As underground mining and excavations continue to expand at a rapid pace, a significant engineering challenge arises in the instability of the surrounding rock masses (Aydan et al., 2017). Rockburst, a representative instability phenomenon, is caused by the abrupt release of accumulated elastic strain energy, posing a grave threat to the safety of workers and causing extensive damage to underground engineering structures (Basnet et al., 2023). A significant number of researchers are committed to addressing this challenge by elucidating the occurrence mechanism (Askaripour et al., 2022; He et al., 2023), enhancing prediction accuracy (Basnet et al., 2023; Pu et al., 2019), and developing effective control measures (Li et al., 2019; He et al., 2018). Rockburst prediction aims to generate accurate signals prior to the occurrence of these disasters, serving as a prerequisite for controlling and managing the rockburst hazards. Nevertheless, due to its unpredictable emergence and numerous influencing factors, rockburst prediction remains a challenging task that has yet to be fully resolved.

Rockburst prediction is conventionally classified into long-term prediction and short-term prediction (Liang et al., 2020). Long-term rockburst prediction endeavors to utilize rock mechanical parameters for constructing a prediction model with the aim of assessing the rockburst probability and types of diverse surrounding rock masses under assorted field conditions. This task is customarily accomplished during the initial stage of engineering design or excavation (Liang and Zhao, 2022). The objective of short-term rockburst prediction, on the other hand, is to predict the time, types, and damage magnitude of dangerous rockburst events by conducting dynamic and static analyses on real-time monitoring data. It is typically carried out during the excavation period (Jinqiang et al., 2021). To attain satisfactory outcomes, researchers are dedicated to applying diverse methods for rockburst prediction, such as empirical analytical (He et al., 2018; Yang et al., 2018), experimental (Hu et al., 2023; Cheng et al., 2023), numerical (Wang et al., 2021; Manouchehrian and Cai, 2018), intelligent (Adoko and Zvarivadza, 2018; Xue et al., 2023), and expert system (Li et al., 2020) methods. Although each rockburst prediction method has its own advantages, in contrast to intelligent methods, traditional prediction methods hinge on extensive expert experience and meticulous judgment. Therefore, the machine learning method represents a promising alternative and has been adopted by numerous researchers to dissect and handle the intricate and nonlinear process of rockburst prediction.

Microseismic (MS) monitoring is a widely acknowledged and highly effective tool for identifying the dangerous signals and key controlling factors of rockburst types for short-term rockburst prediction. It is capable of monitoring the occurrence of rockbursts by extracting valuable signals that propagate from the fracturing process of rock masses. Recently, by capitalizing on the capabilities of machine learning in handling nonlinear problems, scholars have directed their attention to rockburst prediction based on MS data through the application of existing machine learning algorithms. Such algorithms encompass support vector machine (SVM) (Ji et al., 2020; Jin et al., 2022), convolutional neural network (CNN) (Dong et al., 2023; Zhang et al., 2021; Yin et al., 2021a), variants of RNN (Hu et al., 2023; Di et al., 2023a; Di et al., 2023b), convolutional long short-term memory (ConvLSTM) (Chen et al., 2023; Ma et al., 2021), and ensemble-learning (Liang et al., 2020; Yin et al., 2021b; Liang et al., 2021), among others (Yin et al., 2024a; Yin et al., 2024b; Cheng et al., 2024; Yin et al., 2021c). Zhang et al. (2021) and Yin et al. (2021a) have conducted an exploration of the key microseismic indexes that can characterize the development process of rockbursts and have applied the refined convolutional neural network (CNN) to predict rockbursts. In order to depict the spatiotemporal relationship within microseismic data and process the spatiotemporal indexes for rockburst prediction, Chen et al. (2023) developed a deep learning model founded on a ConvLSTM to forecast short-term rockburst risks. Concurrently, the ensemble-learning methodologies (Liang et al., 2020; Yin et al., 2021b; Liang et al., 2021) have also been employed to acquire a highly potent rockburst prediction model relying on MS parameters.

Previous research on the rockburst prediction first analyzed the distribution features of microseismic data before a rockburst occurred. Then, the rockburst prediction indexes are presented, and the machine learning algorithm is used to describe the relationship between the rockburst prediction indexes and the future rockburst risks and types. However, these studies ignore the influence of the main controlling factors of rockbursts on the distribution of MS events before a rockburst occurs. In the rockburst prediction and classification process, there is no quantitative analysis of the influence of the intra-class variance and inter-class gap of prediction vectors on the prediction accuracy. The prediction vectors are the result of rockburst prediction indexes or MS data processing by deep learning encoders.

A quantitative information model of geology and mining is constructed based on the different types of rockburst risk events. The temporal and spatial distribution characteristics of MS data before a rockburst, in the main controlling factors of rockburst (i.e., geological structure, roof, coal pillar, and high-stress coal mass), are studied. The results show that the distribution features of microseismic and rockburst events of the same type may differ, and different types of events may show similarities in space-time distribution. The relationship between the performance of a deep learning prediction algorithm and the prediction vector of rockbursts is studied. The rockburst prediction model must accurately distinguish the differences between different types of data and accurately describe the intra-class variance of monitoring data.

A rockburst risk and type prediction method based on a convolutional neural network (CNN)-gated recurrent unit (GRU) and prototype learning prediction head is proposed. The CNN-GRU model can generate prediction vectors by integrating both implicit and explicit information derived from original microseismic (MS) data and prediction indexes. Through the use of cross-entropy loss, vector-prototype contrastive loss, and inter-vector-prototype contrastive loss, the model can autonomously manage the intra-class variance and inter-class distance of prediction vectors corresponding to different rockburst risks and types. Extensive experimental results demonstrate that our proposed CNN-GRU model with prototype-based prediction outperforms alternative algorithms in forecasting rockburst risks and types using MS data.

Although the factors influencing rockbursts are complex, four controlling factors affecting the structure and stress of rockbursts are commonly recognized: geological structure, hard rock strata, coal pillars, and mining depth. The geological and mining data related to these four factors from typical rockburst events are first quantified. For each working face, the mining progress over a 10-day sliding window throughout the entire mining process is calculated. The geological and mining information models corresponding to different working faces are constructed. Then, by projecting the MS data to the quantified geological and mining information model, MS events and rockbursts are simply divided into four types, that is, the hard rock roof type, geological structure type, coal pillar type, and mining depth type, as shown in Figure 1. Some events are caused by more than one controlling factor.

Figure 2 shows that the distribution of different types of MS events is disordered and irregular within 10 days before the rockburst. Even for the same type of events, microseismic events are clustered in several central regions and do not show similar distribution characteristics. Meanwhile, the same type of MS events is clustered in several central regions and do not show similar distribution characteristics. During long-term mining, most MS events show the characteristics of irregularity and relatively uniform distribution and occur in the area where the main control factors are active. When MS events occur intensively in some special areas, it indicates that there is a rockburst risk in the area. When MS events occur intensively in some special areas during the short term, this indicates that there is a rockburst risk in these areas.

The sequential expansion of the time and MS energy across the entire working face is illustrated in Figure 3. Over the long term, it is obvious that the main active controlling factors of a rockburst are the same as those of most MS events. This is because there are structural planes that hinder energy propagation among these four main controlling factors, and the occurrence area of most MS events can be used to determine the destination or main path of energy propagation. The variation of MS energy over time scales is disordered. From a short-term perspective within a rockburst occurrence, the energy and frequency of MS events have increased slightly, showing a relatively active state.

In summary, the MS events can be divided into four types according to the time and space projection of MS events on the controlling factors. For the entire mining process, all types of MS events are disordered and irregular in terms of temporal and spatial characteristics. When the same type of MS events evolves from disordered and scattered to ordered and intensive, the main controlling factor area may have a rockburst hazard. However, even in the short term of rockburst occurrence, MS events are clustered within several centers on the time and space scale. Therefore, it is difficult and complex to summarize the temporal and spatial characteristics of the same type of MS data and accurately distinguish between different types.

In this paper, the content of rockburst prediction includes the risk and type. There are two rockburst risk prediction results: dangerous and non-dangerous. There are four rockburst type prediction results: gravity type, coal pillar type, roof type, and tectonic type. For the machine learning-based rockburst risk and type prediction method, the algorithms mainly consist of the encoder, decoder, and prediction head. The encoder is mainly responsible for analyzing the input MS data and extracting data features. For most previous rockburst prediction methods, the function of an encoder was replaced by the calculation formulas of prediction indexes. The decoder analyzes the feature space of the data features and processes them into vectors with rockburst risk and type information. The final prediction head maps such vectors to the probabilities of rockburst risk and rockburst types.

This section takes rockburst risk prediction as an example to study the distribution requirement of decoder output vectors and representative vectors in the prediction head. For the traditional rockburst risk prediction head, it is reasonable to assume that there are two representative vectors, that is, a non-dangerous representative vector w1 and a dangerous representative vector w2. After the rockburst prediction vector v is obtained, the essence of this problem is to determine the probability that v belongs to risk class C according to the relationship between v and the representative vector w _c in the prediction head, which can be regarded as follows: f ω c ⊤ v + b c = c o n s t (1)

The least squares error between the dangerous probability calculated based on the vector v and the target yi can be written as follows: L = 1 2 ∑ n = 1 N ω c ⊤ v n + b c − y n 2 (2) where N is the total number of samples. To simplify the calculation, the C label is changed to y _c = N/N ₁, and the other label is changed to y c ¯ = − N / N 2 . To obtain the maximum likelihood estimate for b _c , setting L with respect to b _c to 0, the following result can be obtained: b c = − 1 N ω c ⊤ N c m c + N c ¯ m c ¯ (3) where m _c and m c ¯ are the two category centers of dangerous and non-dangerous vectors. Setting the derivatives of L with respect to w _c : ∂ L ∂ ω c = ∑ n ∈ C v n v n T + ∑ m ∈ C ¯ v m v m T − N c m c m T − N c ¯ m c ¯ m T ω c − N m c − m c ¯ (4) where m = N c m c + N c ¯ m c ¯ / N . In order to analyze the data distribution requirements of vector v data, the intra-class variance is ∑ n ⊂ C ω c ⊤ x n − m c 2 + ∑ n ⊂ C ¯ ω c ⊤ x n − m c ¯ 2 = ω c ⊤ S w ω c (5) where k is the class index. The S _W is S W = ∑ n ⊂ C v n − m c v n − m c ⊤ + ∑ n ⊂ C ¯ v n − m c ¯ v n − m c ¯ ⊤ (6)

The inter-class distance can be expressed as m c ¯ − m c 2 = ω c ⊤ S B ω c (7) where the S _B is S B = m c ¯ − m c m c ¯ − m c ⊤ (8)

According to Equation 6, the following equation can be obtained. ∑ n ∈ C ¯ v n v n ⊤ + ∑ m ∈ C v m v m ⊤ = S W + N c m c m c ⊤ + N c ¯ m c ¯ m c ¯ ⊤ (9)

Combining Equation 8 and Equation 9, Equation 4 can be transformed to ∑ n ∈ C ¯ v n v n ⊤ + ∑ m ∈ C v m v m ⊤ = S W + N c m c m c ⊤ + N c ¯ m c ¯ m c ¯ ⊤ (10)

Setting Equation 10 to zero, the following equation can be obtained: ω c ∝ S W − 1 m c ¯ − m c (11)

Due to the differences in the monitoring data, SW is positive and not proportional to the unit matrix. Therefore, in the task of rockburst risk prediction, the vector v should not only maximize the variance between categories ( m c ¯ and m _c ) but also minimize the variance within each category SW. These two objects are interactive. This conclusion is also applicable to the prediction of rockburst type. If there is no proper training method, reducing the intra-class variance of vectors within the same class will also lead to the reduction of inter-class distance of vectors. Therefore, in this paper, the decoder output vectors of rockburst risk and type are controlled using a prototype-based prediction head that can accurately control the intra- and inter-class variances of vectors.

According to the analysis in Section 2 and Section 3, the distribution of rockburst data is very complicated, and it is difficult to describe the distribution of rockburst prediction vectors belonging to one rockburst risk class or type by a representative vector in the traditional classifier head. Moreover, the traditional prediction headers based on single or multi-layer perceptrons are updated by backpropagation. This results in the distribution space of the rockburst prediction vectors that cannot be reasonably controlled. Therefore, a novel CNN-GRU algorithm for comprehensively analyzing the original MS data and rockburst prediction indexes is presented to project the MS data to rockburst risk and type prediction vectors by automatically analyzing and fusing the spatiotemporal distribution features of MS data. Then, a prototype-based rockburst risk and type prediction head that can control the inter-class distance and intra-class variance of the prediction vectors is constructed.

To overcome the problems of previous rockburst prediction tasks, this paper constructs quantitative information models of geology and mining based on typical rockburst events. By studying the relationship between the spatial-temporal distribution characteristics of MS data before a rockburst and the main controlling factors of rockbursts, the results show that the distribution features may be different for the same type of MS and rockburst events. Different types of events may show similar distribution features. The quantitative research results on the relationship between the deep learning prediction algorithm performance and prediction vectors show that the rockburst prediction model must accurately distinguish among the various types of rockburst events and also accurately describe the intra-class variance of monitoring data.

Based on the above insights, a novel rockburst risk and type prediction algorithm based on a CNN-GRU model with prototype-based prediction is proposed. The CNN-GRU model consists of an encoder and a decoder, which can produce prediction vectors by fusing implicit and explicit information extracted from the original MS data and prediction indexes. The prototype-based prediction uses cross-entropy loss, vector-prototype contrastive loss, and vector-prototype contrastive loss to automatically control the intra-class variance and inter-class gap of rockburst risk and type prediction vectors. The performance superiority of the proposed algorithm compared with the previous algorithm is verified by the comparison experiment on the data of three mines. The ablation experiment also proves the universality of the proposed prototype-based prediction head in different algorithms for rockburst risk and type prediction.