and means, summarizes the research progress of machine learning methods in signal feature extraction, defect classification, recognition and prediction in the laser cladding process, and summarizes typical machine learning models and algorithms for defect detection. The results show that machine learning algorithms can be effectively used for laser cladding coating defect detection and construct the relationship between characteristic signals and coating defects and cladding processes. The machine learning algorithms currently used in the study are mainly supervised learning algorithms. Unsupervised and semi-supervised learning algorithms have low requirements for data annotation and have gained attention in the field of laser cladding process monitoring and have shown great potential. The research results point out the hot spots and directions for the research of machine learning methods in the field of laser cladding.

Laser cladding is an efficient surface engineering technology[1]. By preparing advanced functional coatings on the surface of mechanical parts, it can effectively improve the service performance of mechanical parts in extreme environments. It has been widely used in industrial fields such as aerospace, energy, automobiles and power generation. Typical parts include blades, turbine disks, valves, pistons, superheater tubes and various shaft parts[2-4]. Laser cladding technology has developed rapidly in recent years[5]. It uses high-energy laser beams as heat sources to melt metal powders or wires, and forms a fusion with a thin layer on the surface of the substrate to obtain a cladding layer with metallurgical bonding[6]. It has the advantages of high cladding efficiency, high powder utilization, small heat-affected zone, small coating dilution rate, good metallurgical bonding, high microhardness, good corrosion resistance, and the ability to repair thin-walled and small-sized components[7-9]. However, the quality of laser cladding coating is affected by process planning, material selection and processing environment. Improper selection will cause various defects such as pores, cracks and deformation in the coating, reducing the mechanical properties of the cladding layer [10-11]. How to optimize the laser cladding process, adjust parameters such as laser power, powder feeding rate, scanning speed, cladding material, reduce defects in the cladding layer, and ensure the reliability and repeatability of the cladding layer quality has become the biggest challenge facing laser cladding technology [12-13].

By monitoring the laser cladding process, the discovery and prediction of cladding defects can be achieved, and optimized feedback can be provided for the laser cladding process, reducing the probability of cladding processing failure and shortening the cladding layer development cycle. The laser cladding process is mainly based on complex physical processes, including non-equilibrium kinetics, thermodynamics and microstructure evolution of the melting process [14-16]. In order to effectively monitor the laser cladding process, it is necessary to have a deep understanding of the physical and chemical reactions occurring during the cladding process. The generation of cladding defects is the result of the coupling of process parameters. It is necessary to understand the formation mechanism and distribution law of different types of defects in the cladding layer and summarize the control methods of cladding defects [17]. On this basis, select appropriate sensors to collect process signals such as sound, light, and heat generated during laser cladding to obtain the state characteristics of the molten pool flow and solidification process [18]. Currently, commonly used signal acquisition methods include acoustic emission sensors, high-speed cameras, pyrometers, etc., which can realize non-destructive online acquisition of cladding process signals. Many researchers have established the correlation between defect generation and cladding state and process parameters through information such as acoustic signals, molten pool flow images, and molten pool temperature field distribution during the cladding process [19-22]. However, the massive data generated during the monitoring process and the complex relationship between process-signal-defect-quality make signal processing methods based on traditional statistical methods face huge challenges [23]. Therefore, laser cladding process monitoring and defect detection based on machine learning has become a research frontier and hotspot in the field of laser cladding technology.

Machine learning (ML) has been widely used in the fields of defect detection, fault diagnosis and life prediction [24-27]. Using machine learning algorithms, features can be extracted from the signals collected during the laser cladding process and classified and identified. On this basis, the classification, positioning, prediction and optimization of defects in the cladding layer can be achieved. Figure 1 summarizes the typical machine learning algorithms and main defect types used in laser cladding process monitoring. The defects in the cladding layer are mainly divided into pores, cracks, underfusion and deformation. The machine learning algorithms used for process monitoring mainly include supervised learning, unsupervised learning, semi-supervised learning and reinforcement learning. Different machine learning algorithms have their own advantages and disadvantages, and the algorithm must be optimized according to the nature of the specific problem and the characteristics of the data set [28-30]. At present, the research on laser cladding process monitoring and defect detection by domestic and foreign scholars mainly focuses on the relationship between laser energy field and material in laser cladding process, thermal behavior of molten pool and in-situ monitoring method, processing and feature extraction method of monitoring signal and classification prediction based on machine learning algorithm [31-33]. There is still a lack of comprehensive review on the formation mechanism of laser cladding defects, feature extraction of monitoring signals and classification prediction based on machine learning.

This paper investigates the domestic and foreign literature on the application of machine learning methods in the field of laser cladding process monitoring and defect detection in recent years, summarizes the defect types and their formation mechanisms in the laser cladding process, sorts out the characteristics of the signals generated in the laser cladding process and the main monitoring principles and methods, discusses and summarizes the application characteristics of different machine learning algorithms in laser cladding process monitoring and defect detection, analyzes the problems existing in the current machine learning algorithms in coating defect detection research, and summarizes and prospects the application of machine learning in laser cladding technology research in the future.

1 Cladding defects and forming mechanism
Laser cladding is a dynamic physical metallurgical process in which high-energy laser beams act on alloy powders and substrates, involving multiple links such as powder transportation, laser energy absorption, material melting, and solidification of the cladding layer. It includes the interaction of multiple complex processes such as heat transfer, convection, mass transfer, and crystallization, and is affected by multiple factors such as cladding process parameters and the properties of powder and substrate materials. During the cladding process, energy is concentrated and the solidification time of the cladding layer is extremely short. Due to improper matching and regulation of factors such as heat source characteristics, processing technology, and material properties, defects such as pores, cracks, underfusion, slag inclusions, and geometric deformation are often generated in the cladding layer [34], which in turn affects the quality and microstructure of the cladding layer and reduces the service life of the cladding parts.

1.1 Pore generation and distribution
Pores are the most common defects in laser cladding coatings. Their size, number, shape, and distribution have a direct impact on the hardness, strength, anisotropy, and fatigue properties of the cladding layer [35-39]. As a weak area in the cladding layer, pores are prone to stress concentration and crack formation, which affects the density, bonding performance, fatigue strength and service life of the cladding layer. The mechanisms of pore formation in the cladding layer are different. Pores are the most easily generated pore defects in the laser cladding process. Insufficient powder diffusion, underfusion and keyholes can also lead to pores in the molten pool [40-41]. Pores are mainly divided into two types according to the generation mechanism: entrainment type and reaction type. During the laser cladding process, the cladding powder and the substrate surface will absorb air and moisture. Under high temperature, they will volatilize and be entrained in the molten pool to form pores. The pore escape behavior is affected by surface tension, Marangoni convection effect and recoil pressure [42]. Under the coupling of multiple forces, the gas moves in the molten pool. If the gas fails to escape from the solid-liquid surface before the molten pool solidifies and closes, it will be captured by the molten pool to form pore defects [43]. The molten metal reacts with the escaped water vapor to generate metal oxides and hydrogen, and the hydrogen is further drawn into the molten pool to form pores [44]. The generation of pores in the cladding layer is determined by the escape behavior of the gas during the solidification of the molten pool. As shown in Figure 2, during the laser cladding process, Marangoni convection occurs in the direction indicated by the arrow in the molten pool, and the flow direction is from the bottom to the center surface of the molten pool, and then from the center to the boundary of the molten pool. This convection transfers the gas to the bottom of the molten pool, promotes the movement of the pores, and increases the collision chance of small holes merging into large holes. The pores move along the direction of fluid flow and are distributed in a chain.

1.2 Crack initiation and expansion

Cracks in the cladding layer are defects caused by the coupling of internal stress and metallurgical properties. The cause of crack formation and the location of crack generation are affected by the cladding parameters.

Crack defects directly affect the wear resistance and corrosion resistance of the cladding layer, and may even act as stress concentration points to induce fracture and failure of the cladding parts. Therefore, based on machine learning methods, combined with crack characteristics and distribution, exploring the intrinsic correlation mechanism between materials, processes, defects and performance, and regulating cladding materials and processes are currently hot research topics in the field of defect control in laser cladding coatings. Under the action of the laser energy field, the powder and the substrate melt to form a liquid metal pool. The molten metal in the pool is affected by the combined effects of surface tension, gravity, viscous shear stress and shielding gas pressure, which makes the thermal expansion coefficients of various materials in the molten pool convection solidification process different, thereby inducing thermal stress, phase change stress, constraint stress and other internal stresses in the cladding layer. When the local internal stress accumulation exceeds the stress limit of the material, cracks will be generated in the cladding layer [45-46]. Thermal stress plays a leading role in the crack defect generation process in the cladding layer [47-48]. When the thermal expansion coefficient of the cladding layer material is greater than that of the substrate material, the thermal stress is manifested as tensile stress, and vice versa as compressive stress. Porosity, intergranular liquid film, and segregated hard phase in the cladding layer tend to become local stress concentration areas [49]. Cracks often initiate at these locations and extend along the weak locations of the material in the cladding layer. As shown in Figure 3, crack types can be classified according to orientation and location. According to crack orientation, cracks can be divided into transverse cracks, longitudinal cracks, and mesh cracks in the coating [50]. Transverse cracks usually start at the joint between the coating and the substrate, and then extend perpendicularly to the scanning direction to the coating surface, showing the fracture of the entire coating perpendicular to the scanning direction. The shape of the transverse crack is relatively straight, showing the characteristics of transcrystalline fracture. The initiation of longitudinal cracks is mainly affected by the tensile stress on the coating cross section. When the coating is thick, the longitudinal cracks mainly extend from the coating surface along the metallurgical bonding area to the bottom; when the coating becomes thinner, the interface cracks along the metallurgical bonding area become nearly vertical longitudinal cracks on both sides of the coating. Network cracks usually initiate in the coating and develop in multiple directions in three-dimensional space driven by thermal stress. They can extend to the coating surface or interface and have the characteristics of intercrystalline and transcrystalline extension [50]. According to the crack initiation location, cracks can be divided into cladding layer cracks, overlap zone cracks and metallurgical bonding zone cracks. During the laser cladding process, the temperature distribution of the molten pool gradually decreases from the center to the edge. Driven by the unbalanced surface tension gradient, the molten metal flows from the low tension area (center) to the high tension area (edge), triggering the Marangoni convection effect [51], which leads to stress concentration and crack initiation in the cladding layer. The overlap area of ​​the cladding layer produces a significant temperature gradient due to repeated melting of adjacent welds, and the accumulation of thermal stress causes cracks in the overlap area. Since the cladding layer and the substrate have different laser energy absorption characteristics, thermal conductivity and thermal expansion coefficient, during the flow and solidification of the molten pool, heat conduction will gradually transition to the thermal convection stage, making the thermal stress in the center of the molten pool significantly smaller than that at the edge of the molten pool. The maximum thermal stress appears at the bottom of the molten pool, resulting in cracks in the metallurgical bonding zone. 1.3 Underfusion and slag inclusion After the powder flow is ejected from the nozzle and coupled with the laser beam, the powder is captured by the molten pool or forms splashes after absorbing the laser radiation energy [52]. As shown in Figure 4, the splashed powder can be divided into fully melted and partially melted states according to its physical properties. Partially melted powder will cause underfusion. When the laser power is low, the underfused powder particles on the surface of the cladding layer will increase the surface roughness [53]. In addition, when the powder feed rate is small, the molten pool width is narrow, which causes the coating overlap rate to decrease, resulting in a large amount of unmelted powder remaining between multiple welds. The slag remaining in the weld is called slag inclusion. Slag inclusions reduce the plasticity and toughness of the weld, and their sharp corners easily cause stress concentration, leading to crack initiation [54]. The remelting effect in the multi-layer cladding preparation process can improve the under-fusion state of powder between layers [55], but the layer roughness will affect the overall roughness of the cladding layer, causing interlayer gaps, inducing the formation of pores, and affecting the interlayer bonding [56-57]. Partially melted powder will form under-fusion pores in the molten pool, which will hinder the adhesion of the molten pool. Once the flowing molten pool is blocked, the surface tension caused by the temperature gradient in the molten pool on the surface of the molten droplet will cause the molten droplet to tend to roll into a ball, resulting in a spheroidization phenomenon [50, 58]. The cavity area formed between the molten droplets is difficult to be completely filled, which will also lead to defects such as pores [50, 59].

1.4 Geometric deformation of cladding layer Common laser cladding geometric defects include flatness defects, melt collapse, deformation, cracking and delamination. Geometric deformation is a macroscopic defect, and its main cause is the accumulation of stress and error during the cladding process. Geometric deformation directly affects the accuracy of cladding forming. In severe cases, it will lead to direct scrapping of parts [34]. As shown in Figure 5, the stress accumulation in the cladding layer will lead to warping deformation and form flatness defects [62]. Large cladding parts usually have flatness defects. When the heat is unevenly distributed during the cladding process, an unstable molten pool will be formed on the surface, and molten collapse often occurs at the edge of the cladding layer [28]. The residual stresses that cause deformation of the cladding layer mainly include thermal stress, structural stress, constraint stress and solidification shrinkage force [63]. These stresses are mainly caused by the temperature gradient formed by the rapid cooling of the molten pool during the cladding process. The rapid heating and cooling of the molten pool produces uneven thermal expansion and contraction, thereby forming thermal stress [64-65]. The different grain growth rates during the solidification of the molten pool cause the difference in grain size, which makes the time and degree of phase transformation different, thus leading to the generation of structural stress. The degree of deformation and position distribution of the cladding layer are related to the constraint mode, and the constraint stress is mainly related to the material properties, temperature and part size. Solidification shrinkage refers to the volume shrinkage caused by the reduction of melt volume and uneven temperature distribution during crystallization and cooling, which produces residual stress [66]. Periodic multiple remelting causes cracking and delamination defects in laser cladding parts. Delamination usually occurs between the substrate and the cladding layer or between continuous cladding layers. Interlayer cracks are usually caused by different shrinkage rates caused by different interlayer temperature gradients. Effective measures to prevent cracking and delamination include effective heat dissipation, appropriate process parameters and material adaptability [67].

2 Cladding process signals and detection
As shown in Figure 6, during the preparation of laser cladding coatings, the interaction between the laser and the powder and substrate will generate acoustic, optical, thermal and other signals. These signals contain rich cladding process information and are important for monitoring cladding process anomalies and defects in the cladding layer, predicting cladding layer performance and optimizing process parameters. Efficient and accurate acquisition of signal data is the prerequisite for achieving cladding layer quality control and machine learning prediction.

2.1 Acoustic signal
Elastic waves are generated during the melting of powder and substrate and the flow of molten pool, which carry many acoustic wave signals related to internal characteristics, including information such as defect type, defect location and cladding layer quality in the cladding layer. Selecting appropriate sensors to collect and process acoustic signal characteristics can realize defect monitoring and cladding layer quality prediction.
As shown in Figure 7, LI et al. [68] found that the amplitude of the acoustic emission (AE) signal in the normal cladding process is within 0.05 dB, while the amplitude of the acoustic emission signal in the cracked cladding process can reach up to 1.5 dB, which is 30 times that of the normal state, indicating that the acoustic signal can be effectively used to monitor crack defects. The noise signal in the cladding process is complex, and the collected signal must be denoised and feature extracted. CHEN et al. [69] used a microphone sensor to monitor the laser cladding process of martensitic aging steel C300 powder, studied the relationship between the acoustic emission signal and defects such as cracks and pores, and established an acoustic signal classification model for laser-material interaction through automated in-situ acoustic noise reduction and feature extraction, realizing cladding defect detection based on acoustic signals. Identifying characteristic signals during the cladding process is the basis for establishing the relationship between defects and acoustic emission signals. GAJA et al. [70] studied the acoustic emission signals during the laser cladding process of titanium alloy and tool steel composite powder, extracted seven features from them to analyze the laser cladding process, and used machine learning algorithms to cluster the characteristic data, effectively distinguishing the acoustic emission signals of cracks and pores. Acoustic signals can monitor the state of the molten pool and defects during the cladding process. The acoustic features obtained through extraction, classification and other processing can establish the corresponding relationship between acoustic signals and defects. However, the type, monitoring position and angle of the acoustic sensor have a direct impact on the acquisition of acoustic signals. The difficulty lies mainly in the accurate acquisition of acoustic emission signals and the feature extraction during signal processing. The use of machine learning methods can efficiently establish the relationship between signal features and defects. At present, the research on acoustic emission signal monitoring of laser cladding mainly focuses on single-pass cladding, and further development of acoustic signal monitoring methods suitable for multi-pass, multi-layer and the entire processing process is required. 2.2 Optical signal
The image information in the cladding process can be used to study the phenomena of powder, molten pool, spatter, pores, cracks, etc. Common optical signal acquisition devices include industrial cameras, high-speed cameras, spectrometers and photodiodes. By extracting the features of the collected signals, the quality and defects of the cladding coating can be monitored. Industrial camera image acquisition has low cost and high resolution, and is often used to evaluate the surface quality and defects of the coating. High-speed cameras have a high sampling frequency and can capture transient molten pool changes, monitor spatter, pores and other features, and identify defects online. Photodiodes can convert the collected optical signals into electrical signals, so that the radiation signals generated by the molten pool, spatter, etc. are converted into analog electrical signals, enriching the evaluation information of the coating quality. X-ray spectrometers can penetrate samples and reflect the number, morphology and location of defects inside the cladding layer. They can be combined with crack extension models to effectively evaluate the fatigue life of the coating.

As shown in Figure 8, LI et al. [71] used a high-speed camera to monitor the dynamic characteristics of the molten pool during laser cladding, studied the effects of different laser modes on the stability of the molten pool boundary and the solidification microstructure, and found the advantages of quasi-continuous wave laser in enhancing the geometric morphology stability and promoting the continuous epitaxial growth of columnar dendrites. ASSELIN et al. [72] designed a molten pool visual online measurement system based on three sets of industrial cameras, and developed a corresponding image analysis algorithm to extract the geometric shape of the molten pool from the image and use it as a closed-loop signal to achieve process control of laser cladding of complex curved surfaces. HOJJATZADEH et al. [73] used an in-situ X-ray method to study the process of powder melting to form a molten pool. By converting the X-ray signal into image data, in-situ observation of the pore generation process was achieved, and a new pore formation mechanism was proposed, which provided a basis for reducing pore defects. ZHANG et al. [74] filtered the optical signal radiated by the molten pool through an attenuation plate and a filter, and then converted it into a current signal through a photodiode. They studied the stability of the collected signal under different laser powers and found that the relative distance between the molten pool and the photodiode and the incident angle have a great influence on the data accuracy.

Optical signal measurement technology has the characteristics of non-contact, high efficiency, high accuracy and easy automation. It has been widely used in laser cladding process monitoring and can be used as a closed-loop feedback control signal to reduce the generation of defects. It is of great significance to improve the quality of laser cladding layers and flexible manufacturing. However, the laser cladding process is fast and has high requirements for feedback control. It is necessary to improve the accuracy and speed of optical signal recognition and optimize the cladding process recognition algorithm.

2.3 Thermal signal High-energy laser beams can melt powder and form a molten pool on the surface of the substrate. The formation, movement and solidification of the molten pool are all heat transfer processes. The complex heat conduction in the laser cladding process has a direct impact on the microstructure, residual stress, defects and deformation of the cladding layer. Therefore, studying the temperature distribution during laser cladding is of great significance for optimizing the quality of the cladding layer. During laser cladding, the sensors used for temperature monitoring are mainly pyrometers and thermal imagers. Pyrometers can usually only measure the temperature of local areas, while thermal imagers have a wider temperature measurement range. Compared with traditional contact thermocouple temperature sensors, pyrometers and thermal imagers are both non-contact and can detect the temperature distribution of the moving molten pool online. As shown in Figure 9, MUVVALA et al. [75] used a pyrometer to monitor the reaction characteristics of TiC particle-reinforced metal-based composite coatings during laser cladding. They found that low scanning speed and slow cooling process would cause TiC particles to decompose and generate dendrite structures, proving that non-destructive analysis of the decomposition state of ceramic reinforcement phases can be achieved by studying the temperature field of the molten pool. MISRA et al. [76] used an infrared pyrometer to monitor the temperature field distribution of the molten pool and explore the relationship between process parameters and molten pool characteristics. By collecting the temperature field distribution, microstructure of the cladding layer and deposition characteristics, the prediction of grain morphology, phase structure and coating defects is achieved. MAZZARISI et al. [77] used a thermal imager to study the average temperature, maximum temperature, thermal cycle, cooling rate and thermal gradient characteristics of the molten pool during the preparation of a single-layer multi-pass cladding layer. On this basis, the formation mechanism of cracks and pore defects and the law of hardness change were analyzed, and the cladding process optimization strategy was proposed. D’ ACCARDI et al. [78] used an infrared thermal imager to track and monitor the molten pool temperature during laser cladding, extracted thermal features from the selected area of ​​the molten pool, and analyzed the correlation between temperature data and cladding parameters such as laser power, scanning speed and powder flow rate through variance analysis statistical methods, and verified the feasibility of infrared thermal imagers to monitor defects and quality abnormalities during laser cladding online. The distribution of the temperature field is closely related to the morphology and thermal emissivity of the material. During the cladding process, there are liquid molten pools, solid powders, solidified metals and gaseous evaporated metals. The flowing molten pool also makes the temperature difference at different positions obvious. Therefore, the thermal emissivity of the cladding layer is not a constant, but varies with the material state, spatial distribution and temperature.

2.4 Multi-signal fusion
The signal measured by a single sensor is difficult to fully reflect the processing status and defect information during the laser cladding process. With the development of monitoring technology, many studies on laser cladding process monitoring using multi-sensor integration have emerged in recent years. By collecting multiple signals at the same time, it can more comprehensively reflect the cladding characteristics and improve the accuracy of monitoring. As shown in Figure 10, MAFFIA et al. [79] constructed a multi-sensor fusion molten pool characteristic monitoring system to simultaneously monitor the molten pool height, area and temperature characteristics during the cladding process online. The results show that the molten pool height is mainly affected by the scanning speed, while the molten pool area and temperature are mainly affected by the laser power, and the molten pool area is linearly related to the temperature. BI et al. [80] developed a new laser cladding head with integrated camera and photodiode, which realized the online monitoring of molten pool characteristics and used the controller to realize the closed-loop control of distance and temperature during the cladding process, avoiding cladding accumulation and realizing cladding wall thickness control. CHEN et al. [81] designed a multi-sensor spatiotemporal information fusion method based on visual images, acoustic emission signals, and temperature field signals. They used visual signals to label defect areas and combined acoustic emission signals with molten pool thermal characteristic signals to train defect prediction models, achieving higher accuracy than traditional single-sensor defect prediction methods.

3 Defect evaluation machine learning algorithm

During the laser cladding process, the molten pool contains rich cladding information. The molten pool morphology, temperature field distribution, spectrum, and acoustic emission signals are closely related to the internal structure, defects, and geometric accuracy of the cladding layer after solidification. By associating these characteristic variables and optimizing process parameters, the quality of the cladding layer can be improved. Machine learning aims to use algorithms to find patterns through data samples, thereby building a model with the ability to generalize and draw inferences from one example. As shown in Figure 11, machine learning methods can be used for defect detection, performance evaluation, process optimization, and exploration of complex relationships between process parameters, microstructure, and macroscopic performance of laser cladding layers. The machine learning process mainly includes data collection, data preprocessing, feature engineering, data set division, model training and verification, etc.

The machine learning methods used in the study of laser cladding processes are mainly supervised learning, unsupervised learning, and semi-supervised learning. Supervised learning uses labeled feature data for training to establish a mapping relationship between process parameters, molten pool characteristics, cladding layer defect characteristics, and cladding layer performance to predict and classify defect categories. Unsupervised learning uses unlabeled feature data for training to cluster or reduce the molten pool and defect information with similar characteristics, and to explore the intrinsic relationship between process parameters and cladding layer performance. Semi-supervised learning combines the prediction of supervised learning with the clustering of semi-supervised learning, and uses a small amount of labeled data and a large amount of unlabeled data for training to infer the defect category of unknown data.

3.1 Supervised learning
Supervised learning is the most widely used machine learning technology.
Common supervised learning algorithms include logistic regression (LR), decision tree, support vector machine, K-nearest neighbor learning, artificial neural network and convolutional neural network.
By combining multiple basic supervised learning models, better prediction can be achieved.
Ensemble learning methods mainly include bagging algorithm, boosting method and stacking ensemble learning method [82].
According to the continuity or discreteness of the data, supervised learning tasks can be divided into classification problems and regression problems. The labels of classification problems are discrete values, and the labels of regression problems are continuous values ​​[83-84].
Table 1 lists the literature related to defect analysis and prediction of laser cladding coatings in recent years, including supervised learning algorithms, materials and defect types, data set types and prediction results.
GAJA et al. [85] used logistic regression model to detect defects in Ti6Al4V and H13 mixed powder cladding coatings. By collecting acoustic emission signals during the cladding process, the peak amplitude, rise time, duration, energy and count of acoustic emission were extracted, and the two defects of cracks and pores were effectively marked and monitored. DANG et al. [86] used the ridge regression model to conduct a posteriori analysis of fatigue life of titanium alloy, extracted microstructure indicators and stress intensity characteristics, established a posteriori analysis model of fatigue life, and obtained good generalization prediction results. The regression model predicts the results by analyzing the mathematical relationship between variables. According to the mathematical relationship, it can be divided into linear regression and nonlinear regression. According to the input laser cladding parameters, material properties and defect types, the regression model can be used to predict the key performance indicators of the cladding layer. The decision tree represents a mapping relationship between object attributes and object values, and its basic process follows the “divide and conquer” strategy. The abnormal powder flow caused by the blockage of the nozzle internal flow path and the luminescence and spatter ejection in the molten pool area make it difficult to monitor abnormal powder feeding. LEE et al. [88] proposed a cladding abnormality and powder feeding abnormality recognition technology based on coaxial camera images, as shown in Figure 12. The coaxial camera was used to collect the characteristics of the molten pool under normal and nozzle blocked conditions, which were used to identify abnormal powder feeding during the cladding process. Through model training and verification, it was found that with the increase of the exposure time of the collected image, the prediction accuracy of the decision tree method increased. At an exposure time of 300 μs, the prediction accuracy reached 93%. KHANZADEH et al. [87] established the correspondence between the molten pool image and the pore position based on the decision tree model, and compared the prediction accuracy of the coating porosity using the molten pool thermal characteristics and the simple indicators of the molten pool (length, width, peak temperature, area, etc.). It was found that the morphological model containing the molten pool thermal characteristics combined with the supervised learning model showed better prediction accuracy. The error value of the decision tree model identifying the normal molten pool as pores was as low as 0.03%. The decision tree can process both numerical features and categorical features. It is more relaxed in processing raw data, does not require a lot of data preprocessing work, and does not require special data conversion. It is more flexible when processing different types of data. Support vector machine (SVM) is a model that classifies sample space based on a training set. Its decision boundary is the maximum margin hyperplane for solving learning samples. It can also be classified nonlinearly by kernel method. As shown in Figure 13, SEIFI et al. [89] used multilinear principal component analysis to extract layer-by-layer features from the molten pool image, and then used the SVM model to predict layer-by-layer quality. Two cross-validation techniques were used to evaluate the prediction performance of the proposed method, proving that the technology can achieve online monitoring of part quality. CHEN et al. [90] established a ceramic-based laser cladding coating quality performance prediction model based on SVM to explore the relationship between cladding process parameters and coating quality characteristics. The results show that the preset powder thickness, laser spot diameter and laser power are the key process parameters affecting the performance of the cladding layer. The support vector regression (SVR) model uses the SVM model for regression analysis to find the optimal fitting curve. HAO et al. [91] used the SVR model to predict the height, width and peak shift point of SS316L laser cladding coating on a tilted substrate. The model input includes tilt angle, laser power, powder feeding rate and scanning speed. After training the model, it was found that the prediction accuracy of the SVR model can be improved by adjusting the model hyperparameters (penalty coefficient, kernel coefficient). DANG et al. [92] established a fatigue life prediction model for titanium alloy parts based on the SVR model, taking stress intensity factor and micropore characteristics as input variables, and verified that the predicted results were close to the measured fatigue life data. The SVM model has a good classification effect on small sample data. By maximizing the interval of the classification boundary, it has a strong generalization ability and can achieve good performance when facing new data. The K-nearest neighbors (KNN) model classifies samples by measuring the distance between different feature values. This method only determines the category of the sample to be classified based on the category of one or several nearest samples in the classification decision. As shown in Figure 14, CHEN et al. [62] conducted a study on surface defect prediction based on the surface point cloud contour information of the cladding layer, and divided the surface defects into protrusions, dents and wavy defects. The KNN model was used to obtain an accuracy of 93.15%. WU et al. [60] used the KNN model to study the powder splashing phenomenon during laser cladding. The splashing powder 5 mm above the molten pool was collected by a high-speed camera, and the splashing powder was labeled according to the three states of unmelted, semi-melted and fully melted. The coating quality was divided into four levels according to the porosity of the cladding coating. On this basis, by assigning inverse distance weights that meet the quality level, the coating quality prediction accuracy obtained was above 95%. KHANZADEH et al. [87] extracted feature information from the temperature field distribution of the laser cladding molten pool and used the KNN model to predict the porosity of a single thin-walled Ti6Al4V sample. After parameter tuning and K-fold cross-validation method, the KNN model classification accuracy was as high as 98.44%. The KNN model is intuitive and simple, suitable for multi-category classification problems with small sample sizes, but is not sensitive to outliers.

The artificial neural network (ANN) model consists of multiple node layers, usually including an input layer, one or more hidden layers, and an output layer. Nodes are connected to one another through relevant weights and thresholds, and the weights are changed based on training data to reduce the cumulative error on the training set. The error back propagation (BP) algorithm based on the gradient descent strategy is a typical example.
In coating defect detection, the ANN model can improve the efficiency and accuracy of clustering and classification.

As shown in Figure 15, FEENSTRA et al. [93] used the ANN model to study the influence of laser cladding process on the geometry and dilution rate of cladding welds. The input layer includes seven processing parameters such as laser power, beam diameter, and scanning speed, and the output layer is the geometric size of the cladding layer.
The trained model has a prediction accuracy of 91%, 95.5% and 92.7% for the height, depth and dilution rate of the cladding layer, respectively, which effectively reveals the relationship between the cladding process parameters and the geometric characteristics of the molten pool solidification. BHARDWAJ et al. [94] established an ANN model coupled with process parameters by collecting the cross-sectional geometric dimensions of the cladding layer to study the repeatability of the single-pass Ti15Mo material cladding layer. After optimizing the model parameters, the high-precision prediction of the cladding layer dilution rate was achieved. LI et al. [95] studied the formation mechanism of the microstructure during the solidification process of the Ti6Al4V laser cladding molten pool and established an ANN model based on factors such as grain boundary inclination angle and thermal gradient, crystal orientation and Marangoni effect, which effectively described the competitive grain growth behavior and quantitative simulation of microstructure generation. The ANN model performs well on large-scale and high-dimensional data sets, can learn and model complex nonlinear relationships, is adaptive, and can extract features from data through learning without manual feature design. CIAMPAGLIA et al. [96] trained two models, feedforward neural network and physical information neural network, to study the effect of cladding layer microstructure and defects on sample fatigue life. The average error between fatigue strength prediction results and experimental results for AlSi10Mg dataset was 4%, and the maximum error was 17%. In this study, the advantages of physical model and data-driven model were integrated into a new type of physics-driven machine learning method. The peak temperature of the molten pool was inferred by using the physical characteristics derived from the temperature field and molten pool flow, so that the method can effectively learn the molten pool dynamics, and the Marangoni index was designed as an indicator of the change of molten pool flow morphology, and the complex nonlinear relationship between process parameters and molten pool width and layer height was established. WANG et al. [110] established a physics-driven temporal convolutional network (TCN) model, designed a physical model based on the molten pool dynamics and molten pool flow mechanism, extracted the inherent physical characteristics closely related to the molten pool width and molten pool layer height, and realized the prediction of molten pool width and layer height. Traditional machine learning methods have difficulty in mining hidden physical information from the temperature field and molten pool flow, resulting in a lack of physical interpretability of the complex thermomechanical phenomena of laser cladding, which may lead to an erroneous relationship between input and output. By integrating the physical knowledge or rules extracted from the physical model into the input of the machine learning-driven model, the transparency, interpretability and analytical ability of the model can be improved [111]. The convolutional neural network (CNN) model has excellent image processing and pattern recognition performance and has been widely used in the fields of laser cladding molten pool feature classification and cladding layer defect detection. The CNN model mainly includes input layer, convolution layer, activation layer, pooling layer, fully connected layer and output layer. Through local connection and parameter sharing, the mapping between input image and output target is established to realize image classification. At present, the CNN model has been applied to the detection and analysis of defects such as pores, cracks, lack of fusion and geometric deformation in laser cladding. GONZALEZ-VAL et al. [97] used the CNN model to study the quality of laser cladding layers of three types of steel. The model input is the feature parameters and quality indicators extracted from the coaxial infrared image of the molten pool. The defect dataset includes 24,444 images obtained from 50 welds, and the defect types are manually annotated. After training, the model predicts the F1 (harmonic mean of precision and recall) score of defects such as pores and lack of fusion to reach 0.975.

As shown in Figure 16, Zhang et al. [98] used a coaxial high-speed camera to monitor the characteristics of the molten pool during the laser cladding process of titanium sponge powder, and used X-ray cross-sectional tomography to extract the pore size in the cladding layer. The molten pool parameters and pore size were used as input parameters of the CNN model. The model can predict micropores below 100 μm, and the prediction accuracy of pores in the cladding layer reached 91.2%. TIAN et al. [99] developed a CNN-based PyroNet model to study the relationship between the temperature field distribution of the molten pool and the interlayer porosity. The samples with pores larger than 0.05 mm were marked as defects by tomography. The constructed data set contained 840 photos. The overfitting problem was avoided by cross-validation. The obtained algorithm had a prediction accuracy of nearly 100% for the pore defects of Ti6Al4V thin-walled parts. As shown in Figure 17, CHEN et al. [69] developed a defect recognition technology for laser cladding layers based on acoustic emission signals. The positions of pores and cracks in the cladding layer were collected by optical microscopy, and the positions were spatially and temporally matched with the collected acoustic emission signals to extract the acoustic features corresponding to the defect-free, pore and crack features. By training CNN, adaptive boosting (AdaBoost) and gradient boosting (GB) models on the corresponding noise dataset, compared with other classic machine learning models, the overall accuracy of the CNN model in predicting defects in the cladding layer reached 89%. HOSSAIN et al. [100] used acoustic emission sensors to collect acoustic emission signals in five states: equipment idle, powder spraying only, optimal cladding process, low laser power and low powder feeding rate, and used wavelet transform to obtain the time-frequency spectrum of the acoustic emission signals under each process condition. On this basis, the CNN model was used to correlate the acoustic emission signal with the defects observed in the scanning electron microscope image of the cross-section of the cladding layer. The model achieved a 96% accuracy in predicting defects in the Ti6Al4V cladding layer. XIE et al. [103] used laser cladding technology to prepare multilayer nickel-based thin-walled samples. By adjusting the cladding interval time and optimizing the wall geometry, the relationship between the preparation process and the mechanical properties of the samples was studied. As shown in Figure 18, the data set includes the thermal history collected during the wall forming process and the ultimate tensile strength of the selected area sample. On this basis, the CNN model was used to predict the tensile strength of the samples at different geometric positions, with a prediction error of less than 3%. The results showed that the cladding interval helps to improve the mechanical properties of the thin wall. PERANI et al. [101] conducted research on online monitoring and prediction of geometric deformation of laser cladding layers to avoid large deformation of the cladding coating. Using coaxial molten pool images and process parameters as CNN model training data, through model optimization, it is found that the model with more than three convolutional layers shows better prediction ability. The model provides support for adjusting the laser cladding process and optimizing the width, height and geometry of the cladding weld. FRANCIS et al. [102] pointed out that the geometric deformation of the cladding layer is mainly caused by the residual stress generated by the thermal cycle. Studying the thermal history of the local area during the cladding process can realize the prediction of the geometric distortion of the cladding layer. By collecting the temperature field and deformation data of different positions in the Ti6Al4V disc cladding preparation process, a CNN model including the cladding process parameters is established. The model can predict the deformation error of the disc manufacturing within 56 μm. CNN models focus on extracting spatial features from input data and are good at processing tasks such as image classification, target detection and image segmentation, in which the spatial relationship between pixels or features is crucial. Unlike recurrent neural networks (RNN) or long short term memory (LSTM) models, CNN models have no inherent concept of temporal sequence. In order to solve the gradient vanishing problem in the monitoring time of the cladding process and optimize the real-time process control, it is necessary to establish a hybrid machine learning model [104-106] to improve the prediction accuracy of the model. Ensemble learning combines multiple basic learners to improve the overall prediction performance and generalization ability. According to the generation method and the dependency between learners, it can be divided into serial methods and parallel methods. The representatives of serial methods are Boosting method and Stacking method, and the representatives of parallel methods are Bagging algorithm and Random forest (RF). As shown in Figure 19, GARCÍA-MORENO et al. [109] used the RF model to evaluate the porosity defects in laser metal deposition (LMD) coatings. Through image denoising and pore segmentation, 15 pore defect-related features were extracted from the image. The pores were divided into three categories according to their size: micropores, large pores, and elongated pores. They were manually annotated and a data set of 6,552 samples was established. The classification accuracy of the trained model was greater than 94%. The RF model can reduce the risk of overfitting and improve the generalization ability of the model by combining the prediction results of multiple decision trees. It has strong processing capabilities for large-scale data sets. ZHU et al. [107] studied the influence of laser cladding process parameters on the shape of 304 stainless steel cladding layer. The cross-sectional morphology of 210 cladding layers was collected using an optical microscope, and the height and width of the molten pool were extracted for constructing the data set. The prediction model was constructed using the extreme gradient boosting (XGBoost) algorithm. The model’s prediction accuracy for the height and width of the molten pool reached 97.0% and 96.3%, respectively. XGBoost reduces the risk of overfitting of the model by introducing a regularization term in each iteration, and can provide an assessment of the importance of features to the model, which helps to understand how the model makes prediction decisions. LI et al. [108] used the Stacking method to establish a cladding layer height prediction model to address the problem of material stacking at corners during laser cladding, in order to reduce defects such as collapse, bulging, and deformation caused by heat accumulation at corners. After molten pool feature extraction and cladding layer contour processing, the processing parameters and molten pool features were used as model inputs, and the cladding layer thickness was used as output. The training results showed that the prediction accuracy of the Stacking model was higher than that of the single learner SVR model, with an average absolute percentage error of 2.369 7%. The Stacking method is suitable for basic models with diversity. By selecting basic models with large performance differences, different aspects of the data can be better captured and the generalization ability of the integrated model can be improved. The disadvantage of supervised learning is that it has extremely high requirements for the quality of the dataset. Due to the relatively high cost of laser cladding and the limited data available, it is challenging to build high-quality models based on physics or data. Transfer learning (TL) is a new type of machine learning method that can connect various data sources with limited new data and build models with high transferability. As shown in Figure 20, instance-based and feature-based TL methods are only used for quality prediction and process optimization, while model-based TL methods and multi-task learning are widely used in quality prediction and detection, defect detection and process monitoring. For laser cladding datasets, most studies aim to explore the transferability from process to process, from coating to coating, and from material to material. At present, the research on laser cladding process mainly adopts offline training mode and uses limited data to verify the constructed model. However, the prediction accuracy of the trained model for the target process is limited by the number of training samples. The online TL method-assisted modeling framework may use in-situ data to improve the prediction performance of the model online [112]. 3.2 Unsupervised learning
Unsupervised learning usually explores the internal structure, relationship and law of data without prior knowledge of unlabeled data. It mainly includes clustering, self-organizing map (SOM) and deep belief network (DBN) algorithms. Table 2 summarizes the literature on laser cladding defect detection based on the above algorithms. K-means is the most commonly used clustering algorithm. It divides sample data into K groups, randomly selects cluster centers, and classifies data by calculating distance functions. It can be used to study data distribution structure or classify data objects. SOM is a competitive learning type of unsupervised learning algorithm. It calculates the data in the input space, reduces the dimension to generate a low-dimensional, discrete mapping (Map), and maintains the topological structure of the input data in the high-dimensional space, thereby realizing data classification.

As shown in Figure 21, REN et al. [116] used LSTM-autoencoder to extract the characteristics of spectral signals during the cladding process of 7075 aluminum alloy, and collected the characteristics of pore defects from the SEM photos of the cross-section of the cladding layer. On this basis, K-means clustering was used to classify and predict the quality of the cladding layer. The model can effectively distinguish between unqualified cladding layers and qualified cladding layers based on surface roughness and high porosity. The LSTM model can capture sequence characteristics and improve the overall prediction performance. The DBN model can capture the nonlinear relationship in the data and has better fitting ability for complex data structures. The SVM model constructs a hyperplane in the feature space and has good processing ability for linear inseparable problems. Combining the DBN and SVM models can be classified in a higher dimensional space, thereby better processing acoustic emission signals.

3.3 Semi-supervised learning
Supervised learning requires a large amount of labeled data to train the model, and data labeling is time-consuming and expensive. Unsupervised learning has high requirements for data quality. The data after clustering or dimensionality reduction lacks interpretability. The lack of targets and labels makes the performance and accuracy of the model subjective and difficult to quantify, and there is a risk of overfitting. Semi-supervised learning can be assisted by unlabeled data based on partially labeled data. Unlabeled data is usually easier to obtain, so the scale of the training set can be expanded to improve model performance and generalization ability. Table 3 summarizes the literature on laser cladding defect detection based on semi-supervised learning algorithms. YADAV et al. [117] used tomography technology to study the drift phenomenon in the laser cladding process. They used an unsupervised K-means clustering algorithm to label the unlabeled data with the help of a small amount of labeled data and helped select the most appropriate distance metric. They trained the KNN model with the labeled data set and successfully classified the cladding layer into “drift” and “no drift” using a semi-supervised method. K-means clustering can be seen as a way to reduce the dimensionality of data. By selecting the cluster center as a feature, the input dimension is reduced, which helps the KNN model to classify more effectively. As shown in Figure 22, PANDIYAN et al. [118] used a coaxial camera to monitor the quality of the laser cladding process in situ. The quality of the cladding layer was divided into six levels according to the cladding process. The CNN model based on contrastive learning was used to cluster the collected molten pool images. On this basis, the logistic regression algorithm was used to supervise the classification of low-dimensional data. The obtained semi-supervised model had an average accuracy of 97% in predicting the defect type in the cladding layer. YUAN et al. [119] designed a semi-supervised CNN model for monitoring the 316L stainless steel powder laser cladding process. By collecting the geometric features and molten pool features of a single cladding layer, the semi-supervised CNN model was trained using partially labeled data and a large amount of unlabeled data. The results show that the regression and classification performance of the semi-supervised method are better than those of the fully supervised method. JAFARI-MARANDI et al. [120] used an infrared thermal imager to extract thermal data of the Ti6Al4V laser cladding pool, and used X-rays to capture the microstructure and defect information in the cladding layer. The multi-layer perceptron (MLP) was used to predict the position of the data on the SOM algorithm mapping. The constructed semi-supervised self-organizing error-driven neural network can effectively predict the spatial distribution of pore defects in the cladding layer, providing support for optimizing and controlling the mechanical properties of the cladding layer. MLP can automatically perform feature learning during the learning process and extract useful information from the input data. By mapping the output of MLP on the SOM algorithm, it helps to reduce the data dimension and visualize the data structure, so that the model can better adapt to different data features. However, the combination of MLP and SOM algorithms requires relatively more hyperparameters to be adjusted, which may increase the complexity of the overall calculation. 4 Conclusion and Prospect
The application of machine learning algorithms in the field of laser cladding defect assessment is reviewed, and a comprehensive and in-depth analysis of common defects and their formation mechanisms in the laser cladding process is conducted; the sound, light, and heat signals generated in the cladding process are summarized, the corresponding relationship between the signals and cladding defects is explained, and the commonly used laser cladding process monitoring methods, sensors, and signal characteristics are summarized; the classification and characteristics of machine learning algorithms are sorted out, and their application in laser cladding process signal processing is summarized. Through the analysis of the literature, the summary is summarized as follows:
(1) The laser cladding process is complex, and the defects generated directly affect the quality of the cladding layer. The formation mechanism and distribution law of the defects are affected by many factors, and the defects also affect and evolve each other. At present, domestic and foreign researchers have conducted research on defects such as pores and cracks from multiple scales through experiments and simulations, but the generation mechanism of related defects and the mechanism of the defects on the cladding quality are still not in-depth enough, and more abundant methods must be used to study the laser cladding process.
(2) Constructing a quantitative evaluation system of laser cladding process process-signal-defect-quality is a key challenge to ensure the reliability of laser cladding quality. At present, various sensors such as sound, light and heat have been used in laser cladding process monitoring to study the relationship between signal and process, defect and quality. However, due to the limitations of sensor accuracy and defect feature extraction efficiency, it is still challenging to establish a quantitative relationship between process-signal-defect. It is necessary to develop online laser cladding process monitoring technology and defect feature extraction technology with multi-sensor and multi-signal fusion to obtain comprehensive, reliable and accurate cladding information and defect status, and realize real-time quality monitoring of the whole process, which is an important development direction of laser cladding process monitoring. (3) Machine learning algorithms have been applied in laser cladding defect detection. Usually, a data set is constructed based on the features extracted from the collected signals, cladding process and defect features, and then a machine learning algorithm is used to establish the relationship between signal and defect and process. However, the current research on monitoring of laser cladding process mainly focuses on single-pass or small-area cladding layers. The small data set collected will cause overfitting of the model, resulting in reduced accuracy of actual defect detection. It is necessary to design a universal standard defect detection database for the laser cladding process. In addition, the selection of appropriate machine learning algorithms is crucial for different defect detection in the laser cladding process. Different algorithms have their own advantages in processing image data or sensor signals. CNN is the preferred method for processing various defect image data. The SVM method is suitable for multi-class classification problems of sensor signals or images. K-means clustering is widely used in unsupervised and semi-supervised learning. In order to promote laser cladding technology to become a new quality productivity in the field of mechanical manufacturing and remanufacturing, the application of machine learning methods in laser cladding technology is prospected as follows: (1) Laser cladding technology has a wide range of application prospects in the manufacturing industry, and the introduction of machine learning technology can effectively improve the efficiency of laser cladding and reduce defects in cladding coatings. The currently reported literature mainly uses supervised learning algorithms, but supervised learning has high requirements for data annotation and takes a lot of time and cost. Therefore, unsupervised and semi-supervised learning algorithms have gained attention in the field of laser cladding process monitoring, and new models have emerged one after another, showing great potential. (2) Through machine learning technology, it is possible to realize the automatic control and online monitoring of laser cladding equipment. By analyzing and mining a large amount of laser cladding data, optimizing and adjusting parameters such as laser power, scanning speed and powder injection, the laser cladding process can be automated and intelligent, improving production efficiency and reducing costs. Machine learning technology will bring more innovation and development opportunities to the field of laser cladding, and help the technology to be widely used in the manufacturing industry.