Intelligent Online Sensing of Defects Evolution in Metallic Materials during Plastic Deformation

Metallic plastic deformation involves complex microstructural changes and defect evolution, posing challenges in predicting and controlling the quality and performance of formed parts. Therefore, a pressing demand exists for a proficient online defect‐sensing system to monitor defects evolution continuously within components during plastic deformation in real time. This article proposes an intelligent online sensing approach for detecting defects in metallic plastic forming based on acoustic emission (AE) and machine learning. A comparative analysis is conducted on AE amplitude signals, stress–strain curves, and defect evolution during the tensile process of TA15 titanium alloy specimens under different stress states. It is found that the defect formation process can be divided into four stages based on the AE amplitude signals. A convolutional neural network model for intelligent defect sensing is established. It leverages transfer learning and is grounded in the relationship between AE signals and the evolution of internal defects. The prediction accuracy using different pretrained models is investigated and compared. It is discerned that utilizing GoogleNet as the pretrained model offers the swiftest training pace with a prediction accuracy of 97.57%. This approach enables intelligent online sensing of internal defect evolution in metal plastic deformation processes.


Introduction
Metal plastic forming, a key manufacturing process for high-performance components, requires precise process control to ensure product quality, which is significantly affected by the evolution of internal defect.The evolution of defects is highly complex and closely coupled with microstructure and stress conditions.3][4] Currently, online sensing of metal forming primarily relies on ultrasonic, eddy current and acoustic emission (AE) detection technology.Ultrasonic detection offers advantages such as the ability to inspect large-sized components, high sensitivity, speed, and cost-effectiveness, enabling both quantitative and qualitative defect analysis.[11] AE detection technology provides an integral solution, demonstrating high sensitivity to the appearance of microcracks and enabling qualitative and positional defect analysis.14] Several studies have shown that the AE detection holds the potential to identify both the occurrence and location of macroscopic cracks during the plastic deformation. [15,16]Behrens et al. [17] reported that the AE signals evolve throughout the forging process, and can indicate the eventual development of macroscopic cracks.Moreover, the reactivity of AE signals varies in response to distinct lubrication conditions.The activity of AE signals depends on the material's plasticity and process conditions.The lower the material's plasticity, the more pronounced the AE signals are. [17]Additionally, changes in the microstructure such as dislocation, twinning, and cracks influence the variations in the characteristic parameters of AE signals, which can be used to predict the material's strength. [18,19]Different plastic deformation mechanisms produce AE signals with distinct characteristics.For example, AE signals from dislocation slip are continuous, while those from twinning are more abrupt. [20]In the process of plastic deformation, macroscopic cracks typically originate from the nucleation of microvoids and the progressive growth of microcracks.Due to the significant energy release associated with the formation of macroscopic cracks, they can be readily identified through AE detection, whereas the formation of microvoids, microcracks, and other microscopic defects often remains inconspicuous.Nevertheless, the detection of these microscopic defects holds paramount significance, as they can evolve into macroscopic cracks during service, hastening component failure.Therefore, the detection of entire process from formation of microscopic defects to their evolution into macroscopic cracks is crucial for manufacturing defect-free components.Regrettably, there are currently no reported findings on this matter.
In contrast, AE signals contain abundant information and large data volumes.They are susceptible to environmental noise and stray signals, while manual analysis requires a substantial amount of labor and time.Moreover, it is challenging to discover the intrinsic connections between AE signal characteristics and material deformation behavior.Machine-learning-based detection models offer unique advantages in dealing with noisy and perturbed signals while also enabling the exploration of nonlinear relationships between variables.They have been widely applied to uncover hidden relationships between AE signal waveforms and material damage mechanisms, particularly in damage type identification.The fundamental process includes data collection, feature extraction, identification, and error analysis. [21]achine-learning-based defects identification includes unsupervised clustering and supervised classification.Unsupervised clustering can automatically distinguish different locations in the feature parameter space of AE signals, corresponding to different sources of AE signals, i.e., different material damage mechanisms. [22]But it has some degree of fuzziness, and the results of each clustering may not be entirely consistent.In contrast, supervised classification requires a predefined set of AE signal features for each class, followed by the classification of the collected signals.This approach enables the precise quantification of information for each category of AE signals. [23]n recent years, some researchers have combined AE testing with machine-learning techniques for defect detection in component manufacturing. [12,24]However, most of these studies involve real-time acquisition of AE signals during forming process, followed by post-analysis of these signals, which does not enable online sensing of forming defects.Currently, there have been no reports on online sensing of defect evolution during metal plastic deformation.
To address the challenge of real-time online sensing of defect evolution in metallic-plastic-forming processes, this study proposes an intelligent online sensing method based on AE and machine learning.Initially, experiments were conducted using TA15 titanium alloy sheets, collecting AE signals during tensile deformation under different stress conditions.By establishing the correlation between the AE signal characteristics, engineering stress-strain curves, and the formation of micro-voids or microcracks, this research classified and delineated the basic stages of defect formation during TA15 plastic deformation.Subsequently, a convolutional neural network (CNN) was constructed using transfer learning to identify the signal images collected by the AE system.This enables real-time determination of the stage of defect formation, providing an effective method for intelligent online defects evolution sensing in metallic materials plastic deformation.

Intelligent Sensing Framework
The framework of intelligent online defects evolution sensing in metal plastic forming based on AE and machine learning is illustrated in Figure 1.First, as illustrated in Figure 1a, a group of metal specimens was stretched to fracture, during which real-time AE amplitude signals were recorded to observe their distinctive characteristics.Another group of metal specimens was stretched until the appearance of specific features, allowing for the observation of internal defects.This helped establish the intrinsic relationship between AE signals and the evolution of internal defects in the material.Subsequently, based on the AE amplitude signals, the internal defect evolution during metal plastic deformation was identified.From the complete signals recorded in the first group, the process of collecting AE amplitude signals was reverse engineered.This involved redrawing the amplitude signal graph at regular time intervals, with each graph containing all the AE amplitude signals collected up to a specific moment.Each amplitude signal graph was then labeled with the corresponding stage of internal defect evolution during metal plastic deformation, forming an AE-defect dataset.
Next, as shown in Figure 1b, a CNN was constructed using transfer learning.This CNN was trained using the dataset constructed in the previous step to enable it to identify the stage of internal defect evolution corresponding to the amplitude signal graph at a particular moment.Finally, the trained CNN was applied to the real-time-recorded AE amplitude signal graphs, enabling intelligent online defects evolution sensing in the metal plastic deformation.
It was worth to note that this framework could be applied to various alloys, and this study demonstrates it using TA15 titanium alloy as an example.The stress state for plastic deformation, the method of collecting AE signals, the characterization of internal defect evolution in the metal, and the CNN model used in transfer learning were all flexible within this framework.The approaches outlined in the following sections represent the recommended methods used in this study.Section 2.2 describes the materials and sample geometries used in this study.Section 2.3 details the specifics of AE signal collection and the means of characterization internal defects in the metal.Section 2.4 explains the selection of the CNN model.

Materials and Testing Specimens
The materials used in the experiments were 2 mm thick annealed TA15 (Ti-6Al-2Zr-1Mo-1V) titanium-alloy-rolled sheets.The chemical composition of the alloy, presented as mass percentages (wt%), is shown in Table 1.
The stress triaxiality could reflect the complex stress state of the plate during the forming process.Therefore, five representative tensile specimens with different stress triaxiality values were selected.Their geometric shapes and dimensions are shown in Figure 2, with stress triaxiality values varying from left to right as 0.25, 0.33, 0.35, 0.43, and 0.53.Increasing stress triaxiality indicates a higher level of stress concentration during deformation.The specimens were cut from the TA15 sheets in the rolling direction using wire cutting.Each specimen had a thickness of 2 mm.Subsequently, the specimen surfaces were polished to a smooth finish according to standard metallographic sample preparation procedures to ensure a smooth cross section without surface microcracks that could potentially lead to premature specimen failure.Figure 3a-e shows the scanning electron microscopy (SEM) images at the gauge section edges of original specimens and no microcracks or microvoids were observed.Figure 3f shows the EBSD results of raw grain morphology on original specimen, there were strip-like grains from the  rolling of the plate, which were oriented in the direction that the specimen was about to be tension.

AE Signal Collection
As shown in Figure 4, a dual-sensor setup was chosen for the AE sensors, and silicone oil (Grade 6) was used as the coupling agent between the sensors and the specimens.The uniaxial tensile tests were conducted on an 100 kN material universal testing machine AG-IC manufactured by Japanese Shimadzu company.The tests employed a constant loading rate, with a uniform stretching speed of 1 mm min À1 .Each stress condition was tested with five repetitions to minimize experimental errors.The acquisition of AE signals was divided into two groups.In the first set of tests, AE signals were continuously recorded throughout the entire tensile process until the specimen fractured.In the second set, the specimens were stretched to different degrees of deformation, and SEM was employed to observe the formation of internal microvoids, microcracks, and macroscopic cracks within the material.
Prior to each experiment, a "pencil break" test was conducted.In this test, an Hardness Black (HB) pencil head was broken on the middle surface of the specimen's gauge length.If both channels of the AE-sensor-recorded signals with amplitudes greater than 80 dB, it indicated good coupling between the AE sensor and the specimen's surface, allowing the experiment to proceed.Otherwise, coupling was reestablished before conducting the experiment.
The AE signal collection system used in this study was MISTRAS Express-8 system, which was built upon digital signal processing.A schematic diagram of the system was illustrated in Figure 4a.After being sensed by the AE sensors, the signals were initially processed through a preamplifier before entering the AE system.Subsequently, they passed through a series of components, including filters, the main amplifier, and an A/D converter, before being available for analysis.The preamplifier's amplification factor was set to 40 dB.
Before the tensile tests, the background noise level of the AG-IC 100 kN material testing machine during unloaded tensile tests was measured to be approximately 40-45 dB.Consequently, the signal threshold for this experiment was set at 45 dB.
Previous research indicated that the AE signals generated by different deformation mechanisms of titanium alloys fall within the range of 200-600 kHz. [25]Therefore, the frequency band of the AE sensors used in this study was set between 200 and 750 kHz.Coupled with a preamplifier and an analog filter, it could detect AE signals in the frequency range from 100 kHz to 1 MHz, demonstrating the capability to capture all AE signals generated during the deformation of TA15.The sampling rate was 5 MSPS (MSPS-mega-samples per second), and the sampling length was 5 k.Based on empirical values, the peak definition time, hit definition time, and hit-lock time were set to 300, 600, and 1000 μs, respectively.The maximum duration time was left at the default value of 1000 ms.

Built-up of the CNN Model
Due to the inherent ambiguity and randomness in AE signals, coupled with the complex propagation of elastic waves within materials, there existed a nonlinear coupling relationship between the feature parameters of AE signals and the evolution of internal defects in materials.In such cases, the data-driven model had significant advantages.CNNs, with their unique ability to handle image data and the advantages of local weight sharing, were particularly well suited for image-based data samples.Compared to directly analyzing data from the backend of the AE system, AE signal amplitude images more intuitively depicted the temporal distribution characteristics of amplitudes.Using AE signal amplitude images as input allowed the CNN to accurately identify the information contained behind the amplitude distribution, facilitating intelligent online perception.
The framework proposed in this study for online monitoring of internal defects in metal-forming processes using AE and CNN was divided into a training phase and a prediction phase, as shown in Figure 1b.The AE system we used had the capability to output images based on the AE signals collected at regular intervals during the AE signal acquisition process.It could be configured with specific parameters such as the features represented by the Â and y axes, time interval length, scatter plot shape, and color of the output images.With these features, it was easy to directly extract real-time AE amplitude signal images for defect detection during plastic deformation.Therefore, to simulate image recognition during actual detection, the images used to train the CNN model must also be the same as the automatically generated AE signal amplitude images during actual detection.However, to ensure that the AE amplitude signal images used for training came from different stages of defect evolution and were evenly distributed, resulting in a CNN with higher recognition accuracy, we evenly sliced the complete records of AE amplitude signal images along the time axis.The accuracy of CNNs trained using raw image data collected in real time by the AE system and CNNs trained on image data drawn from the system's backend might be slightly different, but this would not affect the effectiveness of the proposed intelligent online sensing framework.
The time interval between each slice was set to 1 s, and all the data before the current slice were plotted on a single image, creating an image dataset that mimicked the process of collecting AE signal amplitudes during TA15 deformation.In the training phase, AE signals collected during the deformation process of TA15 were used as inputs, and the internal defect evolution stages were used as labels.This phase aimed to train a CNN capable of identifying the nonlinear relationship between the two.In the prediction phase, AE signals generated in real time during the deformation of TA15 were captured.At regular intervals, the images of these AE signals were input into the trained CNN, allowing the CNN to predict the current internal defect formation stage of TA15.

Analysis of the AE Signals during Plastic Deformation of TA15
AE signal features can reflect various information, with key parameters comprising amplitude, count, energy, duration, rise time, center frequency, and peak frequency, etc.Among these parameters, peak frequency and amplitude are considered the preferred AE features for material fracture identification. [16]In this research, the amplitude of AE signal was chosen as the feature parameter to analysis because amplitude features in the time domain can directly reflect acoustic phenomena resulting from various phenomena in the metal plastic deformation process, such as dislocation slip and crack propagation.In addition, amplitude features in the time domain are easier to interpret and establish a direct connection with the stress-strain of the deformation process, and their changing characteristics are sufficient for judging whether defects occur during plastic deformation.In contrast, frequency domain analysis usually requires complex mathematical calculations, such as Fourier transforms or wavelet transforms.These calculations may increase the complexity and time cost of processing, posing a challenge for real-time online sensing.
The engineering stress-strain curves and the history of AE amplitudes profiles during the tensile testing of the five different-shaped TA15 specimens are shown in Figure 5.
From Figure 5, it is evident that the specimens emit strong AE signals at the point of fracture, coinciding with a decline in engineering stress.This abrupt surge in AE signals at the final stages signifies the macroscopic fracture of the specimen.Prior to the ultimate failure, AE amplitude signals evolve with deformation, manifesting two distinct characteristic peaks.In Figure 5a,b, engineering stress-strain curves for the shear specimen (shear) and dog bone specimen (DB) samples show clear yield points, and the first peak in AE amplitude appears right at the yield point, which is consistent with previous research results. [26]This phenomenon is attributed to the fact that material's nonuniformity leads to the formation of numerous local stress concentration points during the elastic deformation stage, resulting in intense dislocation movement and an increase in the AE signal.After the material yields and enters the plastic deformation stage, dislocations gradually accumulate, leading to a decrease in the AE signal.
In Figure 5c-e, there are no obvious yield points in the engineering stress-strain curves for the specimens with a central round hole (CH), specimens with a notch radius of 6.35 mm (RN6.35), and specimens with a notch radius of 3.175 mm (RN3.175).However, significant peaks in the AE signal are observed, representing the yield point of TA15.This indicates that AE can detect intrinsic information that deformation force alone cannot reflect.As the deformation increases to a certain extent, a second peak in the AE amplitude can be observed in Figure 5a-e.The generation of these two peaks is related to dislocation movement within TA15 and the final sharp increase is associated with macroscopic fracture.The specific reasons will be discussed in Section 3.2.
The different stress states of the specimens all underwent a period of plastic deformation before entering the necking stage.This indicates that TA15 exhibits a certain degree of plasticity under both low and high stress triaxiality conditions.Table 2 shows the time, true strain, and true stress of two peaks in AE signal amplitudes as well as fracture for the five TA15 specimens.The stress-strain data are calculated from the loaddisplacement curves.It can be observed that the shear fractured first, with deformation concentrated in a smaller region, reaching a high fracture strain of 0.142 in a short time.This is because it underwent shear fracture rather than the ductile fracture observed in other samples.Different fracture mechanisms can have varying effects on the characteristics of AE signals.Therefore, in subsequent discussions, the shear will be analyzed separately from the other specimens.For the remaining four specimens, as the stress triaxiality increased, stress concentration during tensile deformation occurred more easily, resulting in earlier fracture and lower ultimate fracture strain.[29] For example, the DB specimen with a stress triaxiality of 0.33 had a fracture strain of 0.196, while the RN3.175 specimen with a stress triaxiality of 0.51 had a fracture strain of 0.089.
According to Table 2, except for the shear, the higher the stress triaxiality, the earlier the first peak of the AE signal amplitude appears, indicating earlier yielding of TA15.Additionally, the second peak in the AE signal amplitude also appeared earlier, and corresponded to a smaller true strain.This is consistent with the order of fracture occurrence and the trend of decreasing fracture strain among the different specimens.However, for the stress level at which the second peak in the AE signal occurs, as shown in Table 2, due to the influence of stress concentration, specimens with higher stress triaxiality reached the same stress level as samples with lower stress triaxiality at smaller strain conditions, around 1100 MPa or even higher.This makes it easier for microvoids and microcracks to nucleate and grow, which is the reason why they fracture earlier. [30]

Dislocation Motion
The AE signals generated during the plastic deformation of metals originate from various dislocation movements, including dislocation breaking away from pinning points, dislocation cooperative motion, dislocation annihilation, and escape to free surfaces. [26]For near-α titanium alloys like TA15, the plastic deformation mechanisms involve different types of dislocation slip, primarily including basal slip, prismatic slip, and pyramidal slip, with dislocation basal and prismatic slip being the main deformation mechanisms. [31]tudies have shown that the peak values of AE signals near the yield point in metals and alloys are closely related to the internal dislocation movements. [32]When TA15 begins to deform, significant dislocation activities occur inside the material due to the initiation of prismatic slip, resulting in an increase in the amplitude of AE signals.As strain increases after yielding, dislocations start to pile up, gradually forming dislocation walls.And then, the free migration distance of dislocations decreases, leading to a gradual reduction in the activity of AE signals. [33]he appearance of the second peak in TA15 signifies the emergence of new microdeformation mechanisms during its deformation process, reinvigorating the activity of AE signals.
As the deformation level in TA15 gradually increases, individual slip systems within the grains become insufficient to accommodate coordinated deformation.At this point, single slip often transforms into multiple slip, involving basal slip, prismatic slip, and pyramidal slip working in concert.Additionally, with increasing strain levels, the initiation of the pyramidal slip system, which has a relatively high critical resolved shear stress, becomes more pronounced. [34]The vigorous dislocation slip movements result in the reactivation of AE activity within TA15, leading to a resurgence in the amplitude of AE signals.
In specimens with higher stress triaxiality, the degree of stress concentration within the material increases.Consequently, it is more likely to reach the critical resolved shear stress required for the initiation of the pyramidal slip system, and multiple slip activation occurs earlier.
However, as shown in Table 2, except for the shear sample, the higher the stress triaxiality in the other four samples, the lower the fracture strain and the higher the hardening rate.This indicates that dislocation pileup and entanglement occur earlier, hindering the AE signal.Therefore, the relationship between stress triaxiality and the amplitude of the second peak in AE signals depends on the competition between the activation of AE signals due to the initiation of multiple slip systems and the hindrance of AE signals due to the rise in dislocation entanglement.Specifically, in the case of low to medium stress triaxiality, the second peak amplitudes of the shear, DB, and CH samples are 57, 66, and 88 dB, respectively, showing an increasing trend.This is because a higher concentration of stress also leads to a higher degree of multiple slip, resulting in greater activation of AE signals and an increase in the amplitude of the second peak.In the case of medium to high stress triaxiality, the second peak amplitudes of the CH, RN6.35, and RN3.175 samples are 88, 82, and 79 dB, respectively, showing a decreasing trend.This is because, under high stress concentration, the faster appearance of dislocation entanglement and pileup leads to a dominant hindrance of AE signals, causing a decrease in the amplitude of the second peak.
In particular, for the DB sample and CH sample with similar stress triaxiality, the second peak amplitude of the CH sample is 22 dB higher than that of the DB sample.This is because the stress triaxiality used in this study is the average value on the cross section of the deformation center.For the DB sample, the stress triaxiality is relatively uniform on the cross section, and although it has a higher fracture strain and lower hardening rate, the lower stress concentration leads to a smaller degree of initiation of multiple slip systems, ultimately resulting in a smaller second peak amplitude.In contrast, the stress triaxiality of the CH sample is not uniformly distributed on the cross section, with the stress triaxiality near the edges of the circular hole perpendicular to the loading direction being higher than the average value.This leads to intense initiation of multiple slip systems in the local region of the CH sample, resulting in a higher degree of activation of AE signals.Therefore, the CH sample exhibits a higher second peak amplitude, and this region of the CH sample is also most prone to cracking, as detailed in Section 3.2.2.

Formation of Voids and Cracks
To determine the stages corresponding to the generation of internal microcracks and microvoids in TA15, two additional groups of experiments were conducted.In one group, TA15 specimens were stretched until reaching the ascending phase before the second peak in the AE signal amplitude.For each of the five specimen types, tensile durations were as follows: 65, 230, 160, 85, and 60 s.The SEM images at the gauge section edges of these specimens are illustrated in Figure 6.In the other group of experiments, TA15 specimens were tensioned until the descending phase after the second peak in the AE signal amplitude before fracture.For each of the five specimen types, tensile durations were as follows: 80, 310, 200, 110, and 90 s.The macroscopic necking and the microstructures near the gauge section edges of these specimens are shown in Figure 7.The choice of observing the edges is because microcracks and microvoids are more likely to initiate in this region.
As shown in Figure 6, there are no apparent microcracks or microvoids generated on the specimen's surface at this deformation stage.However, when the AE signal enters the descending phase of the second peak, as seen in Figure 7, noticeable macronecking and shear bands appearance were observed.At the microscopic scale, clear microcracks and microvoids become evident.In Figure 7a, for the shear, the shear stress condition results in a distinctive morphology of internal defects, characterized by elongated, slender microcracks propagating along the shear direction.In Figure 7b-e, some microvoids are observed, as well as the coalescence of voids, the formation and propagation of microcracks.All oriented at a typical 45°angle relative to the loading direction.Consequently, based on the results of microscopic defects at different moments, the descent phase of the AE signal amplitude before fracture corresponds to the formation of microvoids and microcracks.In this stage, the number of mobile dislocations decreases, and microvoids and microcracks start to initiate, propagate, and coalesce within the material.These developing defects features gradually impede the propagation of AE signals, resulting in a reduction in AE signal activity. [32,35]Therefore, the presence of a descending phase in AE signal amplitude before the fracture of TA15 indicates the accumulation of internal damage and defects, serving as an early warning for the onset of macroscopic cracks during TA15's plastic deformation.
Furthermore, except for the shear, the proportion of time during which the AE signal amplitude's second peak descends relative to the total deformation duration increases with an increase in the stress triaxiality.For instance, the DB specimen accounts for 15.2%, the CH specimen for 21.7%, the RN6.35 specimen for 30.7%, and the RN3.175 specimen for 36.4% of the total deformation time.This indicates that with higher stress triaxiality, where stress concentration is more pronounced, microvoids tend to nucleate, grow, and coalesce earlier, resulting in the earlier decline of AE signal activity.

Defect Evolution Stage Division Based on AE Amplitude
The analysis provided earlier indicates that changes in AE amplitude can reflect the evolution of internal defects during the plastic deformation process of metals.Therefore, it becomes feasible to distinguish various stages of internal defect evolution during plastic deformation based on the amplitude history.Figure 8 represents the division of defect evolution stages during the plastic deformation of TA15 sheet metal based on the amplitude history.The first amplitude peak corresponds to the yield point of TA15, marking the transition from elastic to plastic deformation.Between the first and second amplitude peaks, the variation in amplitude primarily results from the motion of dislocations, corresponding to stable plastic deformation without the formation of internal defects.After the second amplitude microdefects begin to form within TA15, including the initiation and propagation of microcracks and voids.The AE activity decreases for a period before suddenly increasing when TA15 experiences macroscopic fracture.Based on the amplitude history of the AE signal, we can divide the defects evolution of TA15 sheet metal during plastic deformation process into four stages: 1) elastic deformation stage; 2) uniform plastic deformation stage; 3) initiation and propagation of microcracks stage; and 4) macroscopic crack formation and fracture stage.

Elastic Deformation Stage
In this stage, the material undergoes elastic deformation, and the initiation of dislocation slip systems with relatively low critical resolved shear stress becomes the source of AE.The AE signal exhibits an increasing trend at local maxima, while the AE signals between local maxima show high fluctuations.As the load continues to increase, the material reaches the yield point and begins plastic deformation, at which point the amplitude of the AE signal first reaches its peak.

Uniform Plastic Deformation Stage
In this stage, the material is in the early to mid-stage of plastic deformation, and the amplitude of the AE signal initially decreases and then increases.When the material enters the early stage of plastic deformation, there is limited coordinated plastic deformation by individual slip systems within the grains, and dislocation motion is not very intense, resulting in low AE signal activity and a decrease in amplitude.As the load further increases, in the mid-stage of plastic deformation, more coordinated plastic deformation is achieved through the activation of multiple slip systems, leading to an increase in dislocation density and the release of the material's plastic deformation energy, causing the amplitude of the AE signal to rise.

Initiation and Propagation of Microcracks Stage
In this stage, the material is in the late stage of plastic deformation, and the amplitude of the AE signal decreases.At this point, there is severe dislocation blocking inside the material, leading to reduced AE activity.Microvoids and microcracks begin to nucleate, grow, and coalesce.When microcracks and voids within the material reach a certain size, they also hinder the transmission of transient elastic waves from the material's interior to the surface, resulting in a decreasing trend in amplitude during this stage.

Macrocrack Formation and Fracture Stage
This stage involves the formation of macroscopic cracks and fracture.When cracks stably propagate to a certain length, they start to coalesce and converge.Once they reach a critical length, unstable propagation begins, leading to macroscopic crack formation and immediate unstable fracture.At this point, the material rapidly releases elastic waves outward in a short time frame, and the AE source becomes highly active, causing a rapid increase in the amplitude of the AE signal.

Intelligent Sensing of Defect Evolution Based on Transfer Learning
As stated in Section 2.4, the amplitude-time history plot of the AE signal was sliced based on the time sequence.Figure 9 shows a partial representation of amplitude history plots sliced along the time axis.This slicing method takes into account the temporal aspect, allowing the CNN to implicitly capture the time-series information.Each slice represents the AE signal collected in real time at a specific moment during the deformation of the TA15 specimen, encompassing its deformation process, thus contributing to the intelligent sensing of defect evolution.
In this study, AE detection was performed during the tensile testing of the five different-shaped TA15 specimens, with each type of specimen tested five times.While the general trends in the amplitude changes of AE signals were similar for repetitions of the same type of specimen, there were slight differences in specific details.Therefore, all amplitude-time history plots were sliced and included in the dataset to ensure that the subtle variations between different individuals would not affect the CNN's ability to identify the stages of TA15-forming defects evolution.Finally, a total of 4937 sliced images of TA15 AE signal  3. It's worth noting that due to the varying time lengths of each defect evolution stage, the unevenness in the number of labeled images for each defect stage is unavoidable in a real-time signal acquisition scenario.
Due to the limited number of experiments conducted in this study, the obtained training dataset of 3948 images is relatively small.However, pretrained CNNs can efficiently leverage such small image datasets for deep CNN training. [36,37]In this research, the pretrained VGG16, VGG19, GoogleNet, and Xception were employed.These models have undergone pretraining on millions of images, equipping them with strong image feature identify capabilities.By defining a new classifier for the pretrained CNN, it can be trained to identify the internal defect formation stages of TA15 corresponding to the amplitude history of the current AE signal.Three types of classifiers were defined in this study, each consisting of two layers of neurons: the first layer with 256, 512, and 1024 neurons, respectively, and using the Rectified Linear Unit (ReLU) activation function; the second layer serving as the output layer for the entire CNN, with only 1 neuron, employing the Softmax activation function.This output layer calculates the probabilities of the amplitude history image belonging to each internal defect formation stage.The cross-entropy loss function was chosen, and the RMSProp optimizer was selected.A comparison of their prediction results was conducted to identify the CNN model with the best performance.
Figure 10 illustrates the training loss and accuracy variations for different CNN models with classifiers consisting of 256 neurons, and similar patterns were observed for classifiers with different structures.It can be observed that GoogleNet exhibits the fastest learning rate, achieving a training accuracy of 95% within only 5 epochs and quickly reaching saturation.Xception also achieves a 95% accuracy in 10 epochs due to its unique parallel convolutional layer structure, which is not elaborated on in this paper.In contrast, VGG16 and VGG19, with linearly stacked convolutional layers, show relatively slower training progress but still reach satisfactory results after around 40  epochs.The prediction accuracy of different CNN models on the test dataset is presented in Table 4.
From Table 4, it can be observed that all CNN models, after training, achieve a prediction accuracy of approximately 97% on the test dataset with a margin of error of only about 1%.The bestperforming model is VGG19 with a classifier consisting of 1024 neurons.However, it's worth noting that increasing the number of neurons in the classifier does not significantly improve the prediction capabilities of the CNN model.In practical manufacturing processes, significantly more data can be collected compared to the dataset used in this study, which would further enhance the training of CNN models.Considering the computational cost, GoogleNet, which trains faster, could be a better choice.Furthermore, CNNs can make predictions on each image in just 9-15 ms.This indicates that after training, CNNs are capable of rapidly identifying the stage of internal defect evolution in TA15 from real-time collected AE signal amplitude graphs.In summary, pretrained CNNs can effectively provide real-time perception and sensing of defect evolution in metal plastic deformation processes.
It is noted that the proposed intelligent online sensing framework primarily focuses on defect detection during the plastic deformation process of metallic material, while this framework can potentially be adapted for defect detection under in-service conditions.But it cannot be directly applied to identifying defects in materials that already possess existing defects during service.This is due to the different boundary conditions and environments that materials experience during service compared to deformation conditions.Additionally, the morphology and distribution of the formed defects, which vary from those in the deformation process, can lead to different evolution patterns for subsequent defects.These differences would result in distinct outcomes when applying AE detection to materials in service.To adapt our framework for defect perception in materials with preexisting flaws during service, it would be necessary to collect AE signals produced under specific service conditions.Subsequently, studying the evolution of defects in the material during service, and reestablishing the correlation between AE signal variations and defect evolution, would allow for the use of CNN to identify the defect evolution stage based on AE signals.

Conclusion
This article systematically investigates the plastic deformation behavior, and AE signal characteristics of TA15 titanium alloy sheets under different stress conditions.It divides the defect evolution process into four stages based on the characteristics of AE signals for the first time and introduces a novel approach for intelligent online defects evolution sensing in metallic plastic deformation.The conclusions of this study are as follows.1) AE signal amplitudes from specimens under different stress conditions generally exhibit a pattern of two peaks followed by a rapid increase.The first peak corresponds to the yield point of TA15, while the second peak indicates the internal formation of microdefects.The period between the two peaks corresponds to uniform plastic deformation of TA15; 2) Based on the amplitude-time plots of AE signals, the defect evolution during TA15 deformation can be divided into four stages: 1) elastic deformation stage; II) uniform plastic deformation stage; III) initiation and propagation of microcracks stage; and IV) macrocrack formation and fracture stage; 3) Using transfer learning, several CNN models were established, including VGG16, VGG19, GoogleNet, and Xception.These models were then compared in their ability to identify the defect evolution stages based on amplitude-time plots from the test dataset.All models achieved an accuracy of around 97% with inference times below 15 ms.GoogleNet exhibited the fastest training accuracy increase, followed by Xception, while VGG16 and VGG19 had slower training convergence.The constructed CNN models can effectively identify the defect stages in AE signals, enabling intelligent online sensing of material plastic deformation processes.

Figure 1 .
Figure 1.The framework of intelligent online defects evolution sensing in metal plastic forming based on acoustic emission (AE) and machine learning: a) constructing the AE-defect dataset and b) training and applying the CNN model for intelligent online sensing.

Figure 3 .
Figure 3.The scanning electron microscopy (SEM) images at the gauge section edges of original specimens: a) shear, b) DB, c) CH, d) RN6.35, e) RN3.175, and f ) the EBSD results of raw grain morphology on original specimen.

Figure 4 .
Figure 4. AE signal collection during tensile deformation: a) schematic of the AE signal collection system for tensile deformation, b) the AE sensors mounting on the sample, and c) on-site AE signal collection.

Figure 5 .
Figure 5.The engineering stress-strain curves and the history of AE amplitudes during the tensile testing of the five different-shaped TA15 specimens: a) shear, b) DB, c) CH, d) RN6.35, and e) RN3.175.

Figure 6 .
Figure 6.The SEM images at the gauge section edges of specimens which were stretched until reaching the ascending phase before the second peak in the AE signal amplitude: a) shear, b) DB, c) CH, d) RN6.35, and e) RN3.175.

Figure 7 .
Figure 7. Macronecking (left) and microscopic edge morphology (middle and right) of different specimens when stopping tensile testing during the descent phase of the AE signal amplitude: a) shear, b) DB, c) CH, d) RN6.35, and e) RN3.175.

Figure 9 .
Figure 9. Partial time-amplitude history slices of the DB sample's AE signal: a) amplitude history at 80 s, b) amplitude history at 160 s, c) amplitude history at 240 s, and d) amplitude history at 330 s.

Figure 10 .
Figure 10.The training process of different convolutional neural network (CNN) models: a) loss and b) accuracy.

Table 1 .
The chemical composition of TA15.
Figure 2. Geometric shapes and dimensions of tensile specimens with different stress triaxiality values: shear represents shear specimens, DB represents dog bone specimens, CH represents specimens with a central round hole, RN6.35 represents specimens with a notch radius of 6.35 mm, and RN3.175 represents specimens with a notch radius of 3.175 mm.

Table 2 .
The time, true strain, and true stress of two peaks in AE signal amplitudes as well as fracture for the five TA15 specimens.

Table 3 .
The distribution of images divided by the defect evolution stage.

Table 4 .
The prediction accuracy of the CNN models on the test set.