Acoustic emission-based intelligent identification and dual early warning for coal fatigue failure under multistage cyclic loading Authors: Luyi Xing, Chao Wang, Min Liang, Baoliang Zhang Published: 2026 ABSTRACT AbstractTo address the limitations of insufficient accuracy and timeliness in the early warning of coal instability under multistage cyclic loading, this paper proposes an intelligent recognition and dual early warning method for coal damage based on acoustic emission parameters. The multistage cyclic compression test of coal is performed to obtain five acoustic emission parameters, and the average event intensity (ERR) is used to characterize the sudden changes in coal energy release. The Sparrow Search Algorithm (SSA) is adopted to optimize the hybrid model of the Transformer and Gated Recurrent Unit (GRU) to achieve high-precision recognition of coal damage stages. On this basis, the early warning coefficient (EW) is calculated by using the Isolation Forest algorithm and the CRITIC-TODIM method, and a dual early warning system of “preliminary prompt with ERR and comprehensive confirmation with EW” is established. The results show that the recognition accuracy of the optimized model for the fourth damage stage is 11.7% higher than that of the SSA-GRU model. The early warning lead time of ERR is between 60.5 and 185.7 s, and that of EW is between 10.5 and 101.5 s. The proposed method integrates physical mechanisms and data-driven technology, effectively improving the accuracy and reliability of coal damage early warning under multistage cyclic loading and providing theoretical support and technical reference for the early warning of dynamic disasters in deep coal mines.Keywords: Coal, Multi-level cyclic loading, Acoustic emission, Damage identification, Machine learning, Average event intensitySubject terms: Energy science and technology, Engineering, Natural hazards Introduction As shallow mineral resources dwindle, major coal-mining countries worldwide are increasingly transitioning from shallow to deep mining operations1–3. In deep environments characterized by high in-situ stress, high-intensity mining significantly expands the influence range of mining-induced stress on coal masses. These conditions make coal more susceptible to fatigue damage and failure, potentially leading to dynamic disasters such as rock bursts and coal and gas outbursts, which pose serious threats to mine safety. Consequently, systematically studying the precursor information of coal fatigue failure under cyclic loading has become crucial. Accurately identifying and capturing these failure precursor characteristics is essential for providing technical support for early warning systems aimed at preventing coal instability. This research area is a critical and forward-looking topic in mining engineering. Acoustic emission (AE) technology, a well-established nondestructive testing method, is widely used in rock mechanics research. It has been particularly effective in revealing the stress and failure characteristics of rocks4–6. In studying rock behavior under cyclic loading, AE technology has become crucial for identifying precursors to rock failure and understanding damage evolution laws. Wang et al.7 conducted cyclic loading tests on water-bearing grotto sandstone and discovered that changes in the correlation dimension of AE absolute energy effectively identify instability precursors. They observed that when the water content in sandstone ranges from 2.63 to 3.93%, crack activity shifts from low-frequency order to high-frequency disorder. Water accelerates damage accumulation and significantly reduces the time to instability. Chen et al.8 performed cyclic loading and unloading experiments and noted that sandstone failure is typically accompanied by numerous high-amplitude, high-count AE events. They identified the transition of wave velocity from increasing to decreasing as a critical precursor. They established a linear relationship between wave velocity reduction and dissipated energy, providing a basis for quantitative damage assessment. Zhang et al.9 found, through cyclic loading and unloading experiments, that an instability precursor index constructed from the energy rate difference, characterized by its peak value and post-peak decline rate, can quantitatively reflect the proximity to failure in sandstone with holes. This index correlates highly with AE characteristics and exhibits more significant and sensitive precursor identification capability than traditional AE parameters. Song et al.10 performed triaxial confining pressure unloading and differential cyclic loading tests and found that AE activity in sandstone increases significantly after confining pressure unloading. They reported that the minimum radial-to-axial energy consumption ratio serves as a failure precursor, and that the confining pressure unloading rate plays a dominant role in determining the triggering time and energy release mode of AE. Shen et al.11 observed in cyclic loading and unloading experiments that, at low stress levels, more low-frequency, high-amplitude AE signals appear before sandstone failure, with the b-value initially increasing and then decreasing. In contrast, high lower stress limits suppress AE activity. The rapid advancement of machine learning has significantly propelled interdisciplinary research. Traditional analysis methods often falter when dealing with the high-dimensional and voluminous data from AE tests, making it challenging to uncover deep-level evolutionary laws. Consequently, researchers have turned to machine learning techniques to better reveal the complex internal mechanisms underlying rock damage evolution12–14. Wu et al.15 conducted uniaxial loading tests on tight sandstone and shale, with loading orientations perpendicular and parallel to the bedding. Utilizing AE technology and unsupervised machine learning, they categorized waveforms into three micro-fracture modes, including tensile, shear, and mixed. Their findings indicated that tensile fractures predominate at low stress levels, transitioning to shear fractures near peak stress. Additionally, they discovered that the bedding structure strength of tight sandstone surpasses that of shale, offering valuable insights for reservoir development. Di et al.16 introduced a deep-learning early-warning method for coal mine rock bursts, integrating microseismic, AE, and electromagnetic radiation data. By employing Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), they intelligently identified precursors in multi-source signals and predicted risk coefficients. Field verification demonstrated that this method provides earlier and more accurate warnings than those relying on a single signal. Song et al.17 performed true triaxial compression tests and AE monitoring on sandstone. By combining transfer learning with a CNN, they discovered that introducing the concept of waveform sequence length (specifically 10) enhances model recognition accuracy for the rock fracture stage to 90.4%. This finding underscores that fracture is a continuous process with a sequence correlation, providing a new method for automatic monitoring of deep-engineering rock-mass fractures. Xie et al.18 conducted uniaxial compression tests on granite specimens with persistent, serrated joints at varying strain rates and joint dip angles. They proposed a method for obtaining AE precursors using waveform clustering and CNN. Their study revealed that traditional indicators are not universally applicable to jointed rock masses, and effective fracture precursor waveforms are characterized by low frequency and high energy. Employing Birch clustering and the s-value indicator can more accurately identify precursors, making this method applicable to three types of failure modes and providing a novel approach for early warning in cross-fault engineering. In summary, AE technology effectively studies the evolution of rock damage and the characteristics of failure precursors under cyclic loading. While machine learning excels at processing high-dimensional, nonlinear AE data, research systematically applying it to analyze rock AE behavior under cyclic loading remains limited, offering significant opportunities for exploration. To address this gap, this study proposes a hybrid classification model (SSA-Transformer-GRU) based on the Sparrow Search Algorithm (SSA), the Transformer, and the Gated Recurrent Unit (GRU). This model accurately identifies coal damage stages using multi-dimensional AE parameters. The Local Interpretable Model-agnostic Explanations (LIME) method is introduced to improve model interpretability. On this basis, an early warning system integrating outlier detection and dynamic comprehensive evaluation is further constructed. The Isolation Forest algorithm is used to identify abnormal signals in the fourth damage stage (fatigue failure). The early warning coefficient (EW) is calculated by combining the CRITIC objective weighting method and the TODIM dynamic decision-making method, and an early warning is triggered when EW exceeds 100. Meanwhile, the average event intensity (ERR) index is proposed to characterize energy release, forming a dual early warning mechanism that combines ERR and EW. This study systematically reveals the evolutionary law of coal from micro-damage initiation and stable propagation to macroscopic instability and failure, and realizes the dynamic identification and early warning of coal damage states under multi-level cyclic loading. It provides a data-driven solution for the monitoring and early warning of dynamic disasters in deep engineering, with sound engineering applicability and promotion value. Model construction Figure 1 illustrates the early warning framework designed for detecting failures during multi-level cyclic loading fatigue tests. This framework is structured into three essential stages, comprising damage stage division, damage stage classification and identification, and early warning triggering. The detailed process unfolds as follows. Multi-level cyclic loading early-warning framework. Damage stage division involves applying cyclic loading to coal specimens in the multi-level fatigue test, with an AE acquisition system monitoring their damage and failure processes in real time. Based on the loading process and the evolution of AE signals, the damage process of coal is categorized into five distinct stages, namely the linear damage stage, the first damage stage (micro-damage initiation), the second damage stage (stable crack propagation), the third damage stage (accelerated crack propagation), and the fourth damage stage (fatigue failure). Given that the linear damage stage fundamentally differs from the other stages in terms of mechanism under cyclic loading, it is excluded from subsequent classification and identification analyses. Damage stage classification and identification involve inputting AE parameters into the SSA-Transformer-GRU classification model, which then determines the specific damage stage of the coal. The Isolation Forest algorithm is used to detect outliers in AE data classified as the fourth damage stage. On this basis, the CRITIC-TODIM method is applied to calculate the EW of coal under multi-level cyclic loading. Finally, an early warning is triggered when the EW value reaches 100, and a dual early warning system is constructed in combination with the ERR. Experimental process Coal samples for the experiment were sourced from the Yangcun Coal Mine, part of the Yankuang Group in Shandong Province. These samples were processed into standard cylindrical specimens measuring 50 mm × 100 mm. Initially, triaxial compression tests were performed under a confining pressure of 5 MPa, revealing an average compressive strength of 37.73 MPa for the coal. To replicate the loading and unloading effects experienced during mining, a multistage cyclic loading fatigue test was subsequently conducted at a frequency of 0.5 Hz. Figure 2 illustrates the specific loading method and the experimental apparatus used. Experimental loading equipment. The experiment employed the MTS815.02 electro-hydraulic servo rock mechanics testing system, known for its high precision and stability. The cyclic loading test followed a fixed-lower limit stress and stepwise increase in upper limit stress scheme, conducted entirely in stress-control mode. To ensure safety, the experiment was programmed to automatically stop if the axial displacement of the coal reached or exceeded 3 mm. Initially, an axial force of 0.5 kN was applied to ensure proper contact between the specimen and the indenter. Subsequently, confining pressure was applied at a rate of 0.05 MPa/s until it reached 5 MPa. Once the confining pressure stabilized, axial and circumferential displacements were reset. To minimize initial linear loading damage to the specimen, it was first loaded at a rate of 0.01 kN/s to the upper-limit stress of the first-stage cyclic loading, then switched to a rate of 0.1 kN/s to complete the loading–unloading cycle for that stage. After each stage, loading continued at the same rate to the upper-limit stress of the next stage, repeating the loading–unloading process. During the test, the fatigue state of the coal specimen was judged in real time according to the AE activity. The coal is considered to enter an obvious fatigue state when the ring-down count of AE increases continuously and significantly. When the coal does not exhibit obvious fatigue characteristics, the stress-level increment is 5% of the conventional compressive strength. Once entering the fatigue stage, the stress-level increment is adjusted to 2.5% to avoid sudden non-fatigue failure caused by rapid damage accumulation and to realize refined observation of the fatigue evolution process. The test proceeded step by step until the specimen failed. Given the long time consumption and high cost of fatigue tests, a cosine wave with a frequency of 0.5 Hz was selected as the loading waveform, and 3,000 cycles were preset for each stress level. The cycle number serves as a core parameter controlling damage accumulation. Too few cycles result in inadequate damage evolution and cannot reflect the differentiated effects of various stress gradients, whereas excessive cycles greatly extend the experimental duration and may induce abnormal brittle failure under high stress. Accordingly, 3000 cycles were reasonably determined to ensure sufficient fatigue damage accumulation at each stress stage, balance experimental efficiency, and maintain the stability and validity of test data. Table 1 presents the mechanical test results for three representative coal specimens under multistage cyclic loading at 0.5 Hz. Table 1Results of triaxial cyclic loading tests on representative coal specimens under subjected to various stress levels.Loading frequency (Hz)Confining pressure (MPa)Sample numberingCyclic loading stress levelLower limit stress ratio (%)Upper limit stress ratio (%)Number of cyclesDamage or failure0.55B1130703000No failure230753000No failure330803000No failure43082.51174FailureB2130703000No failure230753000No failure330803000No failure43082.5942FailureB5130703000No failure230753000No failure330803000No failure43082.5706Failure Results of triaxial cyclic loading tests on representative coal specimens under subjected to various stress levels. Determination of AE parameters AE technology, a widely used nondestructive testing method, involves over a dozen common parameters19–21. Using all these parameters to classify and identify coal damage stages would significantly increase time and computational costs. However, five key parameters, ring-down count, rise time, duration, amplitude, and energy, effectively capture the main characteristics of AE waveforms. These parameters can be directly obtained from the instrument without further processing, making them ideal for damage-stage classification. Figure 3 illustrates the physical meanings of these parameters. The ring-down count indicates the number of signal oscillations exceeding the threshold voltage. Rise time measures the interval from when the AE signal first surpasses the threshold to reaching its peak amplitude. Duration is the time span from the arrival of the signal at the sensor to its conclusion. Amplitude represents the peak value of the AE waveform, while energy, depicted as the area under the signal envelope, reflects the energy released by the event22. Definition of AE parameters. Construction of the classification model Data preprocessing After acquiring the time series of fatigue AE parameters from coal under multi-level cyclic loading, the data from three repeated cyclic loading–unloading tests were first normalized. Data normalization, a widely used preprocessing technique, scales each feature value to a specific range, ensuring a zero mean and unit variance. This approach serves several critical purposes, namely eliminating the influence of dimensional differences between features, enhancing the predictive capability of the model, facilitating model convergence during training, accelerating training speed, and improving prediction accuracy. Additionally, it standardizes feature values across samples, enhancing data comparability for subsequent analysis23. Among various normalization methods, this study employed the standard normal method based on standard deviation, calculated as follows. In the formula, denotes the normalized data. Here, x represents the original data, μ refers to the mean, and σ denotes the standard deviation. In classifying damage stages, the original labels are discrete values of 1, 2, 3, and 4, each representing a distinct damage stage. Most machine learning algorithms struggle with discrete labels that have categorical attributes, necessitating an encoding method to convert them into a numerical format suitable for model processing. One-Hot Encoding is employed for this purpose, transforming categorical variables into binary vector representations. This approach not only facilitates model processing but also prevents misunderstandings24. Although the original labels appear to have an ordinal relationship, they merely denote different categories without implying any numerical order. One-Hot Encoding effectively removes these false order relationships by creating separate binary feature columns for each category, thereby preventing the model from developing biases due to misinterpretation of label values and enhancing its ability to understand categorical features, thereby improving the accuracy and generalization performance of the classification task.25 Hyperparameter optimization In machine learning model training, hyperparameter settings significantly influence model performance. Hyperparameters, such as the learning rate and regularization coefficient, are parameters that must be preset before training, as they cannot be automatically learned during the process26–28. These parameters directly affect the final performance and generalization ability of the model, yet determining their optimal values in practice is often challenging. The SSA, a swarm intelligence optimization method inspired by the foraging and anti-predation behaviors of sparrows, addresses this challenge. By emulating key strategies such as individual behaviors, group interactions, position update mechanisms, and fitness evaluations during the foraging process of sparrows, SSA efficiently searches for the global or approximate optimum in the solution space. It boasts strong global search capabilities and high convergence stability, making it suitable for a wide range of complex optimization scenarios29. Consequently, SSA is chosen as the core method for hyperparameter optimization in this study. Transformer-GRU classification model Transformer-GRU is a deep learning classification framework for time-series data. By integrating a Transformer and a GRU, it captures global temporal dependencies and extracts local dynamic features, enabling high-precision identification of coal damage stages. The core of transformer is the self-attention mechanism, which directly captures long-range global dependencies within input sequences and breaks through the sequential processing limitations of traditional recurrent neural networks. Using the multi-head attention mechanism, the model can extract information in parallel across multiple feature subspaces, enhancing its representation of complex temporal patterns. Combined with positional Encoding, the model retains temporal order and relative position information between sequence elements, thereby preserving the integrity of temporal features. As a lightweight variant of recurrent neural networks, GRU achieves selective information retention and forgetting via update and reset gates. It can efficiently model temporal dynamics and mitigate the vanishing gradient and exploding gradient problems, making it suitable for continuous, gradual time-series signals such as acoustic emission. To highlight the rationality and superiority of the proposed method, it is subsequently systematically compared with the single GRU model to verify the effect of introducing the transformer on improving temporal feature modeling and classification performance. Evaluation of model generalization ability K-fold cross-validation is a widely used statistical method for evaluating machine learning model performance30–32. In this approach, the original dataset is repeatedly divided into training and validation sets. Through multiple training and validation rounds, the generalization ability of the model on unknown data is estimated by averaging the evaluation results from each round, as illustrated in Fig. 4. In K-fold cross-validation, each sample is used for both training and validation, thereby enhancing data utilization efficiency. By averaging performance indicators across multiple validation results, the method significantly reduces evaluation fluctuations caused by variations in single data divisions, resulting in a more stable assessment of the generalization ability of the model. Additionally, training and validating the model on various data subsets help identify potential overfitting risks, thereby enhancing the robustness of model construction. Principle of K-fold cross-validation. Interpretability analysis This paper explores the decision-making mechanisms of complex classification models for identifying coal failure stages using the LIME method. LIME offers a local interpretability analysis of the classification behavior of the SSA-Transformer-GRU model. It operates independently of specific model structures, constructing an interpretable surrogate model within a local neighborhood. The process begins by introducing perturbations in the original feature space of the sample to generate neighboring pseudo-samples. The predictions of the original model for these samples are then obtained. Next, weights are assigned to the pseudo-samples based on their similarity to the original sample, and a simple, interpretable linear model is fitted using weighted regression in the local neighborhood of the sample33. This approach transforms the complex nonlinear decision-making of the model into an intuitive explanation, highlighting key AE features and their contribution weights. By doing so, the method not only reveals the primary features the model relies on, verifying their consistency with rock mechanics principles, but also identifies potential atypical sensitive features or abnormal dependencies in the decisions of the model, thereby significantly enhancing the reliability of the classification results. Construction of early warning model Outlier detection Outlier detection identifies abnormal data points in a dataset that significantly differ from other samples. Isolation Forest, an ensemble learning-based method, exemplifies this approach34,35. Its core concept draws from the construction strategy of random forests, determining whether a data point is an outlier by assessing the ease with which it can be isolated. The process begins with a dataset T containing n samples. The algorithm randomly selects a feature and a splitting point to divide the data into two subsets. This partitioning is recursively applied to each subset until a stopping condition is met, such as when a subset contains only one sample or the tree reaches a predetermined height. These steps are repeated to construct multiple isolation trees, collectively forming an isolation forest. For each data point, the algorithm determines its path length in each tree, defined as the number of edges traversed from the root to the corresponding leaf node. It then averages these path lengths across all trees to measure the degree of isolation of the point. A shorter path suggests easier isolation, indicating a higher likelihood of the point being an outlier. Ultimately, the algorithm flags samples with an isolation degree below a predetermined threshold as outliers. Figure 5 illustrates the fundamental principle of the Isolation Forest detection method. Principle of the Isolation Forest detection method. EW early warning trigger The CRITIC-TODIM method integrates the CRITIC weight determination approach with the TODIM interactive multi-criteria decision-making technique to address complex decision-making problems involving multiple indicators and alternatives. Initially, the CRITIC method objectively assigns weights to each evaluation indicator. Subsequently, the TODIM method ranks and selects the alternatives, ultimately yielding a more scientifically grounded decision-making outcome. The CRITIC weight determination method objectively assigns weights to each indicator by evaluating both the contrast intensity and conflict among them. Contrast intensity, typically measured by standard deviation, reflects the variation in values of a single indicator across different evaluation objects. Conflict assesses the correlation between two indicators, a strong positive correlation suggests significant overlap in the information provided, indicating a low level of conflict. The weight calculation process of the CRITIC method emphasizes contrast and conflict to ensure objective and rational weight distribution36,37. The steps of the CRITIC weight determination method are as follows. Establish the original data matrix. In the formula, X denotes the original matrix, which represents the set of importance indices for evaluation indicators. Here, xij refers to the value of the j-th indicator for the i-th object being evaluated. Matrix standardization entails adjusting the data to a common scale. Positive indicators are standardized using Eq. (3), while negative indicators are standardized using Eq. (4). For simplicity, the standardized matrix is still referred to as Matrix X. In this formula, represents the maximum value among all elements in the j-th column of the matrix. Here, denotes the minimum value among all elements in the j-th column. Calculate the variability of indicators. In which denotes the average value of the j-th indicator, while σj denotes the standard deviation of the j-th indicator within the evaluation matrix. Calculate the conflict of indicators. In this formula, Tj denotes the conflict index, while rkj signifies the correlation coefficient between the indicators. Calculate the amount of information. In the formula, the information quantity Cj represents the product of the variability and conflict of the indicator. Calculate the weights of indicators. The TODIM interactive multi-criteria decision-making method, grounded in Prospect Theory, is designed to account for the psychological and behavioral traits of decision-makers, such as loss aversion and risk preference, when confronted with risks and uncertainties. By emulating the cognitive decision-making process, the TODIM method evaluates the relative dominance of each alternative across various criteria and ranks them comprehensively. This process helps decision-makers select options that align with their psychological expectations in complex, dynamic environments. The method is particularly effective in scenarios involving multiple attributes and alternatives, where risk preferences play a significant role. It adeptly integrates the subjective judgments and risk attitudes of decision-makers, thereby enhancing the practical applicability and rationality of decision outcomes. The steps involved in the TODIM interactive multi-criteria decision-making method are outlined below. Calculate the pairwise dominance degrees of calculation schemes. Local dominance degrees of schemes Ai and Ak with respect to indicator w. In the formula, θ represents the loss attenuation coefficient, set at 2.5 in this model. This coefficient indicates the level of loss aversion of the decision-maker. A smaller value of θ signifies greater sensitivity to losses. Calculate the global dominance. Calculate the overall advantage value of the calculation scheme. Results analysis Damage stage division In this experiment, the evolution patterns of AE parameters in various coal samples subjected to fatigue under multistage cyclic loading were generally consistent. Due to space constraints, we illustrate this with sample B2. AE signals were collected using the PCI-2 AE detection and analysis system. Drawing from numerous prior conventional compression AE experiments, we determined the AE parameters for coal under cyclic loading, including a sampling frequency of 10 MHz, a gain of 38 dB, a threshold of 47 dB, a hit definition time of 50 μs, a hit interval time of 300 μs, a sampling time of 0.05 s, and an adjusted threshold voltage of 1.0 V. The specific results are depicted in Fig. 6. Based on the dynamic evolution of AE parameters and changes in stress levels during cyclic loading, the fatigue damage process of coal and rock is categorized into five stages: crack closure, linear elastic, stable crack growth, unstable crack growth, and failure. Evolution of acoustic emission parameters: (a) Ring-down count; (b) Rise time; (c) Duration; (d) Amplitude; (e) Energy. The linear damage stage occurs during linear loading up to the upper limit stress of the first stress level. The slow loading rate results in low, stable AE activity, and all parameters remain at low levels, indicating only minor initial damage to the coal at this stage. The first damage stage (micro-damage initiation stage) takes place at the first stress level (30% ~ 70%). Internal fractures in the coal undergo periodic compaction and closure, with AE activity remaining weak and parameter values staying low. The second damage stage (stable crack propagation stage) corresponds to the second stress level (30% ~ 75%). The elevated upper limit stress facilitates the initiation and stable propagation of new cracks inside the coal, resulting in a marked increase in strain energy release and an initial enhancement of AE signals. The third damage stage (accelerated crack propagation stage) is associated with the third stress level (30% ~ 80%). A further rise in the upper limit stress induces continuous propagation and interconnection of new cracks. AE signals become extremely active, with the ring-down count and energy increasing significantly, and the rise time and duration becoming prolonged. As the number of loading cycles increases, the release of internal strain energy becomes more pronounced, driving the coal into a state of deep fatigue. The fourth damage stage (fatigue failure stage) occurs at the fourth stress level (30% ~ 82.5%). The combined action of linear and cyclic loading leads to continuous damage accumulation. The coal specimen gradually loses its load-bearing capacity, and internal elastic energy is nearly completely released, with all AE parameters reaching their peaks at the point of fatigue failure. Damage stage classification and identification When employing AE signals to identify the fatigue damage stages of coal intelligently, the linear damage stage is typically excluded from the classification system. This decision is primarily based on several considerations. The physical mechanisms exhibit essential differences. The linear damage stage aligns with the initial monotonic loading process, during which the AE response primarily reflects the compaction behavior and elastic deformation of the material upon first-time force application. This stage represents a one-time static mechanical response. Conversely, the subsequent four stages illustrate the progressive accumulation, expansion, and evolution of damage under cyclic loading, characterized by typical fatigue damage traits. These two stages fundamentally differ in their mechanical mechanisms and AE behaviors. The practical value of engineering early warning is significant. In rock mass engineering safety monitoring, capturing the initial loading stage is often challenging or has already occurred, limiting its relevance. Instead, understanding the dynamic evolution of damage under fatigue loading is crucial for assessing structural stability and predicting potential catastrophes. This research aims to enhance the relevance and applicability of the model for engineering early warning by identifying the typical fatigue damage sequence, comprising compaction of primary cracks, stable propagation, accelerated propagation, and unstable failure. Hyperparameter optimization results Following the SSA hyperparameter optimization, the key parameter configuration of the model was established, including a position encoding dimension of 5, 16 attention heads, and a key vector dimension of 32. The Dropout coefficient is set to 0.1, and the learning rate is 0.0001. The GRU hidden layer contains 64 neurons. The overall processing flow of the Transformer-GRU classification model optimized by SSA is as follows. The model takes as input a feature sequence of 5 AE parameters. First, the data is position-embedded through the positional encoding layer, which has 5 output channels and a maximum sequence length of 5. Then, the embedded features and the original input are fused element-wise through an addition module and fed into the first self-attention layer. This layer has 5 output channels and 16 attention heads, with 32 channels for keys, queries, and values, used to extract global dependencies within the sequence. Next, the model applies regularization via dropout_1 with a 10% dropout rate to suppress the activation of some neurons and prevent overfitting. Then, the second self-attention layer is connected to further model sequence features, and regularization is applied via dropout 2 with a 10% dropout rate. After that, the features are fed into a GRU layer with 64 hidden units to capture dynamic temporal patterns in the sequence. The GRU output is mapped to a 4-dimensional feature vector via a fully connected layer, and classification is performed using Softmax, completing the inference process from sequence input to classification results. The overall structure of the model is shown in Fig. 7. SSA-Transformer-GRU classification process. Classification results and model classification performance The three groups of experimental data were uniformly split into training, validation, and test sets. Figure 8 shows the classification accuracy of the two classification models, namely SSA-Transformer-GRU and SSA-GRU, across the training, validation, and test sets for the four damage stages. Overall, both models exhibit favorable identification performance across all damage stages. SSA-Transformer-GRU and SSA-GRU Classification Results: (a) Training set, (b) Validation set, (c) Test set. The minimum identification accuracy of the SSA-GRU model for each damage stage is 90.1%, 87.9%, 93.7%, and 82.9%, successively. Specifically, the identification accuracy of this model for the four stages is 90.5%, 88.8%, 93.7%, and 84.4% on the training set, 91.3%, 87.9%, 94.2%, and 84.2% on the validation set, and 90.1%, 88.3%, 94.1%, and 82.9% on the test set, respectively. The minimum identification accuracy of the SSA-Transformer-GRU model for each damage stage is 88.2%, 94.6%, 90.7%, and 90.7%, respectively. Among them, the identification accuracy of this model for the four stages is 89.7%, 94.9%, 91.4%, and 93.2% on the training set, 88.2%, 94.9%, 92.1%, and 90.7% on the validation set, and 89.9%, 94.6%, 90.7%, and 94.6% on the test set, respectively. Compared with SSA-GRU, the classification accuracy changes of SSA-Transformer-GRU in each damage stage are as follows: − 0.8%, + 6.1%, − 2.3%, and + 8.8% on the training set, − 3.1%, + 7.0%, − 2.1%, and + 6.5% on the validation set, and − 0.2%, + 6.3%, − 3.4%, and + 11.7% on the test set. It can be observed that the recognition accuracy of SSA-Transformer-GRU is not higher than SSA-GRU in some early damage stages. This phenomenon is reasonable and can be well explained: the Transformer structure is good at capturing long-range global temporal dependencies in AE sequences, which is particularly effective for the fourth damage stage with obvious abrupt characteristics and strong time correlation. For the earlier damage stages, the AE signals are relatively stable with weak long-term correlation, and the original GRU model already has sufficient capability to model local temporal features. Therefore, the introduction of Transformer may lead to slight overfitting or redundant feature learning on stable local signals, resulting in a small decrease in individual accuracy. Both models have a small number of misjudgments that classify damage signals into adjacent stages, but all misjudgment errors are controlled within 18.0%. Such misjudgments are mainly due to the continuous evolution of the coal-damage process. Each damage stage lacks a strict boundary in actual evolution and exhibits clear transitional characteristics. Considering that the accurate identification of the fourth damage stage plays a decisive role in the construction of the subsequent early warning model, and the accuracy of SSA-Transformer-GRU in this stage (especially on the test set) is significantly improved by 11.7% compared with SSA-GRU, the slight accuracy degradation in individual non-critical stages is acceptable. Therefore, SSA-Transformer-GRU is selected as the final classification model for damage stages. To assess the performance of the model, we selected accuracy, precision, recall, F1 score, and specificity as evaluation metrics. Accuracy indicates the proportion of correctly classified samples. Precision measures the proportion of true positive samples among those predicted as positive. Recall assesses the proportion of actual positive samples that are correctly identified. The F1 score, a harmonic mean of precision and recall, provides a comprehensive performance evaluation. Specificity denotes the proportion of true negative samples accurately identified as negative38–40. Table 2 shows that all evaluation indicators exceed 0.79, demonstrating the exceptional recognition performance of the SSA-Transformer-GRU classification model and its ability to achieve high-precision damage-stage classification. This outcome indicates that the hyperparameter combination optimized by the SSA algorithm significantly enhances the capabilities of the model, including feature extraction discriminability, adaptability to complex patterns, and regularization effect. As a result, the classification accuracy, stability, and generalization performance of the Transformer-GRU model are comprehensively improved, showcasing an effective synergy between the optimization strategy and the model architecture. Table 2Recognition effect of the SSA-transformer-GRU model.Sample partitioning Evaluation metricsTraining setValidation setTest setAccuracy0.923190.922310.92038Precision0.922910.914990.92449Recall0.800400.799320.79628F1 score0.820020.816890.81571Specificity0.971760.971700.97058 Recognition effect of the SSA-transformer-GRU model. K-fold cross-validation and LIME interpretability analysis Figure 9 illustrates the high stability and exceptional generalization ability of the model during five-fold cross-validation. The accuracy for each fold consistently hovers around 0.915, with minimal fluctuation between 0.9083 and 0.9245. The average accuracy is 0.91708, and the very low standard deviation of 0.0061769 indicates the robustness and insensitivity of the model to data partitioning methods. These findings demonstrate that the model not only excels in overall classification performance but also consistently delivers high-precision results. It is less influenced by data variations, suggesting its potential for reliable application in real-world engineering scenarios. K-fold cross-validation. Figure 10 presents the results of the LIME feature importance analysis, revealing distinct variations in how each input feature affects the classification decisions of the model. The ring-down count has the highest importance value at 5.3949, indicating its predominant role in classifying coal damage stages and its substantial contribution to interpreting the model output. Physically, the ring-down count directly reflects the number and frequency of internal micro-crack initiation and propagation in coal, which is the most essential indicator for identifying the evolution of fatigue damage. The energy feature has an importance value of 1.8109, indicating its critical role in the classification process. The energy parameter characterizes the strain energy released during crack propagation and coalescence, which is closely related to the scale and intensity of coal fracture. Duration ranks third, with an importance value of 1.2535, exerting some influence on the model decision, though less than the first two features. Duration reflects the development degree and complexity of internal fractures, but it is easily affected by the propagation path of AE signals. In contrast, rise time and amplitude, with importance values of 0.1856 and 0.0771, respectively, exhibit limited influence on the model decision in this analysis, indicating that their roles in interpreting the model output are not significant. These two parameters are more susceptible to sensor layout, wave attenuation and local stress conditions, so their sensitivity to coal damage evolution is relatively weak. The above feature importance sequence is highly consistent with the physical mechanism of coal fatigue failure under cyclic loading, which verifies the rationality and reliability of the model decision-making process. LIME feature importance analysis. Evolution law of the average event intensity of AE The LIME feature importance analysis results revealed that ring-down count and energy are crucial for identifying damage stages in coal. To better evaluate the damage degree, the average event intensity, defined as the Energy Ring-down Count Ratio (ERR), is introduced in this study. The ERR is calculated as the ratio of the total AE event energy to the total ring-down count within a specified time interval, which can be expressed as: In the formula, Ei represents the energy of the i-th AE event in the interval, and Ni represents the ring-down count of the i-th AE event in the interval. This index physically represents the average energy release per AE oscillation within the interval. Compared with using ring-down count or energy individually, ERR normalizes the event intensity via this ratio, which reduces the interference caused by overlong signal duration (that may lead to inflated ring-down counts) and non-essential oscillations such as environmental noise. Therefore, ERR can effectively characterize the internal fracture behavior of coal and quantitatively evaluate the fracture severity. To verify the effectiveness of the ERR parameter, three specimens from conventional triaxial compression tests and several specimens from multistage cyclic loading fatigue tests were selected for comparative analysis. When calculating the ERR using the sliding window method, selecting the appropriate window length and moving step is crucial for accurate analysis. A window length that is too large may obscure detailed features of local crack development, while one that is too small can be prone to random noise, failing to capture the overall trend of crack evolution within the interval. An inappropriate moving step can result in data redundancy or the omission of key information. To resolve this issue, optimal parameters for two test types were determined by verifying multiple parameter combinations. For the triaxial compression test, a window length of 3,000 and a moving step of 300 were found to be optimal. For the multistage cyclic loading fatigue test, the optimal window length is 5,000 with a moving step of 500. The ERR calculation results for these tests are presented as broken-line graphs in Fig. 11 for the triaxial compression test and Fig. 12 for the multistage cyclic loading fatigue test. ERR value of triaxial compression test: (a) CB1; (b) CB2; (c) CB3. ERR value of multistage cyclic loading fatigue test: (a) B1; (b) B2; (c) B5. In the triaxial compression test, the ERR curve presents a clear three-stage evolution pattern, which is highly consistent with the damage evolution process of coal specimens. Before the load continuously increases to the stress peak, the ERR rises steadily, indicating that the average energy release of internal micro-fracture events in coal is continuously enhanced, and the damage accumulates stably and intensifies gradually. When approaching the stress peak, the ERR reaches the maximum value, marking that cracks enter the unstable propagation stage and the coal is about to undergo macroscopic fracture. In the post-peak stage, with the rapid formation of the macroscopic main fracture surface, the bearing capacity of the specimen drops sharply, and a large number of crushing signals with low energy and high ring-down count appear, so the ERR decreases sharply. Subsequently, the ERR rises again, indicating a significant transition in the post-peak mechanical behavior. After the formation of the main fracture surface, the deformation mechanism of coal shifts from the initiation and propagation of new cracks to frictional sliding along the existing fracture surfaces. Under confining pressure, the fragmented coal mass recombines to form a new load-bearing structure, generating AE signals characterized by concentrated energy release and relatively few oscillations. This process not only reflects the residual strength of coal but also demonstrates that the ERR index can effectively filter out redundant oscillatory interference and accurately capture sensitive changes in fracture strength. The rise–sharp drop-recovery V-shaped feature of ERR can be regarded as an important precursor for coal failure under conventional triaxial compression. Different from conventional triaxial compression, the ERR in the multistage cyclic loading fatigue test shows remarkable stepwise evolution characteristics. During the initial linear loading to the upper limit of the first stress level, the variation trend of ERR is similar to that in the triaxial test, rising gradually with the linear increase of load, indicating linear damage accumulation inside the coal specimen, and micro-cracks begin to propagate under continuous loading, which is consistent with the linear damage mechanism in triaxial compression. After entering the cyclic loading and unloading stage at the same stress level, the ERR remains relatively stable. The microscopic mechanism of this stability is that under cyclic loading with a constant upper stress limit, the activatable micro-cracks inside the coal have been fully initiated and propagated, and the crack system reaches local damage saturation, only performing elastic reciprocating motion of closure-opening without generating large-scale new cracks. Therefore, the average energy release per unit ring-down count tends to be stable, reflecting the fatigue adaptation effect of coal to the current stress level. When the upper stress limit is increased, the original adaptation state is broken, a large number of new cracks initiate and existing cracks further propagate inside the coal, and the energy release intensity increases abruptly, leading to a stepwise rise of ERR. After several cycles, the coal adapts to the new upper stress limit again, crack propagation reaches local saturation once more, and ERR stabilizes at a higher level accordingly. In the macroscopic failure stage, different from the sharp drop and rebound in the post-peak stage of triaxial compression, coal specimens under multistage cyclic loading experience multistage damage accumulation, cracks gradually penetrate, and sudden overall instability occurs under the critical load. Thus, ERR climbs sharply to an extreme value before macroscopic instability. This catastrophic sudden increase feature can be used as a key precursor for coal damage and instability under multistage cyclic fatigue loading. The above differences confirm the universality and pertinence of the ERR index. ERR can effectively reflect the damage evolution laws under different loading modes, and distinguish two different mechanisms, namely one-time peak failure under monotonic continuous loading and stepwise fatigue cumulative failure under multistage cyclic loading, through its evolutionary morphology differences. From a unified theoretical perspective, the evolution of ERR under the two loading modes is essentially controlled by the initiation, propagation, and penetration of internal cracks as well as the transition of fracture modes in coal. Monotonic loading drives the continuous development of cracks until the main fracture forms at one time, whereas cyclic loading causes cracks to undergo a stepwise process of “activation–propagation–saturation” at each stress level before they eventually coalesce and lead to instability. The morphological difference of ERR is essentially an external manifestation of the changes in crack propagation rate, scale and fracture mode caused by different loading paths, which can provide a unified and accurate basis for coal damage early warning in complex engineering scenarios. Early warning triggering Outlier identification Figure 13 illustrates the distribution of outliers in the fourth damage stage of validation set B2. During this stage, several key AE parameters exhibit abnormal fluctuations. Notably, the ring-down count increases significantly in the later stage, with a sharp rise indicating intensified internal fracture activities within the coal mass. After approximately 27,800 s, there is a marked increase in outliers related to rise time, suggesting a slowdown in signal propagation speed due to structural deterioration and crack expansion. Similarly, multiple high-duration outliers emerge after 27,800 s, indicating prolonged AE events and a more complex internal fracture process. While most amplitude values remain within the normal range of 50 ~ 70 dB, several isolated outliers deviate significantly, pointing to local high-energy fracture events. Additionally, multiple high-energy outliers appear as the timeline approaches 27 800 s, indicating substantial energy release from the coal mass, which corresponds to intense fracture behavior. Analysis indicates that between 27,800 and 28,000 s, the coal mass experiences a series of abnormal processes, including local structural instability, accelerated crack expansion, and even small-scale fractures. Outlier identification: (a) Ring-down count; (b) Rise time; (c) Duration; (d) Amplitude; (e) Energy Determination of indicator weights As illustrated in Fig. 13, during the fourth damage stage of coal, the specimen ultimately failed after undergoing multiple cycles of loading and unloading, with all indicators reaching high levels. Consequently, the five AE multistage cyclic loading fatigue indicators are considered positive indicators. Figure 14 presents the weights of each AE indicator as determined by the CRITIC weighting method. The energy indicator, with the highest weight of 0.429, plays a central role in the evaluation system, significantly influencing the overall evaluation result. The ring-down count, with a weight of 0.322, is also a key indicator, contributing substantially to the comprehensive evaluation. The amplitude, weighted at 0.151, still holds considerable influence within the system. In contrast, the duration, with a weight of 0.051, has a relatively minor impact on the final evaluation. The rise time, having the lowest weight of 0.048, indicates its minimal role in the current evaluation system. Weight index of AE parameters. EW early warning time To evaluate outliers for five AE parameters in validation set B2, the CRITIC method was used to assign weights, followed by the TODIM method for assessment. The findings are illustrated in Fig. 15, which depicts the trend of EW scores during multistage cyclic loading. The analysis indicated that at 27,879.6 s, the EW reached the maximum score of 100. The system identified this point as the early-warning time. When compared to the actual rockburst occurrence at 27,981.1 s, the early-warning time was 101.5 s earlier. This demonstrates the strong predictive capability of the system. EW score. The overall trend of the EW coefficient reveals distinct stage-by-stage variations in score distribution. In the early stage, scores are relatively high with densely distributed data points. In the intermediate stage, scores decrease, and data points are relatively sparse. In the late stage, scores rise again to a high level and maintain a dense distribution. This pattern suggests that as stress on coal increases under load, internal damage intensifies. Despite exhibiting fatigue characteristics at this stage, the coal retains some resistance to external forces. However, once cyclic loading reaches a critical point and internal damage surpasses a threshold, the coal ultimately fails. The damage evolution of coal, as indicated by the EW coefficient, aligns closely with the characteristics of the ERR value in the fourth stage. This alignment further confirms the reliability of the ERR value in representing the damage state of coal. Combined early warning based on ERR and EW The EW scores were calculated for specimens B1, B2, and B5, respectively, and a dual early warning system was constructed by combining EW with ERR. As shown in Table 3, the fatigue instability moments of the three specimens were 28,293.6 s, 27,981.1 s, and 27,661.6 s, respectively. Table 3Early warning time and coal instability failure time of each specimen.Warning mode specimenERR warning time/sEW warning time/sCoal fatigue instability time/sB128,173.528,212.928,293.6B227,795.427,879.627,981.1B527,601.127,651.127,661.6 Early warning time and coal instability failure time of each specimen. According to the sudden extreme points of ERR before fatigue instability, the ERR warning moments were determined as 28 173.5 s, 27 795.4 s, and 27 601.1 s, which were 120.1 s, 185.7 s, and 60.5 s earlier than the instability moments, respectively. According to the moments when the comprehensive EW warning coefficient reached the set threshold (100), the EW warning moments were determined as 28 212.9 s, 27 879.6 s, and 27 651.1 s, which were 80.7 s, 101.5 s, and 10.5 s earlier than the instability moments, respectively. The comparative analysis shows that the two indicators exhibit clear complementary characteristics in the sequence of early warning times. The ERR warning time is generally earlier than the EW warning time, mainly because ERR directly characterizes the sudden change in the internal energy release of coal and can capture the critical transition in the damage state earlier. EW is based on a comprehensive evaluation of cumulative anomalies of multiple parameters, its warning response is relatively conservative, but it effectively reduces the risk of false alarms caused by instantaneous fluctuations of a single parameter. Therefore, the two indicators can form an early-warning chain: “preliminary prompt—confirmed alarm”. ERR first issues a sensitive warning to indicate that coal has entered a high-risk evolution stage, and EW then issues a confirmed alarm based on sufficient cumulative anomalies and collaborative evidence of multiple parameters, providing a more reliable basis for engineering decision-making. In addition, the early warning lead times of the three specimens differ to some extent, reflecting the inherent heterogeneity of coal materials and the randomness of the failure process. For example, the ERR and EW lead times of specimen B5 were 60.5 s and 10.5 s, respectively, indicating a relatively sudden failure process with a short duration of precursor signals. In contrast, the lead times for specimen B2 were 185.7 s and 101.5 s, respectively, suggesting a more complete damage evolution and more significant precursor signals. This difference precisely indicates that a single indicator is difficult to apply to all working conditions. The dual early warning mechanism can better adapt to different failure characteristics and improve the universality of the early warning method through mutual verification of the two indicators. In summary, the proposed dual early warning mechanism, ERR and EW, organically integrates a single-parameter precursor indicator with clear physical meaning and a multi-parameter comprehensive evaluation method. It ensures the timeliness of early warning while enhancing the reliability and robustness. Verification results from multiple specimens show that this mechanism has strong adaptability across different coal specimens, providing a solution with greater engineering application value for disaster early warning in coal engineering. Conclusions This study examines the fatigue failure process of coal subjected to multistage cyclic loading and develops a damage identification and early-warning framework using AE parameters. The primary conclusions are as follows. Under both multistage cyclic loading and conventional triaxial compression conditions, the characteristic AE parameters of coal (ring-down count, rise time, duration, amplitude, and energy) exhibit obvious stagewise evolutionary laws highly corresponding to the damage evolution process, indicating that AE technology can effectively characterize the whole process of coal from micro-crack initiation and stable propagation to macroscopic instability and failure. Under both multistage cyclic loading and conventional triaxial compression conditions, the characteristic AE parameters of coal (ring-down count, rise time, duration, amplitude, and energy) exhibit obvious stagewise evolutionary laws highly corresponding to the damage evolution process, indicating that AE technology can effectively characterize the whole process of coal from micro-crack initiation and stable propagation to macroscopic instability and failure. The ERR shows a V-shaped pattern of “first increase -sharp drop-recovery” under conventional triaxial compression. In contrast, it exhibits stagewise stability and gradual elevation under multistage cyclic loading, with a prominent sudden extreme point before instability. ERR can sensitively capture the sudden change in internal energy release with clear physical meaning and can serve as an important precursor indicator of coal damage and instability. The ERR shows a V-shaped pattern of “first increase -sharp drop-recovery” under conventional triaxial compression. In contrast, it exhibits stagewise stability and gradual elevation under multistage cyclic loading, with a prominent sudden extreme point before instability. ERR can sensitively capture the sudden change in internal energy release with clear physical meaning and can serve as an important precursor indicator of coal damage and instability. The SSA-Transformer-GRU classification model constructed in this study achieves high-precision identification of four damage stages of coal. Compared with the SSA-GRU model, the recognition accuracy for the critical fourth damage stage improves by 11.7% on the test set, demonstrating stronger temporal feature extraction and better generalization. The SSA-Transformer-GRU classification model constructed in this study achieves high-precision identification of four damage stages of coal. Compared with the SSA-GRU model, the recognition accuracy for the critical fourth damage stage improves by 11.7% on the test set, demonstrating stronger temporal feature extraction and better generalization. The EW constructed by combining Isolation Forest outlier detection with the CRITIC-TODIM method forms a dual early warning mechanism, “preliminary sensitive prompt-comprehensive confirmed alarm,” together with the ERR indicator. This combination not only ensures sufficient early warning lead time but also reduces the risk of false alarms caused by single-parameter fluctuations, significantly improving the reliability and practicality of the early warning system. The EW constructed by combining Isolation Forest outlier detection with the CRITIC-TODIM method forms a dual early warning mechanism, “preliminary sensitive prompt-comprehensive confirmed alarm,” together with the ERR indicator. This combination not only ensures sufficient early warning lead time but also reduces the risk of false alarms caused by single-parameter fluctuations, significantly improving the reliability and practicality of the early warning system. The proposed AE-based intelligent damage stage identification and ERR-EW dual early warning framework systematically reveals the damage evolution laws of coal under different loading modes, providing a data-driven, mechanism-clear, and practically applicable new method for dynamic disaster monitoring and early warning in deep coal mining engineering. The proposed AE-based intelligent damage stage identification and ERR-EW dual early warning framework systematically reveals the damage evolution laws of coal under different loading modes, providing a data-driven, mechanism-clear, and practically applicable new method for dynamic disaster monitoring and early warning in deep coal mining engineering. Author contributions Luyi Xing: Writing—review & editing, Funding acquisition, Methodology, Supervision. Chao Wang: Writing—original draft, Formal analysis, Data curation. Min Liang: Writing—review & editing, Conceptualization. Baoliang Zhang: Formal analysis. Funding This study was supported by the Shandong Provincial Natural Science Foundation (Grant No. ZR2025QC575), and the Doctoral Research Fund of Shandong Jianzhu University (Grant No. XNBS20116). Data availability The datasets generated and/or analysed during the current study are not publicly available due to the fact that the raw experimental data are core research assets of the research group and will be used for subsequent series of studies. The datasets are available from the corresponding author on reasonable request. Declarations Competing Interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Chao Wang, Email: 2024010111@stu.sdjzu.edu.cn. Min Liang, Email: liangmin@shandong.cn. References Associated Data Data Availability Statement The datasets generated and/or analysed during the current study are not publicly available due to the fact that the raw experimental data are core research assets of the research group and will be used for subsequent series of studies. The datasets are available from the corresponding author on reasonable request.