TemporalAE-Net: A Self-Attention Framework for Temporal Acoustic Emission-Based Classification of Crack Types in Concrete

Featured Application

TemporalAE-Net provides real-time identification of tensile and shear cracks in concrete structures using acoustic emission signals, enabling early warning and improving structural safety.

Abstract

Crack type classification in concrete structures is essential for assessing structural integrity, yet traditional visual inspections and RA–AF parameter-based Acoustic Emission (AE) methods suffer from subjectivity and limited ability to capture temporal signal dependencies. This study proposes TemporalAE-Net, a self-attention-based machine learning framework designed to classify tensile and shear cracks while explicitly incorporating the temporal evolution of AE signals. AE data were collected from axial tension tests, shear-failure tests, and four-point bending tests on reinforced concrete beams, and a sliding-window reconstruction method was used to transform sequential AE signals into two-dimensional temporal matrices. TemporalAE-Net integrates one-dimensional convolution for local feature extraction and multi-head self-attention for global temporal correlation learning, followed by multilayer perceptron classification. The proposed model achieved an accuracy of 99.72%, outperforming both its ablated variants without convolutional or attention modules and conventional time-series architectures. Generalization tests on 12 unseen specimens yielded 100% correct classifications, and predictions for reinforced concrete beams closely matched established crack-evolution patterns, with shear cracks detected approximately 15 s prior to visual observation. These results demonstrate that TemporalAE-Net effectively captures temporal dependencies in AE signals. Moreover, it provides accurate and efficient tensile–shear crack identification, making it suitable for real-time structural health monitoring applications.

1. Introduction

Cracks are a ubiquitous phenomenon in concrete structures and are often used as a key indicator for assessing concrete quality [1,2]. In current civil engineering practice, periodic visual inspections are commonly used to identify and categorize cracks; however, such inspections are inherently subjective and restricted to accessible surface regions [3,4]. Although various sensor-based methods exist for detecting cracks—whether direct or indirect, contact or non-contact—they generally cannot provide detailed information regarding crack type (e.g., tensile or shear), which is crucial for predicting failure modes and evaluating structural stability.

Acoustic Emission (AE) technology has gained increasing attention for its ability to monitor the real-time condition of materials by detecting transient elastic waves generated by internal stress or deformation [5]. Compared with other non-destructive testing methods, AE offers two major advantages: (1) the signals originate from the material’s intrinsic responses rather than external excitation, and (2) AE captures dynamic processes such as motion and strain evolution, which are difficult to observe using most other techniques [6]. Consequently, AE has been widely applied for crack localization [7,8], damage quantification [9,10], and crack mechanism identification [11,12].

For crack type classification using AE signals, the commonly applied standard is JCMS-III B5706 (2003) [13], which primarily relies on average frequency (AF) and Rise time/Amplitude ratio (RA) to distinguish between tensile and shear cracks. However, no unified consensus exists regarding the optimal AF–RA thresholds. Ohno and Ohtsu demonstrated that effective crack identification can be achieved when the RA/AF ratio is within 1–200 [14], whereas Aldahdooh et al. suggested 1:80 as a suitable boundary for RC beams [15]. Such variability highlights the limitations of parameter-based methods. Specifically, the reported RA–AF thresholds differ by more than a factor of two across studies, and their optimal values are highly sensitive to specimen geometry, loading conditions, and material heterogeneity. This lack of robustness limits the transferability and reliability of RA–AF-based crack classification in practical applications. As a result, recent studies have explored machine-learning approaches for crack classification [15,16,17]. Mandal et al. reported that RA–AF methods are inadequate for heterogeneous materials like concrete and suggested that machine learning can provide more reliable crack identification [16]. Similarly, Ju et al. emphasized that RA–AF boundaries vary considerably among specimens and developed a semi-empirical machine-learning method to address this issue [17].

In the pursuit of more accurate crack type classification, machine learning—particularly self-attention mechanisms—offers promising potential [18,19]. As the core building block of Transformer architectures, self-attention has recently been adopted in structural health monitoring (SHM) research, including Vision Transformer-based approaches for crack detection [20,21]. Self-attention enables the model to assign importance weights to different positions in a sequence, allowing it to capture global dependencies beyond the limits of local-only feature extraction [22]. This capability represents a significant improvement over traditional machine-learning methods, which often treat AE signals as isolated instances and overlook temporal relationships. However, crack propagation is inherently time-dependent [23,24], and existing machine-learning studies have yet to fully incorporate temporal dependencies into AE-based crack classification.

To address this gap, this study proposes a self-attention-based framework—TemporalAE-Net—designed to classify crack types in AE signals while explicitly modeling temporal information. The framework integrates four key components: temporal reconstruction, one-dimensional convolution, self-attention mechanisms, and multilayer perceptrons. AE signals within each time window are reorganized into a two-dimensional temporal matrix as input. The 1D convolution extracts local waveform features without disrupting temporal order, while the self-attention module captures global temporal correlations. Finally, a multilayer perceptron produces the classification outputs. To validate the performance and generalization capability of the proposed model, tensile and shear tests on small specimens were conducted to generate training data, and four-point bending tests on reinforced concrete (RC) beams were employed for further verification.

2. Experimental Program and AE Data Preparation

2.1. Experimental Setup and Data Acquisition

To train TemporalAE-Net, two different types of experiments, namely axial tension tests on dog-bone-shaped specimens and shear failure tests on Z-shaped specimens, are performed. The dog-bone-shaped specimen has been extensively adopted for uniaxial tensile testing of concrete, as its geometry ensures uniform stress distribution within the gauge section and facilitates controlled crack initiation under tension [25,26]. Z-shaped specimens are widely employed in shear testing to generate a well-defined pure-shear stress field and capture the evolution of shear cracking [27,28,29]. The experimental designs were utilized to separately acquire AE signal samples for tensile and shear cracks. Detailed information regarding the instruments and equipment used is listed in

Table 1. Data acquisition equipment.

Item	Specification
Equipment	Physical Acoustics
Acquisition Card	PCI-8
Sensor	R6I
Acquisition Threshold	45 dB
Pre-Amplifier Gain	40 dB
Waveform length (samples)	1024
Pre-trigger Time	256 μs
Coupling Agent	502 adhesives

.

To accurately capture the AE signals of tensile cracks, specially designed dog-bone-shaped specimens made from Ultra-High-Performance Concrete (UHPC) were employed. The utilization of this material serves to prevent premature brittle failure of the specimens, thereby ensuring the acquisition of sufficient AE signals. Details of the specimens and their dimensions are elaborated as indicated in the relevant Figure 1a,c. To finely control the generation rate of AE signals, the axial tension tests were conducted under displacement control mode, with a set tensile rate of 0.1 mm/min.

For the acquisition of signals corresponding to shear cracks, a “Z-shaped” specimen made of a combination of normal concrete (NC) and UHPC was designed. This configuration was intentionally selected to simulate realistic strengthening scenarios, where UHPC is commonly applied as a shear-strengthening layer or overlay for existing RC structures. Therefore, the composite shear specimens better represent practical UHPC–concrete interaction under shear-dominated failure modes. The shape and dimensions of this design are depicted as shown in the relevant Figure 1b,d. The experimental procedure also adopted a displacement control mode and was set at a loading rate of 0.5 mm/min. The datasets obtained from two types of experiments are presented in

Table 2. Dataset description.

Item	Tensile	Shear
Sample Type	Tensile	Shear
Number of Specimens	35	11
Pre-processed AE Signal Count	440,402	136,787
Pre-processed Sample Shape	1024 × 1	1024 × 1
Window Size	100	100
Window Interval	60	20
Post-processed Sample Shape	1024 × 100	1024 × 100
Post-processed Sample Count	7287	6791

.

2.2. Data Preprocessing and Temporal Reconstruction

To account for the temporal correlations between AE signals, we implemented a unique data preprocessing strategy that assembles these signals into a two-dimensional matrix to serve as the input for network training. Specifically, we defined a “sliding window” along the time axis that encompasses 100 AE signals and used it to generate a two-dimensional matrix. The window then shifts incrementally at predetermined intervals to produce two-dimensional reassembled samples. Each of these reassembled samples thus encapsulates all the AE signals within the window, enabling TemporalAE-Net to capture temporal information.

Due to the inherent imbalance between the tensile and shear crack datasets—where the number of shear-related AE signals is substantially smaller than that of tensile signals—a smaller window interval was adopted for shear samples during temporal reconstruction. Specifically, the window interval was set to 60 for tensile samples and 20 for shear samples, thereby increasing the number of reconstructed shear samples while maintaining temporal continuity. Table 2 summarizes the dataset composition before and after preprocessing.

To verify that the use of different window intervals does not introduce reconstruction-induced statistical bias or artificial artifacts, additional control experiments were conducted within the same crack category. Specifically, the shear dataset was reconstructed using two different window intervals (20 and 60), and key AE features—including signal energy (RMS), peak amplitude, and average frequency—were extracted from the reconstructed samples. Distribution-level comparisons were then performed using the two-sample Kolmogorov–Smirnov (KS) test.

As illustrated in Figure 2a–c, the reconstructed shear samples obtained with different window intervals exhibit highly overlapping feature distributions. The corresponding KS statistics remain extremely small (Energy: D = 0.0299, Peak amplitude: D = 0.02499, Average frequency: D = 0.0464), with non-significant p-values (p > 0.05) in all cases. These results confirm that reducing the window interval for shear samples increases the number of training samples without altering the intrinsic statistical characteristics of AE signals, thereby avoiding reconstruction-induced artifacts or artificial inflation of classification performance.

Therefore, the adopted temporal reconstruction strategy effectively mitigates dataset imbalance while preserving the physical integrity of AE features, ensuring that the classification performance of TemporalAE-Net is driven by genuine crack-related characteristics rather than reconstruction-induced statistical artifacts.

2.3. Generalization Capability Validation

Generalization capability measures the predictive accuracy of the model when dealing with unfamiliar data. To comprehensively assess the generalization performance of TemporalAE-Net, we conducted validations from two distinct perspectives: small-scale specimens and four-point loading tests on RC beams. It’s worth noting that the small-scale specimen data used for this validation set were generated under the same experimental settings as described in Section 2.1; however, they were not used for model training or within-training testing.

Table 3. Small Specimen Dataset for Generalization Performance Validation.

Specimen	Crack Type	Sample Count
1	Shear	1045
2	Shear	33,207
3	Shear	2509
4	Tensile	27,733
5	Tensile	11,022
6	Tensile	21,101
7	Tensile	18,300
8	Tensile	16,579
9	Tensile	19,159
10	Tensile	2849
11	Tensile	27,581
12	Tensile	13,116

provides a detailed listing of the 12 specimens used for generalization capability validation.

As shown in Figure 3, the shape and loading method of the RC rectangular beams conform to the standard four-point loading test for RC rectangular beams. The dimensions of the RC rectangular beam were 3000 mm × 450 mm × 200 mm. The main structural components consisted of C30 concrete and HRB400 longitudinal reinforcement. The diameters of the top tensile reinforcement and bottom tensile reinforcement were 20 mm and 22 mm, respectively. The tests are performed using a microcomputer-controlled electro-hydraulic columnar pressure testing machine, with a loading rate of 1 mm/min. AE sensors were placed at the mid-span location of the RC rectangular beams. For the four-point loading tests on the rectangular beams, particular attention is paid to the variations in flexural-shear crack characteristics under different loading phases, a point explicitly elucidated in previous research [30]. Given that these crack characteristics have a significant impact on the overall structural performance of the RC rectangular beams, they introduce additional complexities for the verification of generalization capabilities.

3. Architecture and Methodology of TemporalAE-Net

TemporalAE-Net employs one-dimensional convolutional layers and attention mechanisms to achieve classification of crack patterns. The architecture of TemporalAE-Net is elaborately depicted in Figure 4. TemporalAE-Net takes preprocessed samples as described in Section 2.2 as its input and produces outputs for a binary classification task, distinguishing between tensile cracks and shear cracks.

To improve reproducibility, a systematic hyperparameter exploration was conducted prior to final model selection. Key architectural parameters, including convolution kernel size, number of convolution filters, and the number of attention heads, were evaluated over discrete candidate sets, as summarized in

Table 4. Hyperparameter search space and final selection for TemporalAE-Net.

Hyperparameter	Candidate Values	Selected Value
Convolution kernel size	3, 5, 7	3
Number of convolution filters	3, 5, 10, 15	10
Number of attention heads	3, 5, 10, 15	10

. A grid-based search strategy was adopted, and configurations were compared based on validation accuracy and training stability. The final hyperparameters were selected as a balanced trade-off between classification performance, model complexity, and computational efficiency.

3.1. One-Dimensional Convolution

To preserve the temporal continuity of the AE signals while extracting discriminative features for crack type classification, TemporalAE-Net employs one-dimensional (1D) convolutional operations at the initial stage of the network. A convolution kernel of size 3 × 1 with 10 filters was selected based on parameter tuning and preliminary experiments, ensuring an effective balance between feature extraction ability and computational efficiency.

Given an input temporal matrix $X\in {ℝ}^{T\times L}$, where T denotes the number of AE signals within a temporal window and L is the signal length (1024 points), the 1D convolution operation applies a learnable kernel $W\in {ℝ}^{3\times 1}$ across each row. For an input segment $x_{i:i+k-1}$, the convolutional output is computed as

$$z_i=ReLU({\displaystyle \sum _{j=0}^{k-1}{W}_j}\cdot x_{i+j}+b),$$

(1)

where k = 3 is the kernel size, b is the learnable bias term. To introduce non-linearity and improve representation capability, the Rectified Linear Unit (ReLU) activation is applied:

$$ReLU(x)=max(0,x).$$

(2)

The convolution is performed with a stride of 1, allowing the kernel to slide across the temporal matrix and capture local temporal patterns that may correspond to critical AE waveform characteristics. The resulting feature maps encode short-term dependencies while preserving the global temporal ordering.

To further reduce computational cost and mitigate overfitting, a max-pooling operation with a window size of 2 × 1 and a stride of 1 is incorporated after the convolution layer. Max-pooling compresses the feature maps by retaining only the most prominent response within each local neighborhood:

$$p_i=\max \left(z_i,z_{i+1}\right),$$

(3)

thereby emphasizing salient local features while suppressing noise and redundant information.

Through this combination of 1D convolution and max-pooling, the network efficiently extracts localized temporal features essential for distinguishing between tensile and shear cracks.

3.2. Attention Mechanism

The attention mechanism represents a key component of TemporalAE-Net, enabling the extraction of global dependencies across the temporal dimension of AE signals. While one-dimensional convolution primarily captures local features within short temporal neighborhoods, self-attention allows the network to adaptively assign different weights to different positions in the input sequence. This capability is particularly advantageous for AE-based crack type classification, as tensile–shear transition behaviors and waveform characteristics often exhibit long-range correlations across time.

To enhance feature representation capability, a Multi-Head Self-Attention structure is adopted. TemporalAE-Net employs two stacked self-attention layers, each configured with 10 attention heads (num_heads = 10) and a key dimension of 100 (key_dim = 100), based on empirical tuning to optimize performance. For an input sequence X, the multi-head attention operation is defined as:

$$\mathrm{MultiHead}(Q,K,V)=\mathrm{Concat}({\mathrm{head}}_1,{\mathrm{head}}_2,\dots ,{\mathrm{head}}_h)W_O,$$

(4)

where each attention head is computed as:

$${\mathrm{head}}_i=\mathrm{Attention}(Q{W}_Q^{(i)},K{W}_K^{(i)},V{W}_V^{(i)}).$$

(5)

The scaled dot-product attention mechanism is expressed as:

$$\mathrm{Attention}(Q,K,V)=\mathrm{softmax}({\displaystyle \frac{Q{K}^{\top }}{\sqrt{d_k}}})V,$$

(6)

where Q, K, and V denote the query, key, and value matrices, $W_Q^{(i)},W_K^{(i)},W_V^{(i)}$ are learnable projection matrices for head i, $W_O$ is the output projection matrix, and $d_k$ the key dimension used for normalization.

Through this mechanism, the network learns to emphasize temporally important AE signals—such as those containing high-energy events or precursors of shear behavior—while suppressing irrelevant or noisy segments. This ability to model long-range temporal interactions significantly enhances the accuracy and robustness of crack type classification.

3.3. Model Training and Evaluation

Following the temporal reconstruction process, the dataset was randomly divided into a training set (80%) and a test set (20%) to enable supervised learning and unbiased performance assessment. The model parameters were optimized using the Adam algorithm, while the cross-entropy loss function was adopted as the objective function for updating the network weights during training.

To reduce the risk of overfitting and enhance model generalization, an early-stopping strategy was applied. Training was terminated when the validation loss failed to decrease for three consecutive epochs, preventing unnecessary weight updates and ensuring that the model does not overadapt to the training data. Once early stopping was triggered, the best-performing model was preserved and subsequently evaluated using the independent test set.

Model performance was comprehensively assessed using multiple standard classification metrics, including accuracy, recall, precision, F1-score, and the area under the ROC curve (ROC-AUC). These metrics provide complementary insights into the model’s predictive capability, robustness to data imbalance, and sensitivity to shear crack identification—an aspect of particular importance for structural safety. Additionally, training time was recorded to evaluate the computational efficiency and practical applicability of the model. A summary of these evaluation metrics is presented in

Table 5. Performance Metrics and Their Descriptions.

Performance Metric	Description	Application or Impact
Accuracy	Measures the overall proportion of correctly classified samples.	High accuracy indicates reliable overall performance in AE crack classification.
Recall	Measures the proportion of actual positive cases correctly identified.	High recall is essential for detecting shear cracks, which pose greater structural risks if missed.
Precision	Measures how many predicted positive cases are truly positive.	High precision reduces false alarms, improving practical applicability in monitoring systems.
F1 Score	Harmonic mean of precision and recall.	Provides a balanced evaluation when dealing with class imbalance.
ROC-AUC	Area under the Receiver Operating Characteristic curve.	Indicates model robustness under varying classification thresholds, especially useful for imbalanced datasets.
Training Time	Total time required for model training.	Reflects computational efficiency and feasibility for real-world deployment.

.

To further analyze the contribution of individual network components, two ablation models were developed:

• TemporalAE-Net-no1DC, in which the one-dimensional convolutional layer was removed and the input was fed directly into the attention module;
• TemporalAE-Net-noATT, in which the attention mechanism was omitted entirely.

These ablation studies enable a systematic examination of the respective roles of convolutional feature extraction and attention-based global correlation learning, providing deeper insight into their contributions to overall model performance.

4. Results and Discussion

4.1. TemporalAE-Net Performance

To comprehensively evaluate the performance of the model, various performance metrics were employed, as shown in

Table 6. Performance Evaluation of TemporalAE-Net.

Performance Metric	Accuracy	Recall	Precision	F1 Score	ROC-AUC
Value	0.9972	0.9970	0.9970	0.9970	0.9972

. The model’s accuracy, recall, precision, and F1 score are all nearing 100%, indicating exemplary performance in this task. As demonstrated in Figure 5a,b, the model exhibits excellent performance in terms of both loss and accuracy during the training and validation phases. Notably, both the training loss and validation loss exhibit a decreasing trend, while training accuracy and validation accuracy show an increasing trajectory. It is worth noting that the validation accuracy exceeds 0.97 within the first training epoch, indicating rapid convergence of the proposed model. Therefore, the early-stopping criterion with a patience of three epochs is considered sufficient to prevent overfitting while avoiding unnecessary training iterations.

Figure 5c presents the confusion matrix, revealing nearly flawless performance across two categories, namely, tensile and shear. Specifically, the model misclassified only eight instances. Given the exceedingly low number of misclassified instances, these errors are more likely attributable to data noise or outliers rather than systematic model deficiencies.

4.2. Sensitivity Analysis of Window Size Selection

To further investigate the influence of window size on temporal information capture and classification performance, a sensitivity analysis was conducted by evaluating three representative window sizes: 60, 100, and 150 AE signals. As shown in Figure 6, the classification accuracy increases significantly when the window size is enlarged from 60 to 100, improving from 0.9232 to 0.9972. This indicates that a window size of 60 is insufficient to fully capture the temporal evolution of AE activities associated with crack propagation.

Further increasing the window size from 100 to 150 yields only a marginal accuracy improvement (from 0.9972 to 0.9984). However, this slight gain is accompanied by a substantial increase in GPU memory consumption. This behavior is primarily attributed to the self-attention mechanism employed in TemporalAE-Net, whose memory complexity scales quadratically with the sequence length due to the construction of attention score matrices. As a result, excessively large window sizes significantly increase computational cost without providing commensurate performance benefits.

Considering both classification accuracy and computational efficiency, a window size of 100 AE signals was selected as an optimal trade-off. This configuration provides sufficient temporal context for reliable crack type discrimination while maintaining feasible memory requirements for practical deployment.

4.3. Advantages of TemporalAE-Net

4.3.1. Performance Advantages

Although Section 4.1 has demonstrated the strong performance of TemporalAE-Net, the ablation results further highlight the importance of its core components. As shown in Figure 7a, the two variant models—TemporalAE-Net-no1DC and TemporalAE-Net-noATT—exhibit substantial declines across all evaluation metrics, including accuracy, recall, precision, F1 score, and ROC-AUC. Their ROC-AUC values in particular approach near-random classification, underscoring the critical roles of both the one-dimensional convolution layer and the attention mechanism in effective crack type discrimination.

Figure 7b further compares the training behavior of the three models. TemporalAE-Net shows a stable and steadily increasing accuracy trend for both training and validation sets, with no evidence of overfitting or underfitting. In contrast, the two ablation variants display unstable or stagnant accuracy curves, with TemporalAE-Net-no1DC requiring significantly longer training time. These observations confirm that the full model not only achieves superior predictive performance but also maintains better training stability and computational efficiency.

4.3.2. Comparison with Baseline Machine Learning Models

To further evaluate the effectiveness of the proposed TemporalAE-Net, its performance was compared with several representative baseline models commonly used for time-series and AE signal classification, including a purely convolutional network (1D-CNN), a Long Short-Term Memory network (LSTM), and a Gated Recurrent Unit network (GRU). All baseline models were trained and evaluated using the same dataset split and preprocessing strategy to ensure a fair comparison.

The Baseline 1D-CNN adopts a hierarchical convolutional architecture consisting of three one-dimensional convolutional layers with increasing channel numbers (32, 64, and 128) and decreasing kernel sizes (7, 5, and 3). Each convolutional layer is followed by max-pooling for temporal downsampling, and a global average pooling layer is applied before the fully connected layers.

The LSTM and GRU baseline models are designed to capture sequential dependencies through recurrent units. However, directly processing the full temporal length (1024 time steps) would result in excessive computational cost and memory consumption. Therefore, both models employ an initial convolutional layer with 32 filters and a kernel size of 5, followed by max-pooling with a factor of 4, reducing the effective sequence length from 1024 to 256. A single recurrent layer with 64 hidden units is then applied, followed by a fully connected layer with dropout for classification.

Table 7. Performance comparison between TemporalAE-Net and baseline models.

Model	Validation Accuracy	Number of Parameters	Inference Speed (Samples/s)
TemporalAE-Net	0.9972	142,342	801
Baseline 1D-CNN	0.9486	65,761	925
Baseline LSTM	0.9373	45,089	849
Baseline GRU	0.9196	39,073	810

summarizes the quantitative comparison results, including validation accuracy, number of trainable parameters, and inference speed. Compared with the baseline models, TemporalAE-Net achieves the highest classification accuracy (0.9972) while maintaining a moderate model size. Although the 1D-CNN baseline exhibits faster inference speed due to its simpler structure, its classification accuracy is noticeably lower, indicating limited capability in capturing global temporal relationships. The LSTM and GRU models show inferior performance in both accuracy and efficiency, suggesting that recurrent architectures alone are less effective for modeling the complex temporal evolution patterns present in AE signals.

Overall, these results demonstrate that the proposed TemporalAE-Net provides a better balance between accuracy, model complexity, and inference efficiency, benefiting from the combination of convolution-based local feature extraction and self-attention-based global temporal modeling.

4.3.3. Interpretability Analysis of TemporalAE-Net

Although deep neural networks are often criticized for their “black-box” nature, the design of TemporalAE-Net incorporates components—such as one-dimensional convolution and self-attention—that enable a certain degree of interpretability. Understanding how these modules process AE signals is essential for building trust in the model, especially in critical engineering applications like structural health monitoring.

Figure 8 presents a comparison between the input temporal matrix and the output of the 1D convolutional layer for a representative sample. The convolution operation compresses local waveform segments while retaining the global temporal order. Notably, salient temporal features—such as peaks or abrupt changes in signal amplitude—remain aligned in time after convolution (e.g., at time positions 37 and 83). This demonstrates that the 1D convolution not only extracts meaningful local patterns but also preserves the temporal structure necessary for interpreting the evolution of AE events. Such preservation is critical, as the initiation and propagation of cracks are inherently time-dependent phenomena.

Complementing the convolutional analysis, Figure 9 visualizes the attention-weight matrix generated by the multi-head self-attention mechanism for the same input sample. The heatmap reveals that the attention module assigns significantly higher weights to specific time positions—again, notably around positions 37 and 83—indicating that the model actively focuses on these high-information regions. This behavior aligns with human expert intuition, as analysts typically concentrate on strong-amplitude or energy-dense AE events when differentiating between tensile and shear crack activities.

Together, Figure 8 and Figure 9 reveal that TemporalAE-Net possesses an interpretable internal decision process:

• The convolution layer amplifies salient waveform characteristics while preserving their temporal alignment.
• The attention mechanism further identifies and prioritizes key AE signals, effectively suppressing noise and irrelevant information.

These observations demonstrate that TemporalAE-Net not only achieves high classification accuracy but also learns meaningful temporal patterns consistent with physical cracking behaviors. This interpretability strengthens the model’s reliability and suitability for real-time structural health monitoring tasks.

Nevertheless, the current interpretation is primarily qualitative, and the correspondence between attention weights and specific physical damage mechanisms has not yet been quantitatively established. Future work will focus on systematically linking attention responses to detailed AE source mechanisms—such as aggregate bridging, frictional sliding, or fiber pull-out—through combined experimental observations and physics-informed analysis, thereby further enhancing the interpretability and generalizability of the proposed framework.

4.4. Generalization Capability

Generalization capability is a critical indicator of a model’s robustness, reflecting its performance on previously unseen data. To thoroughly evaluate the generalization ability of TemporalAE-Net, two levels of validation were conducted: (1) small-scale specimens not involved in training and (2) four-point bending tests on reinforced concrete (RC) beams exhibiting mixed crack modes.

4.4.1. Validation Using Small-Scale Specimens

Table 8. Classification Results on Small Specimens.

Specimen	Actual Type	Sample Count	Predicted Shear	Predicted Tensile	Prediction Time (s)
1	Shear	1045	1045	0	10.45
2	Shear	33,207	33,207	0	41.47
3	Shear	2509	2509	0	11.67
4	Tensile	27,733	0	27,733	43.31
5	Tensile	11,022	0	11,022	23.56
6	Tensile	21,101	0	21,101	35.30
7	Tensile	18,300	0	18,300	31.70
8	Tensile	16,579	0	16,579	28.76
9	Tensile	19,159	0	19,159	33.71
10	Tensile	2849	0	2849	4.03
11	Tensile	27,581	0	27,581	43.64
12	Tensile	13,116	0	13,116	25.08

summarizes the classification results for the 12 small-scale specimens described in Section 2.3. All specimens used in this validation were excluded from model training and testing and were obtained from independent experimental runs under the same loading configurations. TemporalAE-Net achieved a 100% classification accuracy on these unseen specimens, indicating strong generalization capability within the investigated experimental conditions. All small-scale specimens were obtained from independent tests conducted under the same experimental protocol ensuring specimen-level independence. In addition, A 95% Wilson CI was computed at the sample level; the lower bound remained above 99.9%.

These results confirm that the model achieves a strong balance between model complexity and generalization performance, avoiding both overfitting and underfitting tendencies. Nevertheless, it should be noted that the reported 100% classification accuracy on the small-scale specimens is obtained from a limited number of experimental tests conducted under controlled laboratory conditions. Although these specimens were not used during model training, they were prepared following similar procedures and loading configurations, which may introduce potential correlations. In addition, the finite sample size implies that the observed accuracy is subject to statistical uncertainty. To address this, confidence intervals were estimated for the classification results, and future work will focus on validating the proposed framework using larger and more diverse datasets obtained from different batches, material compositions, and loading conditions.

4.4.2. Validation Using RC Beam Tests

To further evaluate model robustness under more complex structural behavior, TemporalAE-Net was tested on AE data obtained from four-point bending tests on RC rectangular beams. Due to the coexistence of tensile and shear cracks during beam failure, this dataset provides a more challenging validation scenario.

Figure 10a shows the load–time response of the tested RC beam, highlighting key cracking stages:

• Stage 1: Micro-crack formation
• Stage 2: Appearance of first visible crack (Point A, 1065 s)
• Stage 3: Emergence of distributed tensile and shear cracks (Point B, 1400 s)
• Stage 4: Localized damage and final failure (Point C, 1800 s)

Figure 10b–d further illustrate the crack propagation patterns corresponding to these stages.

Figure 11 depicts the predicted shear-crack probability over time. A classification threshold of 0.5 was used: values above 0.5 indicate shear cracks, while lower values indicate tensile cracks. The threshold value of 0.5 was selected as the default decision threshold for sigmoid outputs. This choice provides an intuitive and consistent criterion for binary crack type classification and is commonly adopted in probabilistic classification tasks. Additional threshold sensitivity analyses (e.g., precision–recall trade-offs) may further refine decision boundaries for specific engineering applications and will be explored in future work.

The model outputs reveal the following:

• Stage 1: All AE signals are classified as tensile, consistent with the pure flexural behavior observed experimentally.
• Stage 2: TemporalAE-Net predicts the first shear crack at 1385 s, which is approximately 15 s earlier than visual detection (at 1400 s). This early detection is likely due to the model capturing subtle precursors of shear behavior that are difficult to observe manually.
• Stages 3 & 4: Increasing crack complexity results in a mixed pattern of tensile and shear predictions, consistent with the expected mechanical behavior of RC beams approaching failure.

Table 9. Comparison of Crack Type Proportions with Previous Research.

Stage	Tensile (%) (Model)	Shear (%) (Model)	Tensile (%) (Ref.)	Shear (%) (Ref.)
1	100.00	0.00	100.00	0.00
2	99.90	0.10	100.00	0.00
3	85.24	14.76	87.96	12.04
4	74.72	25.28	82.90	17.10

compares the predicted tensile–shear proportions at different stages with the experimental findings of Aldahdooh et al. [30,31]. The close agreement between predictions and established results demonstrates the physical consistency and reliability of the model.

The model’s predictions closely match documented crack evolution trends, reinforcing that TemporalAE-Net effectively captures both the onset and progression of tensile and shear cracking under realistic structural loading conditions.

Although the present study focuses on four-point bending tests of rectangular RC beams, the proposed TemporalAE-Net framework is not inherently restricted to this loading configuration. Since the model relies on temporal patterns of AE signals rather than geometric crack features, it has the potential to be extended to other loading regimes, such as cyclic fatigue or impact loading, where crack initiation and evolution are also accompanied by characteristic AE sequences. In such scenarios, the self-attention mechanism may be particularly beneficial for identifying critical AE events within highly non-stationary signal streams.

Furthermore, the framework is expected to be applicable to different concrete types, including fiber-reinforced or ultra-high-performance concrete, although material heterogeneity may alter AE signal characteristics. In these cases, retraining or fine-tuning with material-specific data would be required to ensure reliable performance. It should be noted that the current validation is limited to a specific loading mode and specimen geometry, and future studies will focus on expanding the dataset to encompass a wider range of structural configurations and loading conditions.

4.5. Potential Deployment and Integration with SHM Systems

Although the present study primarily focuses on model development and experimental validation, the proposed TemporalAE-Net framework is designed with potential practical application in mind and can be integrated into existing structural health monitoring (SHM) systems. A conceptual deployment workflow is illustrated in Figure 12 to demonstrate how the proposed method could be incorporated into an AE-based SHM framework. In such a scenario, AE sensors installed on concrete structures continuously acquire raw waveform data, which are processed by standard AE acquisition units to generate time-series signals. These signals are subsequently reconstructed into temporal matrices using a sliding-window strategy and fed into the trained TemporalAE-Net model for real-time crack type classification.

From a computational perspective, TemporalAE-Net exhibits moderate model complexity and high inference efficiency, indicating that near-real-time processing could be achieved on commonly available hardware, such as industrial PCs or edge-computing devices equipped with GPUs or high-performance CPUs. Depending on system configuration and computational resources, crack type identification may therefore be performed either locally at the monitoring site (edge inference) or remotely via cloud-based SHM platforms.

With respect to system integration, the proposed framework is intended to operate as a modular analytical component within existing SHM architectures. As illustrated in Figure 12, the model outputs—namely, probabilistic predictions of tensile and shear cracking—can be transmitted to a central monitoring interface or digital twin system. Such outputs have the potential to support condition assessment, early warning, and decision-making by providing mechanism-oriented information, thereby complementing conventional threshold-based AE parameters.

Despite this potential for integration, several challenges should be considered for real-world SHM applications, including variations in sensor layouts, environmental noise, material heterogeneity, and structural configurations, which may lead to discrepancies between laboratory and field AE data. These challenges could be mitigated through site-specific calibration, model fine-tuning using limited field data, or transfer learning strategies to improve adaptability across different monitoring scenarios.

5. Conclusions

This study proposed TemporalAE-Net, a novel machine-learning framework incorporating temporal reconstruction, one-dimensional convolution, and multi-head self-attention mechanisms for classifying tensile and shear cracks from Acoustic Emission (AE) signals. Based on comprehensive experimental validation—including small-scale specimens and full-scale RC beam tests—the following conclusions can be drawn:

• TemporalAE-Net achieves highly accurate crack type classification through effective temporal feature learning. By combining temporal reconstruction, 1D convolution, and self-attention, the model captures both local waveform features and long-range temporal dependencies. It reaches 0.9972 accuracy, 0.9970 precision/recall, and an ROC-AUC of 0.997, misclassifying only eight samples, while visualization confirms that it focuses on physically meaningful temporal positions.
• The model demonstrates strong generalization capability across specimen types and loading scales. TemporalAE-Net correctly classified 100% of over 180,000 unseen AE signals from small specimens. In RC beam tests, its predictions aligned with documented crack evolution and even identified shear cracks ~15 s earlier than visual observation.
• TemporalAE-Net provides high computational efficiency suitable for real-time monitoring. The model processes approximately 800 AE samples per second, classifying 33,207 samples in 41.47 s, enabling real-time application. Ablation studies show that removing convolution or attention components leads to severe performance degradation, confirming their necessity for reliable and efficient crack classification.

Nevertheless, this study still has certain limitations. First, the experimental dataset primarily involves a limited range of concrete materials (UHPC for tensile specimens and composite concrete for shear specimens), which may influence the direct transferability of the proposed model to other concrete types or material configurations. Future work will therefore focus on extending the validation to a broader range of materials, such as normal-strength concrete, fiber-reinforced concrete, and strengthened composite systems, to further assess the robustness and transferability of the proposed framework.