Approach for Microseismic Monitoring Data-Driven Rockburst Short-Term Prediction Using Deep Feature Extraction and Interpretable Coupling Neural Networks

Abstract

Rockburst disasters have become increasingly prevalent as distinct forms of subsurface geotechnical engineering advanced to the deep earth. Confronted with such a threatening subsurface geopressure disaster that poses a risk to personnel and equipment safety, the microseismic monitoring technology has been employed to track signals generated from rock fracture and collapse in the field. To guide the prevention and control of the hazard, the investigation conducted an effective microseismic data mining method. Through deep feature engineering and interpretable intelligence, a practical and available short-term prediction approach for the rockburst intensity class was developed. On the basis of rockburst case database collected from various underground geotechnical engineering, the neural network-based feature extraction method was conducted in the process of model training. The optimized model was obtained by combining the K-fold cross-validation approach with the structural parameter search methodology. The evaluation among the considered artificial intelligence models on the testing dataset was conducted and compared. Through analyses, the interpretable coupling intelligent model combining convolutional and recurrent neural networks for rockburst prediction were demonstrated with the most robust performance by evaluation metrics. Among them, the proposed adaptive feature extraction method leads the benchmark method by 6% for both accuracy and precision; meanwhile, the proposed metric generalization loss rate (GLR) for accuracy and precision in the validation–testing process reached 1.5% and 0.2%. Furthermore, the Shapley additive explanations (SHAP) approach was employed to verify the model interpretability by deciphering the model prediction from the perspective of the fined impact of input features. Therefore, the investigation demonstrates that the proposed method can predict rockburst intensity with robust generalization and feature extraction capabilities, which possess substantial engineering significance and academic worth.

1. Introduction

As a typical ground pressure hazard, rockburst is brought on by excavation or other external disturbances that cause the underlying rock’s elastic strain energy to suddenly release [1,2]. A rockburst typically causes the surrounding rock to exhibit a variety of rock mechanical phenomena [3,4]. The security of construction personnel and facilities in the field is seriously threatened by the sudden, swift, and destructive nature of rockburst, which can occasionally result in destruction that could completely damage the site and inflict enormous losses [5]. The intensity and likelihood of rockburst can significantly grow if the geotechnical engineering of subsurface space carries on, getting progressively deeper [6,7]. Therefore, research into precise and trustworthy rockburst prediction techniques is crucial to guiding the secure conduction of subsurface geotechnical engineering.

At present, there exist two types of investigation on rockburst prediction, including short-term and long-term prediction [8,9,10,11,12,13]. In order to detect potential rockburst threats, the short-term rockburst prediction employs site monitoring data acquired in real time [14,15,16]. This type of prediction investigation possesses various forecasting targets, such as the location, timing, and intensity level, in contrast to long-term rockburst prediction [17,18], which relies its estimation of the likelihood of rockburst on intrinsic rock mechanical properties. Long-term rockburst prediction primarily targets the design phase of engineering projects (such as before the extraction of working faces or extraction tunnels), assessing the likelihood of ground pressure impacts over the entire engineering operation area. Long-term rockburst prediction encompasses the entire scope of the project and the entire duration of the construction period, representing a macro-level evaluation of the occurrence of ground pressure impacts. Its prediction window typically spans several months to years, significantly longer than the short-term prediction, which generally covers a period of hours to days. Common evaluation indicators for long-term rockburst prediction include the impact energy index, elastic strain energy index, deformation brittleness index, and so on, which belong to the intrinsic characteristics of rocks. In the field, electromagnetic radiation, acoustic emission, and microseismics are often employed methods [19,20,21,22,23,24]. In the area of mining engineering, acoustic emission monitoring technology primarily captures high-frequency, low-energy signals, while microseismic monitoring technology mainly collects low-frequency, high-energy signals. These two methods share similar physical principles, as they, respectively, monitor the micro-fracture within rock mass and the failure and vibration of the rock mass. Consequently, they provide localized and global monitoring and early warning for the ground pressure phenomena in mines. It is significant that the most popular and successful of these methods is microseismic monitoring [1,25,26]. Microseismic monitoring records microseisms generated by the microcracks in rocks, and the monitoring system can analyze the location and intensity of the current damage and vibration. The historical temporal variation in these parameters obtained through microseismic monitoring can be used to predict and issue early warnings for these ground pressure activities.

By analyzing the precursors from microseismic data prior to and during a rockburst, numerous scholars have already presented a number of microseismic characteristics, such as the amount of microseismic events, microseismic energy, moment magnitude, apparent volume, fractal dimension, and b value, and proposed drastic changes or a quiet period of the parameters to predict the danger of a rockburst [27,28,29,30,31,32]. The parameters mentioned, such as moment magnitude, are common phenomena extracted from the microseismic monitoring system of mine ground pressure, which are calculated based on the energy released during the fracturing of rock masses under mining activities and obtained from microseismic monitoring systems and devices. They are commonly used in the study of mine ground pressure disasters and the actual monitoring of mine seismic intensity. In microseismic monitoring, the apparent volume refers to the ratio of the energy released by microseismic events to the volume of the rock mass, which is used to quantify the scale of rock mass fracturing and the intensity of energy release. It reflects the scale of rock mass fracturing; a higher value indicates a wider fracturing range and stronger energy release. The b value is the proportional coefficient in the relationship between the seismic magnitude and frequency, representing the proportional relationship between the frequency of different magnitude earthquakes in a certain area. For example, Ma et al. adopted microseismic energy, microseismic moment, and apparent stress, and investigated the law of the evolution process of ground pressure disasters based on the parameter space and microseismic pathways, which is named by the EMS (Energy–Moment–Stress) method [33]. By examining microseismic events that occur both before and after rockbursts, Xue et al. presented an experience criterion for evaluating rockburst severity and likelihood [34]. They found that the likelihood of rockburst occurrence was related to the abnormal variation in the event amount or b value. The frequent occurrence of high-energy microseismic events and the sharp increase in cumulative apparent volume are important indicators of the onset of violent rockburst [35]. The aforementioned prediction methods of rockburst based on precursory variation in microseismic monitoring parameters prior to the occurrence of the disaster ultimately still pertain to empirical forecasting. They rely on manual analysis to determine the rockburst production pattern, which has a significant subjective nature, inefficiency, and weak precision. An urgent breakthrough in novel technologies is imperative.

The emerging machine learning (ML) is a data-driven and strong-representing artificial intelligence approach [36,37,38], which possesses powerful feature extraction and complicated relationship representation capabilities. The increasing number of rockburst cases gathered from microseismic monitoring has led to a significant body of investigation on rockburst prediction algorithms using artificial intelligence techniques [3,39]. For instance, Kadkhodaei and Ghasemi developed two models based on the logistic model tree algorithm to quantitatively determine the probability and consequences of rockburst occurrence [40]. Numerous machine learning models, such as the random forest (RF) and Support Vector Machine (SVM), have demonstrated encouraging outcomes in the prediction of rockbursts. For instance, Zhou et al. found that gradient boosting and random forests work well in this situation [41]. In order to increase the prediction accuracy of rockburst, several researchers have also used optimization methods or heuristics to improve the developed models. Two new hybrid models based on a random forest (RF) were proposed for short-term rockburst intensity prediction [42]. The hyperparameters were adjusted using the coati optimization algorithm (COA) and whale optimization algorithm (WOA). These studies have helped to advance machine learning for predicting rockbursts from the prediction performance.

Several studies have established methods for assessing rockburst prediction models, such as the uncertainties by confidence interval uncertainty analysis [43,44,45,46,47] and interpretation methods. Sun et al. employed an improved cloud model algorithm integrating a microseismic monitoring feature and proposed an interpretable RF-CRITIC method to predict rockburst disasters from a short-term perspective [31]. Shen et al. used the RF as the rockburst hazard prediction model, optimizing model structure parameters using the Optuna architecture and the Shapley value approach combined to describe the prediction procedure and evaluate the impact of model features [48]. Additionally, a few number of studies have built short-term rockburst prediction models utilizing the feature engineering approach thus far [49,50]. Nevertheless, all of these above traditional intelligence approaches were all predicated based on machine learning algorithms and feature correlation, lacking an adaptive feature engineering strategy to choose the optimal input feature for models and ignoring the direct relationship between model prediction and input parameters. Therefore, investigating novel methods for forecasting short-term rockburst is crucial.

In order to forecast short-term rockburst hazards more precisely, a practical and available prediction method based on the deep feature extraction method and interpretable artificial intelligent algorithms was investigated in this study utilizing microseismic monitoring parameters as the model input. On the basis of rockburst cases gathered from various underground geotechnical engineering databases, the feature extraction based on neural networks in a direct model prediction manner was conducted in the process of model training. The optimized model was obtained by combining the K-fold cross-validation approach with the structural parameter search methodology. The evaluation among the considered artificial intelligence models on the testing dataset was conducted and compared. Through analyses, the interpretable coupling deep neural networks for rockburst prediction were demonstrated with the most robust performance by evaluation metrics. Among them, the proposed metric generalization loss rate (GLR) for accuracy in the validation–testing process reached 1.5%. Furthermore, the Shapley additive explanations (SHAP) approach was employed to verify the model interpretability by deciphering the model prediction from the perspective of the fined impact of input features. The flow chart of this investigation is demonstrated in Figure 1.

2. Data

2.1. Rockburst Dataset Construction

Numerous underground geotechnical projects at home and abroad have been subjected to rockburst disasters, and a total of 119 sets of examples of rockburst engineering in the mining process were collected here as a sample database for the short-term prediction of rockburst intensity classes. They came from the characteristics of rockburst microseismic data variations collected from the underground engineering sites at home and abroad, respectively [51,52]. Each set of samples contained six microseismic parameters, including the accumulation number of microseismic events (N), accumulation energy of microseismics (E), accumulation of apparent volume (V), accumulation number rate of microseismic events (R_N), accumulation energy rate of microseismics (R_E), and accumulation rate of apparent volume (R_V). Among them, N reflected the number and density of microscopic cracks generated from the rock fracture under the condition of high stress, E and V reflected the strength and size of microscopic fractures, and R_N, R_E, and R_V reflected the degree of change in microscopic fractures with the time effect. These six characteristics could accurately forecast the severity of a rockburst in real time and characterize the stress condition and failure of the rock body, which had been widely used in the study of short-term rockburst prediction. Through checking the literature related to rockburst intensity classification standards, according to the field damage characteristics, the rockburst intensity could be divided into none (I), slight (II), moderate (III), and violent (IV) [53].

Table 1. Microseismic parameter dataset for rockburst intensity classes.

No.	N	lg(E/J)	lg (V/m3)	RN/d−1	lg(RE /(J·d−1))	lg(RV/(m3·d−1))	Rockburst Intensity
1	41	5.968	4.694	3.727	4.926	3.653	IV
2	14	5.841	4.622	1.556	4.887	3.668	III
3	17	4.754	4.397	1.889	3.8	3.443	III
…
117	5	3.154	3.309	2.5	2.853	3.008	I
118	18	5.602	4.779	1.8	4.602	3.779	III
119	2	1.94	3.25	1	1.639	2.949	I

Note: d represents day.

demonstrates the microseismic data of the rockburst cases. Among them, 96 sets of data are randomly selected as the training set for training, and the remaining 23 sets of data are used as independent test datasets, which are not involved in the training and validation process at all.

As for the rockburst microseismic datasets in this work, there exist 31, 30, 28, and 23 cases corresponding to none, slight, moderate, and violent rockburst intensity classes, respectively. The figures of statistical boxes from the six input characteristics for distinct rockburst prediction classes have been demonstrated in Figure 2. The measure’s distribution was asymmetrical since, as can be observed, the median of every metric was not situated in the middle of the box. As is illustrated, there exists no fixed relationship between the intensity class distribution and characteristics. Additionally, almost all the characteristics had a relatively discrete value distribution for at least two rockburst intensity classes. This implies that rockburst production is extremely complicated and that sufficient information extraction is required to facilitate rockburst prediction.

2.2. Data Preprocessing

In order to render distinct microseismic features into the prediction model’s input form, the microseismic monitoring dataset conducted data preprocessing for the intelligence method, incorporating data normalization. Figure 3 shows the dataset preparation for preprocessing. The data normalization prepared for inputting is then elaborated as follows:

$$da{t}_i^{*}={\displaystyle \frac{da{t}_i-\mu }{S}},$$

(1)

where $da{t}_i^{*}$ refers to the normalized values. $da{t}_i$ refers to the original values of the time series. S refers to the standard deviation calculated from $da{t}_i$ in all subscripts of i. $\mu$ refers to the mean value of $da{t}_i$ in all i.

Meanwhile, in this procedure of acquiring model datasets suitable for intelligent prediction architecture, the proportion of the amount of rockburst cases in the training and testing datasets was about 4:1. The rockburst sample cases were separated into a training dataset and testing dataset randomly. The testing dataset was divided out thoroughly and reserved to retain independence for the eventual examination.

2.3. Data Balancing

Various methods have been employed to improve model performance, especially for the minority prediction class, such as the synthetic minority oversampling (SMOTE) [54], in order to accommodate unbalanced rockburst intensity data. To balance the dataset and lessen overfitting brought on by simple oversampling in this work, the SMOTE method was employed for creating generated sample cases. Moreover, the SMOTE method created additional samples by interpolating between a minority class sample and its closest neighbors. By using the following equation and balancing the classes in the dataset, c is produced by combining any existing value a with any neighbor b:

c = a + rand(0,1) × |a−b|,

(2)

where c represents the newly acquired sample, a represents the majority sample, b represents the sample selected randomly from the nearest neighbor of sample a, |a−b| represents the absolute variation between sample a and sample b, and rand(0,1) is a random number between 0 and 1.

A total of 96 sample cases were separated randomly as the training dataset, including 31 sample cases marked as none, 25 as slight, 26 as moderate, and 14 as violent. For balancing the class number in the dataset, the SMOTE method produced additional sample cases, generating 124 cases evenly distributed into the four rockburst classes.

3. Methodology

3.1. Structure

In this work, the artificial intelligence architecture included 3 solely utilized algorithms, 4 ensemble learning methods, and coupling deep neural networks. The solely utilized algorithms included Decision Tree (DT) [55], Support Vector Machine (SVM) [56] and Artificial Neural Networks (ANN). The ensemble learning methods included random forest (RF) [57], Light Gradient Boosting Machine (LightGBM) [58], Adaptive Boosting (AdaBoost) [59], Extreme Gradient Boosting (XGBoost) [60], and Gradient Boosting Decision Tree (GBDT) [61]. The coupling deep network structure (CDN) was conducted combining convolutional neural networks (CNN) and time series predictor gated recurrent units (GRUs).

These distinctive predictors rendered diverse rockburst prediction performance in this work. It was simple for DT to comprehend and interpret with high-efficient computation. SVM tended to work effectively with high-dimensional data and possesses the capacity for effective generalization. ANN held high fault tolerance with outliers and accurately mimicked complicated nonlinear representation. The RF architecture was not inclined toward overfitting and excellent at effectively conducting data in high dimensions. XGBoost was more responsive to model input features with association and effectively adjusted to datasets in small sizes. AdaBoost was easy to utilize, efficient, and accurately performed for balanced samples. In addition to being adaptable and interpretable, GBDT boasted outstanding generalization and stable performance. They improved the prediction task’s performance and dependability, permitted a highly varied and effective ensemble manner, and broadly handled pertinent issues in numerous fields.

Here, the intelligent model CDN was proposed with interpretability. The basic structure of the CDN (coupling CNN with GRU) is illustrated in Figure 4. The CDN mainly includes convolutional neural networks (CNN), gated recurrent units (GRUs), and fully connected neural networks (or named as dense neural networks, DNN).

For the CDN model, any one of the convolutional neural network layers is elaborated with

$$Y_{ij}=f(W_{ij}^{**}{X}_{ij}+b_{ij}^{**}),$$

(3)

where $Y_{ij}$ and $X_{ij}$ are referred to as the output and input of ith unit of jth convolutional layer, f is referred to as the activation function, and $W_{ij}^{**}$ and $b_{ij}^{**}$ are referred to as the weight part and bias part of a kernel in the convolutional layer.

For the GRU model, the neural unit at the tth time step in any layer of GRU networks is elaborated with [62]

$$z_t=\mathrm{sigm}(W_z{X}_t+U_z{h}_{t-1}),$$

(4)

$$r_t=\mathrm{sigm}(W_r{X}_t+U_r{h}_{t-1}),$$

(5)

$${\tilde{h}}_t=\tan \mathrm{h}(W_h{X}_t+U_h(r_t\otimes h_{t-1})),$$

(6)

$$h_t=(1-z_t)\otimes h_{t-1}+z_t\otimes {\tilde{h}}_t,$$

(7)

$$Y_t=f(W^{*}{h}_t),$$

(8)

where $W_z$, $W_r$, $W_h$, $W^{*}$, $U_z$, $U_r$, and $U_h$ are referred to weight values in the unit, sigmoid function can be elaborated with $\mathrm{sigm}(x)={\displaystyle \frac{1}{1+e^{-x}}}$, hyperbolic tangent function can be elaborated with $\tan \mathrm{h}(x)={\displaystyle \frac{e^x-e^{-x}}{e^x+e^{-x}}}$, $z_t$ is referred to as the update gate, $r_t$ is referred to as the reset gate, ${\tilde{h}}_t$ is referred to as the candidate activation, X_t and Y_t are referred to as the input and output of each unit of the layer, $\otimes$ is referred to as the Hadamard product between vectors, and f is referred to as the activation function.

For fully connected neural networks, the net layers are elaborated, respectively, with

$$a_j^{(1)}=f(W_j^{(1)}\mathit{x}+b_j^{(1)}),$$

(9)

$$a_j^{(m)}=f(W_j^{(m)}{\mathit{a}}^{(m-1)}+b_j^{(m)}),$$

(10)

$$y_j=f(W_j^{(4)}{\mathit{a}}^{(3)}+b_j^{(4)}),$$

(11)

where f is referred to as the activation function, $W_j^{(k)}$ and $b_j^{(k)}$ are referred to as the weight part and bias part of kth layer and jth unit in the fully connected neural networks, x is referred to as the input vector in the training procedure, and a⁽ⁱ⁾ is referred to as the vector comprising unit output values in ith hidden layer.

3.2. Hyperparameters Optimization

Except for the basic structure illustrated in Figure 4, the coupling intelligent structure setup included activation function, initialization distribution, optimizer, and loss function. The activation function included Rectified Linear Unit (ReLU) [63] and softmax unit; the weights and biases in the network were set according to Gaussian normal uniform distribution; and a dropout layer was added into the model structure to prevent overfitting in the training process. At the same time, in order to cope with the classification problem, one-hot was adopted to encode the hierarchical labeled data of the prediction target, and the categorical cross-entropy was used as the loss function in the network. The one-hot was used to encode the hierarchical labeled data of the prediction target to cope with the multiclassification problem, and the categorical cross-entropy was used as the loss function in the network. The root mean square propagation (RMSprop) algorithm was used for optimization function in the training procedure.

3.3. Evaluation Metrics

The investigation utilized 5 evaluation metrics for intelligent models to assess prediction results, comprising accuracy (ACC), precision (PRE), recall (REC), and F1-score, which could practically evaluate the classifying effect of the intelligent prediction methods. The model’s prediction results across all classes were evaluated using the metric ACC. The model’s predictive ability was evaluated at each rockburst intensity class using the metric PRE. The ratio of accurately predicted samples to all positively labeled samples was measured by the metric REC. The PRE and the REC were combined to generate the metric F1 score. Based on the real label and prediction classes, the symbols from the equations comprised true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). In order to keep the decimal point accuracy of the index value consistent, the value of the metric PRE took 4 decimal places, and the ACC took 2 decimal places after changing into a percentage. Moreover, generalization loss rate (GLR) was proposed in this work to measure the generalization capability of models, where MET represented any metric obtained from training datasets. The formulas for these evaluation metrics are as follows.

$$\mathrm{ACC}={\displaystyle \frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}}\times 100\%,$$

(12)

$$\mathrm{PRE}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}},$$

(13)

$$\mathrm{REC}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}},$$

(14)

$$\mathrm{F}1 \mathrm{score}={\displaystyle \frac{2\times \mathrm{PRE}\times \mathrm{REC}}{\mathrm{PRE}+\mathrm{REC}}},$$

(15)

$$\mathrm{GLR}={\displaystyle \frac{\left|\mathrm{MET}-{\mathrm{MET}}^{\prime }\right|}{\mathrm{MET}}}\times 100\%,$$

(16)

4. Results and Validation

4.1. Feature Extraction

The extraction of feature for prediction is crucial for determining the accuracy of rockburst prediction. A rational method for acquiring input features can reduce the model’s learning difficulty and enhance its generalization ability, with minimal performance loss. Historically, the choice of rockburst prediction parameters relied primarily on correlation analyses, but this sort of approach failed to adequately reveal the relationships between features and the prediction target. Consequently, the feature extraction (FE) method proposed in this section offered greater interpretability. To mitigate overfitting risks and improve the performance of rockburst prediction models, we introduced an artificial intelligence method-based feature extraction method.

The microseismic monitoring data (MS) were integrated, among which the cumulative number of microseismic events, cumulative microseismic energy, cumulative microseismic apparent volume, and their respective rate-of-change parameters occurring before the rockburst event were derived. Firstly, importance ranking of MS characteristics was performed by accuracy, based on inputting each characteristic into a three-layer fully connected neural network, with results listed in

Table 2. Importance ranking results of microseismic parameters.

Microseismic Parameters	ACC (%)
E	57.33
RE	56.90
RN	53.46
N	52.41
V	51.87
RV	46.75

.

According to the principle of adding in the best and reducing the best by the importance ranking of microseismic parameters, we selected combinations into the model and then verified its predictive performance, with combinations demonstrated in

Table 3. Combinations of MS parameters for model inputs.

No.	Input Combination
1	E
2	E, RE
3	E, RE, RN
4	E, RE, RN, N
5	E, RE, RN, N, V
6	E, RE, RN, N, V, RV
7	RE, RN, N, V, RV
8	RN, N, V, RV
9	N, V, RV
10	V, RV
11	RV

. Such a feature engineering process could reveal the relationship between the characteristic combinations and the classification level of rockbursts, visualized in Figure 5. After comparing the predictive performance of the selected combinations of microseismic parameters, the fifth combination was observed as the final optimized input combination. This method greatly reduced the subjective interference and obtained the optimal combination with the advantage of automatic and adaptive feature extraction by a deep neural network algorithm.

In order to validate the advancement of this approach, the correlated feature selection (CFS) algorithm proposed in the previous study was compared here [52]. The CFS approach was similar to the proposed FE approach to reveal the relationship between metrics and categories and was briefly described as centered on the use of a heuristic approach to assess the superior value of a subset of features. The heuristic method removed features that did not work for category prediction and identified features that were highly correlated with other features. The heuristic equation is

$${\mathrm{Merit}}_{\mathrm{S}}={\displaystyle \frac{k{\overline{r}}_{\mathrm{cf}}}{\sqrt{k+k(k-1){\overline{r}}_{\mathrm{ff}}}}},$$

(17)

where Merit_S is the heuristic excellence value of the feature subset S containing k features; ${\overline{r}}_{\mathrm{cf}}$ is the feature–category average correlation; and ${\overline{r}}_{\mathrm{ff}}$ is the feature–feature average correlation. The greater the result of Merit_S becomes, the larger the feature–category average correlation becomes, and the smaller the feature–feature average redundance becomes; otherwise, the opposite occurs. In the previous study, the four microseismic parameters, including E, N, V, and R_E, were selected as the input parameters of rockburst prediction model by CFS method.

After obtaining the respective optimal feature combinations, the prediction performance of the CDN model with these two feature engineering methods in the model training-validation procedure is demonstrated in Figure 6 and Figure 7, respectively. In the meantime, the basic performance of the confusion matrices of the two methods is demonstrated through Figure 8 and Figure 9, and the comparison among the predictive metrics of the two feature engineering methods’ predictions is listed in

Table 4. Comparison among predictive metrics of prediction modes from the two feature engineering methods.

Metrics	CFS	FE
ACC	79.03%	83.87%
PRE	0.7943	0.8448
REC	0.7903	0.8387
F1 Score	0.7903	0.8396

Note: ACC represents accuracy; PRE represents precision; and REC represents recall.

. For further exploration, the comparative performance among the evaluation metrics of the two feature engineering methods at each rockburst intensity class is illustrated in Figure 10.

As is observed, the model’s prediction results are basically around the true values. The predicted values in Figure 7 are surrounding the true values closer than the predicted values in Figure 6. As illustrated in Figure 8 and Figure 9, the accurate classification is shown by the values on the main diagonal, whereas the other values show biased prediction. As is seen in Figure 10, there exists some performance difference among distinct prediction classes. Among them, the violent class performs the best among all the classes.

4.2. Optimization of Model Parameters

The dataset for each characteristic was preprocessed in a standardized manner to adapt the data to the network model input requirements. The model structural parameters and hyperparameters were optimized by a K-fold cross-validation method based on a grid search strategy. Taking the CDN model as an example, through the comparative analyses for distinct values of K based on the prediction performance, the final optimal value of K was determined to be 5.

Table 5. Model structural parameters and hyperparameter optimization results.

Hyperparameters	Range	Optimization
Batch size	{1, 2 4, 8, 16, 32, 64, 128}	32
Input step size	{5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60}	35
Dropout rate	{0.15, 0.3, 0.45, 0.6, 0.75, 0.9}	0.3
Learning rate	{0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05}	0.0005
Rflr	{0.1, 0.3, 0.5, 0.7, 0.9}	0.1

Note: Rflr represents reduction factor of learning rate.

lists the selection range and final choice of structural parameters as well.

Table 6. Hyperparameters optimization results of comparative models.

Models	Optimization of Hyperparameters
AdaBoost	learning rate: 0.2; n_estimator 100; max_depth: 5; min_samples_split: 20; min_samples_leaf: 3
SVM	kernel:‘rbf’; Parameter C: 1; Gamma: 1; shrinking: True; probability: True; degree: 3
DNN	learning rate: 0.001; reduction factor: 0.7
LightGBM	learning rate: 0.1; n_estimators: 250; min_child_sample: 20; max_depth: 4; num_leaves: 7; subsample: 0.8; colsample_bytree: 0.8
GBDT	learning rate: 0.2; n_estimators: 100; max_features: sqrt; max_depth: 3; min_samples_split: 2; min_samples_leaf: 4; subsample: 0.5
XGboost	learning rate: 0.2; n_estimators: 100; gamma: 0.8; min_child_weight: 2; max_depth: 4; subsample: 0.8; colsample_bytree: 0.5
RF	n_estimators: 250; max_depth: 3; min_samples_split: 4; min_samples_leaf: 1

lists the hyperparameters optimization results of comparative models.

In summary, it was observed that the neural network-based uniparameter sorting method combined with the combination selection performed better than the correlated feature selection method, and it was verified that the group, including E, N, V, R_E, and R_N, was the optimal combination for considered characteristics in this rockburst prediction study. Therefore, through performance comparison among combinations of input features and model parameters in the validation procedure, the selected input feature combination and optimal model parameters were obtained and employed for the rockburst prediction in the latter section of the investigation.

5. Discussion

5.1. Contrast Among Rockburst Prediction Models

After the feature engineering method, the optimal input combination was fed into XGBoost, LightGBM, AdaBoost, Gradient Boosting Decision Tree algorithms, Support Vector Machines, random forests, and the CDN model for training and validation, respectively. Figure 11 demonstrates the evaluation metrics of each artificial intelligence algorithmic architecture considered in the work during the training process.

Through the validation of optimizing the structural parameters and hyperparameters of each model, we obtained the optimal prediction model for various intelligent architectures. Finally, the predictive performance of the models was conducted on completely independent testing data to examine the classification prediction of the obtained models for rockburst intensity. Figure 12 shows the comparison of the evaluation of each intelligent architecture considered in this study on the testing dataset.

It can be seen that the CDN model tops the list in terms of accuracy, precision, and other evaluation metrics when compared with other machine learning models. On the training dataset, the model’s overall prediction accuracy was 83.87%; on the testing dataset, the model’s overall prediction accuracy was 82.61%. Therefore, the CDN model, a coupling intelligent structure, possesses a generalization loss rate of 1.5% for accuracy in the validation and testing process, which is ranked as the best among all the models mentioned in the paper. All of these can prove that the proposed model has the best generalization ability. In order to further explore the classification prediction of the model’s specific class, the prediction performance of the CDN model is evaluated for each intensity. Figure 13 and Figure 14 show the prediction performance of the model on the training dataset and the testing dataset, respectively. The CDN model clearly demonstrates excellent prediction for violent rockburst cases but does not have perfect prediction for the rest of the classes. Due to the calculation and definition of the metric ACC, its evaluation values are identical across various categories. Additionally, the predicted classes from the trained short-term prediction model for rockburst hazards on 23 testing samples are listed in

Table 7. The prediction results for rockburst intensity on testing dataset.

No.	N	lg(E/J)	lg (V/m3)	RN/d−1	lg(RE/(J·d−1))	lg(RV/(m3·d−1))	Prediction Results
1	3	3.668	3.609	0.5	2.89	2.831	I
2	11	5.926	4.141	1.222	4.972	3.187	III
3	6	4.837	3.712	0.848	3.983	2.858	III
4	3	4.376	4.079	1.5	4.075	3.778	III
…
12	3	4.448	4.261	0.333	3.493	3.306	II
…
20	11	4.029	4.944	1.222	3.075	3.99	I
21	5	3.154	3.309	2.5	2.853	3.008	I
22	18	5.602	4.779	1.8	4.602	3.779	III
23	2	1.94	3.25	1	1.639	2.949	I

. The predictions are correct for most of the samples, except for the 3rd, 4th, 12th, and 20th samples.

In order to explore the CDN model’s prediction performance in more detail, Figure 15 shows the model’s specific prediction results, and it can be seen that the model’s prediction results are basically around the true values without excessive differences or the results of incorrect predictions occurring across the classes. Meanwhile, the confusion matrix of the model in Figure 16 proves this conclusion through visualization, and it can be seen that the prediction results of the model are basically distributed in the main diagonal and one cell above and below the main diagonal.

Table 8. Comparison among the CDN model and other similar rockburst prediction studies.

Models	Prediction Accuracy	Reference
Supervised Gradient Boosting	61.22%	[38]
Bayesian Network	78%	[47]
Multinomial Linear Regression	75.2%	[8]
Artificial Neural Network	78%	[46]
Ensemble Learning	80%	[15]
CDN	82.61%	Proposed in the article

lists the comparison among the CDN model and other rockburst prediction studies, which demonstrates the prediction effectiveness of the proposed CDN model.

5.2. Model Interpretation by SHAP Method

Even though the constructed model performs well, it is unable to quantify the correlations among various features or evaluate the impact of local sample cases for model prediction results, which leaves the prediction process with inadequate explanations. Moreover, the model is hard to interpret because of the intricacy from nonlinear network structures. In order to improve comprehension of the rockburst risk prediction mechanism, an integral model-interpretable approach, known as the Shapley additive explanation (SHAP), was presented [64], which combines the locally interpretable approach on the Shapley results obtained from the cooperative game theory. The new technique makes use of Shapley values to explain intelligent prediction. It is feasible to ascertain the precise contribution from the model feature onto the prediction classes and whether or not they hold a positive or negative influence on the outcomes by using the SHAP approach [64]. Although SHAP is excellent at giving local explanations, it is also possible to broaden the perspective by examining the Shapley values [42,48,65].

The overall input feature onto model prediction in the bar plot in Figure 17 is the most common interpretation analysis. For each feature, there exists a color representing the different intensity classification of the prediction model, and the length of each color bar represents the influence value of the feature on each prediction classification [48,66]. Through the analysis of the bar length of each feature, it can be concluded that there exists a difference in terms of each feature onto each class of the prediction model. Meanwhile, it is illustrated that the length of each feature bar in the violent class is the longest, which holds consistent with the conclusion of the best prediction performance for the violent intensity of rockbursts.

The investigation employs the Shapley values to analyze the model prediction from the other manners in the overall perspective. The beeswarm plot in Figure 18 shows the global distribution of features and ranks the features according to the average SHAP value. Each point represents the SHAP value of a feature in the sample, and the color in the figure represents the size of the feature value. Red stands for high value and blue stands for low value; the position of the point is determined by the SHAP value, and the bigger it is, the farther away it is from zero. The advantage of the beeswarm figure is to demonstrate the influence of multiple features at the same time and reveal the relationship between features and predicted values (such as monotonically increasing or nonlinear, etc.) [42]. Ranked features from the top down are listed by their mean absolute SHAP value. Low feature values (blue points) have a negative impact on the prediction, while high feature values (red points) have a high positive impact on the prediction. Therefore, as is observed, there still exist blue points representing the negative effect from the feature onto the prediction class, where the intelligent method requires further improvement for fined feature contribution. Moreover, the vertical stacking that occurs in Figure 18 demonstrates that there is a high density of SHAP values for each feature onto each rockburst intensity classification.

The relationship between the feature value (X axis) and SHAP value (Y axis) has been shown in the dependence scatter plot, Figure 19, where each point represents the feature value and SHAP value of each sample. The gray bars can clearly and comprehensively demonstrate the influence of direction, size, and change in features on the model prediction [66]. Through the scatter plot, it can be observed that the lower the value of the feature, the greater the corresponding SHAP value, indicating that the lower feature value has a greater influence on the positive prediction of the model, while the higher feature value has significant influence on the negative prediction of the model.

Another overall explanation, as seen in the decision plot in Figure 20, exhibits how a complex model makes decisions; that is, how a model obtains its prediction. The decision plot is a literal representation of the SHAP value, which makes it easy to interpret model input and identify the size and direction of the main influence. The gray vertical straight line in the middle of the decision plot marks the basic value of the model prediction, and the colored line is the prediction route, indicating whether each feature moves the output value to a value higher or lower than the average prediction. The feature is located in vertical axis close to the prediction lines for reference. Starting from the bottom of the graph, the forecast line shows how the SHAP value accumulates from the basic value to the final score of the model at the top of the graph. The decision plot is demonstrated and functioned by visual multi-output prediction, the cumulative effect of the interaction display, outlier detection performance, and determination of the typical prediction path.

SHAP is typically utilized for providing overall explanations, while it is also feasible to obtain a local perspective by examining the Shapley values. Force plot is a crucial local manner of explanation for intelligent model interpretation [67]. As is illustrated in Figure 21, it is a force plot of the model prediction for a certain rockburst sample, with the SHAP value accumulated from the base value to the final value 0.38 after five input characteristics. The base value is the original average value of the target variable in all records. Each band shows the influence of its characteristics in pushing the value of the target variable farther or closer to the base value. The wider the stripe is, the higher the absolute value of the effect is. Red stripes indicate that their characteristics push the value to a higher value, while blue stripes indicate that their characteristics push values to lower values. The sum of these contributions pushes the value of the target variable from the start value to the final predicted value [65,67]. In Figure 21, it is illustrated that all five input parameters have red bands, positive contributions, to the predicted value for this particular record; meanwhile, they boost the value of the target variable to reach the final predicted value, which coincides with the conclusion obtained from the section of feature engineering and demonstrates, indirectly, the effectiveness of the FE method.

5.3. Limits and Prospect

Even though the CDN model performed effectively on the gathered rockburst cases from underground geotechnical engineering, there remain some issues that require attention on the prediction model.

(1)

The rockburst sample cases are not sufficiently large yet. Although 119 samples were gathered for the study, this is still an insufficient sample size when compared with other superior intelligent prediction methods. To some extent, an insufficient training dataset may increase the risk of overfitting issues for the model. Therefore, in order to improve the CDN model’s forecasting robustness and generalization, the amount and quality of rockburst cases will keep growing. The applicability of the model in various geotechnical engineering contexts will be investigated, and more rockburst cases must be further gathered.

(2)

In the feature engineering work of rockburst prediction, latter researchers can continue to explore more advanced feature extraction. The microseismic monitoring parameters considered in the model input of this study can be further expanded in future research. Meanwhile, attention should be paid to the time series correspondence and availability of the expanded microseismic parameters in this extension process. This is a very crucial factor in the generalization of models utilizing extended parameters. The point correspondence of time series data will greatly affect the construction of intelligent architecture and the training process of the model and then affect the final prediction performance and generalization ability of the utilized model.

(3)

From the collected database of rockburst intensity, it can be seen that rockburst cases basically occurred in underground metal mines with a high rock strength condition. Through comparative analysis, the prediction based on this sort of mining rockburst data is effective in this study. The CDN model is not suitable for mines with soft lithology because the prediction results may be biased. In the meantime, the applicable input feature in this study is the dynamic time series of microseismic monitoring data during rockburst disaster. However, it will be beneficial for the prediction model input to continue to add up static parameters with a strong correlation to rock lithology by rockburst area. Therefore, it is necessary to further participate in the construction of feature engineering and rockburst prediction intelligence architecture by coupling static parameters with dynamic time series parameters, which may bring more robust and extensive prediction results in the future.

6. Conclusions

In order to forecast short-term rockburst hazards more precisely, a practical and interpretable prediction model based on the CDN model and feature extraction method are developed in this study using six microseismic monitoring parameters as inputs. Therefore, the following conclusions can be drawn:

(1)

On the basis of the rockburst cases gathered from various underground geotechnical engineering databases, the feature extraction method based on neural networks in a direct model prediction manner is conducted in the process of model training to mitigate overfitting risks and improve the performance of rockburst prediction models. The method possesses a better-behaving performance compared with the previous study.

(2)

In the model training and validation process confronted with numerous combination parameter choices, the model structural parameters and hyperparameters are optimized by a cross-validation method based on a grid search strategy. Through comparative analyses based on prediction performance, optimal values for the model parameters are selected.

(3)

By comparing the prediction performance of considered machine learning models and coupling networks CDN, it is concluded that the CDN model not only has advantages in each basic evaluation metric of classification prediction but also has obvious advantages in a novel metric proposed in this paper. The proposed generalization loss rate (GLR) for accuracy in the validation–testing process reaches 1.5%, demonstrating the generalization ability of the rockburst prediction model.

(4)

An interpretable prediction model based on the CDN model coupling CNN with GRU is acquired through model training. The Shapley additive explanations (SHAP) approach is employed to verify the model interpretability by deciphering the model prediction from the perspective of the fined impact of input features. This method greatly expands the understanding for the influence of the model structure and input characteristics of intelligent algorithms on the prediction results.