Microseismic Early Warning Process for Mine Roof Based on Multi-Algorithm Fusion

Abstract

Microseismic early warning for roof disaster in excavated coal roadways often suffers from low pertinence and a high false positive rate. This study establishes an intelligent early warning process based on unsupervised learning and a voting mechanism. True triaxial compression and drilling tests were conducted to characterize the acoustic emission responses of coal and rock during fracture. Using 720 h of field microseismic data from a high-gas mine in Shanxi, high-weight precursor features were extracted from time–frequency indicators. Kernel principal component analysis (KPCA) was used to optimize the indicator system, and 49 indicators with weights above 0.08 were selected as model inputs. Five unsupervised clustering algorithms were integrated to establish an ensemble decision-making early warning model. The results show that the model eliminates the drawbacks of single algorithms, achieves accurate roof disaster warning, and correctly distinguishes disaster events from non-disaster high-energy events. The false positive rate is zero on the 720 h field dataset, and the reliability of early warning is significantly improved. This study enhances the reliability of mine roof microseismic warning, enriches roof disaster prediction theories, provides a complete intelligent early warning process for mine roof disaster, and offers important references for deep mining dynamic disaster warning research.

1. Introduction

Underground mining remains the predominant coal extraction method in China, which involves the construction of extensive underground roadway networks with an annual excavation length of approximately 12,000 km [1,2]. Most of these roadways are driven directly in coal seams. As mining depth continuously increases, roadways are increasingly subjected to severe conditions including high in situ stress, high gas pressure, and complex geological constraints. Consequently, the zones ahead of and behind the excavating face are highly vulnerable to dynamic disasters, such as roof fracturing, roof disaster, large roadway deformation, and even coal and gas outbursts [3,4].

From the perspective of rock fracture mechanics, acoustic emission (AE) and microseismic (MS) have the same physical mechanism: both are elastic signals generated by the release of elastic stress waves during the initiation, propagation, and coalescence of microcracks within rock. Their core differences lie only in monitoring scale, energy magnitude, and application scenarios: Acoustic emission (AE) is a laboratory-scale monitoring technique that captures the entire fracture process of rock samples, characterized by low energy and high frequency, and is mostly used to reveal the mesoscopic mechanism of rock fracture and precursor evolution laws. Microseismic, by contrast, is a large-scale field monitoring signal in mines that detects extensive fracture activities in roadway surrounding rock, with higher energy and longer propagation distance, and is suitable for real-time underground disaster monitoring.

Extensive studies have been conducted to explore the precursory response characteristics of microseismic monitoring during coal–rock deformation and fracture. Wang [5] identified that acoustic emission and electromagnetic radiation signals present evident abnormal fluctuations before the occurrence of coal–rock roof hazards at mining faces, and these two types of signals show highly synchronous responses to disaster evolution. Dou [6] combined laboratory acoustic emission experiments and in situ microseismic monitoring to identify seismo-acoustic precursors during coal–rock disaster fracture and established a comprehensive early warning model for rock bursts using a microseismic precursor index system. Li et al. [7] adopted multifractal theory to analyze microseismic signals during rock bursts and found that the parameter Δα increases continuously while Δf(α) decreases steadily before rock bursts. Li et al. [8] reported that sharp fluctuations or abnormal peaks in the daily cumulative energy and event count of microseismic signals can be regarded as typical precursors of high-energy events. Wu et al. [9] investigated microseismic precursors of rock bursts in steeply inclined coal seams and confirmed that the daily cumulative energy and frequency of microseismic events exhibit abrupt declines and abnormal fluctuations prior to rock bursts. Wang [10] proposed that rock bursts are triggered by the combined effect of accumulated unstable energy E, restraining capacity RE, and disturbance intensity DI, among which sufficient energy accumulation is the key prerequisite for rock burst development. He et al. [11] observed an obvious energy quiet period accompanied by an increasing number of microseismic events before rock bursts, and the electromagnetic–microseismic coupled evaluation method was highly consistent with on-site disaster precursors. Xie [12] analyzed acoustic emission signals from goaf rock masses and found that the fractal parameters Δα and Δf(α) show regular variation trends during the progressive deformation and fracture process. Majid Khan [13] used power spectral density analysis and source parameter calculation to classify microseismic events and characterize stress concentrations and verified that spatiotemporal variations in microseismic parameters can effectively indicate strata pressure evolution. Zhang [14] took a rock burst event as a case study and recognized multiple precursory characteristics, including increases in S and A(t) values, continuous decreases in b and Qt values, and a dominant frequency shift toward low-frequency bands. Ma et al. [15] conducted microseismic monitoring during roadway excavation and found that the daily event count and its rate first increase and then decrease, reaching the maximum at the main shock; meanwhile, the signal amplitude increases, and the frequency band shifts to low frequency. On this basis, a multi-parameter microseismic monitoring model was established to improve the reliability of hazard identification [16,17].

Research on early warning for coal and rock dynamic disasters during roadway excavation has also achieved abundant progress. Jing and Zhu et al. [18] developed a full-process simulation system for deep underground engineering structural instability and quantitatively analyzed the evolution features of stress, surface displacement, acoustic emission, resistivity, and electromagnetic radiation signals during the entire deformation and fracture process of roadway surrounding rock and anchorage structures under different support conditions. Jing [19] adopted multi-source information evolution laws including acoustic, electrical, and magnetic signals to realize quantitative evaluation and multi-physical field characterization of the full load–displacement process for surrounding rock with different structural types. Wu [20] revealed that the internal displacement of surrounding rock during roadway excavation follows an S-shaped variation process, which can be divided into three typical stages: initial growth, rapid growth, and stable convergence. He, Gong et al. [21,22] conducted similar simulation monitoring tests for roadway excavation in strata with different dip angles using infrared thermography and clarified the deformation and fracture evolution characteristics of surrounding rock after excavation. Zhao et al. [23] used endoscopic imaging to trace the entire instability process in physical model tests and revealed the mechanical and deformation properties of deep-buried roadways. Chai et al. [24] embedded fiber Bragg grating sensors in similar models to monitor the dynamic deformation process of rock strata. Zhu et al. [25] applied scanning and imaging techniques in large-scale geomechanical tests and realized accurate measurement of surrounding rock convergence. Zhang and Wu et al. [26] constructed a large 3D similar model and analyzed the acoustic emission response characteristics of surrounding rock deformation under mining disturbance. Zhang et al. [27] combined visible light imaging, acoustic emission, and far-infrared monitoring to reproduce the instability process of granite roadways and proposed a multi-physical field coupled monitoring and early warning method for rock bursts induced by roadway fracturing. Jing [28] pointed out that cohesion and uniaxial compressive strength of surrounding rock increase gradually with burial depth, showing an exponential spatial distribution, and the in situ strength of engineering rock mass also exhibits an obvious time-dependent effect during excavation.

In recent years, machine learning has advanced rapidly in geophysics and has been widely used in seismic data processing, subsurface imaging, and geological hazard monitoring. Global research has focused on physics-guided, unsupervised, and multi-source fusion methods to overcome challenges such as limited labels, strong noise, and high nonlinearity, with unsupervised learning performing well especially in microseismic identification [29]. However, most existing studies focus on large-scale exploration and natural earthquake prediction, while research on mine-scale microseismic monitoring for rock burst and roof instability warning, especially unsupervised clustering-based systems, remains insufficient. This paper establishes an unsupervised learning framework customized for mine scenarios, defines decision variables and decision rules mathematically, and proposes a multi-algorithm fusion early warning strategy. The method is site adaptive, data efficient, label-free, and more suitable for underground engineering with scarce disaster samples. Most existing intelligent early warning indicator systems for coal and rock fracture are developed for application in mining faces [30,31]. Furthermore, indicator selection is usually restricted to a small number of parameters related to event energy, which cannot fully extract and utilize the abundant inherent information contained in microseismic waveforms. Meanwhile, mainstream microseismic early warning technologies for coal and rock fracture are primarily designed for monitoring and forecasting large-scale dynamic disasters such as rock bursts [32,33], and their applicability for early warning of roof disaster in excavating roadways still needs to be further systematically verified and analyzed [34,35].

Roof disaster early warning is a typical dynamic process governed by the evolution of stress, fracture propagation, and microseismic signals. However, most current investigations fail to consider the complete chain from feature extraction to multi-algorithm decision making. In this study, a full monitoring and early warning process for roof disaster in coal roadways is established using feature optimization and unsupervised learning fusion. The proposed method forms a complete early warning process, which can provide a useful reference for the research of disaster warning processes in Processes.

2. Experimental Process and Spatiotemporal Characteristics of Acoustic Emission During Coal–Rock Fracture Under True Triaxial Loading

2.1. Experimental System and Setup

Figure 1a illustrates the true triaxial loading system for coal–rock masses, which consists of a main loading frame, vertical and horizontal loading assemblies, independent loading hydraulic pumps, a triaxial pressure chamber, acoustic emission sensor fixtures, and computer-based control software. This equipment is from China University of Mining and Technology (Xuzhou, China). The system adopts electro-hydraulic servo control, with a maximum vertical loading capacity of 1600 kN and a maximum horizontal loading capacity of 600 kN. The loading assemblies in different directions can independently perform loading and unloading operations without mutual interference, supporting a variety of test processes, including continuous loading, graded loading, and cyclic loading–unloading tests under uniaxial, biaxial, and true triaxial stress conditions.

Figure 1b shows the dedicated experimental system for studying the sudden instability process of coal–rock dynamic disasters, which mainly includes a control unit, a loading unit, and a drilling unit. The drill pipe used in the test is 350 mm long and 20 mm in diameter, with a feed rate of 5 mm/min and a rotational speed of 150 rpm. The vertical and horizontal components of the loading system both adopt electro-hydraulic servo control technology, with a maximum vertical loading force of 2000 kN and a maximum horizontal loading force of 800 kN, respectively.

During drilling, typical AE signals were effectively captured, while the mechanical signals generated in the process are classified as conventional background noise and do not require additional specialized processing.

The uniaxial compressive strength of the sandstone used in this study is approximately 30 MPa, and that of the raw coal is about 14 MPa. To meet the requirements of different experimental processes and working conditions, bulk raw coal and sandstone were processed strictly in accordance with the standards recommended by the International Society for Rock Mechanics (ISRM). Five groups of standard cubic specimens with a size of 150 mm × 150 mm × 150 mm were prepared. These specimens include two types: sandstone–raw coal–sandstone composite specimens with a prefabricated simulated roadway of 20 mm in diameter (Figure 2a), and sandstone–raw coal–sandstone composite specimens without any prefabricated structure (Figure 2b). The specimens with prefabricated roadways were used for conventional loading tests, while those without prefabricated structures were adopted for drilling-related tests. Each layer in the composite specimens was designed to have a uniform thickness of 50 mm to ensure stable and controllable mechanical properties between adjacent interlayers. All specimens were fabricated in full compliance with ISRM standards. The dimensional error was controlled within ±0.2 mm, and the surface flatness deviation was less than 0.05 mm. After precise grinding, all specimens exhibited uniform texture, compact structure, and flat surfaces, with no obvious natural joints or cracks observed. In addition, homogeneity tests were performed on all coal and rock specimens by means of density measurement and wave velocity testing. Only specimens with small deviations in physical parameters were selected for formal tests, thereby guaranteeing the reliability and comparability of the experimental results obtained in this study.

2.2. Experimental Procedure and Loading Path

After roadway excavation, the surrounding rock mass behind the excavation face is repeatedly disturbed by continuous mining operations. The originally stable surrounding rock thus undergoes secondary loading in the vertical direction, which further induces local stress concentration, internal deterioration, and progressive damage within the rock mass. In response to this process, a series of true triaxial stress schemes with variable axial pressure and confining pressure are designed in this study, and the detailed arrangement of acoustic emission sensors is specified in

Table 1. Spatial layout and coordinates of acoustic emission sensors.

Sensor Number	Coordinate (mm)			Sensor Number	Coordinate (mm)
	X	Y	Z		X	Y	Z
N1	123.5	0	123.5	N5	93.5	147	93.5
N2	123.5	0	26.5	N6	123.5	147	23.5
N3	26.5	0	56.5	N7	23.5	147	123.5
N4	0	23.5	123.5	N8	0	123.5	23.5

.

2.2.1. Vertical Loading and Lateral Unloading Scheme

During the excavation of roadway surrounding rock, the original in situ stress state is disturbed, and the surrounding rock transforms from a state of isotropic triaxial compression to a triaxial stress state characterized by vertical loading and lateral unloading. Therefore, different relationships between axial and confining pressures were set under true triaxial conditions (Figure 3), and preloading and formal loading were conducted on prefabricated composite specimens simulating the roadway (sandstone–raw coal–sandstone and its composite structures) as follows: Preloading: Triaxial stresses were uniformly applied at a rate of 0.05 MPa/s up to 2 MPa, followed by a constant load holding for 200 s. Formal loading: Stresses were uniformly increased at 0.05 MPa/s to raise the vertical stress σ₁ and lateral stresses σ₂ and σ₃ all to 6 MPa, with σ₃ kept constant until the end of the experiment. Subsequently, σ₁ and σ₂ were continuously loaded to 8 MPa at the same rate, and σ₂ was maintained constant thereafter. After that, σ₁ was loaded to 12 MPa at 0.05 MPa/s and held for 300 s and then further loaded at the same rate until the complete fracture of the specimens. Meanwhile, σ₃ was unloaded to 0 MPa at a rate of 0.02 MPa/s during the loading of σ₁. Acoustic emission (AE) location monitoring was performed throughout the entire loading–unloading fracture test of the coal–rock specimens, with the AE sampling frequency set at 2 MSPS.

The stress–time curve exhibits obvious staged characteristics corresponding to the loading procedure. The vertical principal stress σ₁ increases stepwise at a uniform loading rate and shows a typical ladder-shaped rising trend, while the lateral stress σ₂ rises linearly to a predetermined level and then remains constant, appearing as a horizontal segment on the curve. Meanwhile, the lateral stress σ₃ maintains a stable value in the initial stage and then decreases slowly and linearly at a fixed unloading rate until it is reduced to zero. The combined stress curve clearly reflects the true triaxial stress path of vertical step loading, one-side constant confinement, and the other-side gradual unloading, which leads to a steady increase in deviatoric stress and determines the progressive propagation of internal cracks and the stepwise evolution of acoustic emission signals.

2.2.2. Experimental Scheme of Vertical Loading and Lateral Unloading During Drilling

The drilling equipment was activated to perform the drilling test, and the vibration signals generated by the drilling rig were monitored synchronously. The samples adopted in this test scheme did not contain prefabricated structures to simulate roadway. The drilling tests were conducted using the dedicated dynamic drilling module matched with the true-triaxial geomechanical test system. The drilling device was mounted on the external support of the triaxial pressure chamber, and the drill rod was horizontally aligned perpendicularly to the specimen center to ensure accurate centering and avoid eccentric load interference. The system synchronously collected real-time signals of stress, displacement, and acoustic emission. Drilling and loading were terminated immediately once macroscopic fracture and abrupt load drop occurred in specimens to guarantee experimental safety and data integrity. The integrated apparatus can stably realize continuous drilling under confined true-triaxial stress conditions, ensuring controllable test procedures and credible experimental results.

Both the vertical and horizontal directions were loaded to 2 MPa using a force-control loading scheme at a loading rate of 0.05 MPa/s, and the stress state was maintained for 200 s to achieve stabilization.

Subsequently, uniform loading was applied at a rate of 0.05 MPa/s until the vertical stress σ₁, lateral stress σ₂, and lateral stress σ₃ all reached 6 MPa. The lateral stress σ₃ was maintained constantly until the end of the test.

Afterwards, the vertical stress σ₁ and lateral stress σ₂ were further loaded to 8 MPa at a rate of 0.05 MPa/s, while lateral stress σ₂ was held constant until the end of the test.

Subsequently, the vertical stress σ₁ was loaded to 12 MPa at a rate of 0.05 MPa/s and maintained for 300 s under a constant stress condition.

The drilling system was then activated with a drilling speed of 5 mm/min and a rotational speed of 150 rpm. During the drilling process, vertical loading was applied at a rate of 0.05 MPa/s, while simultaneous unloading was performed in the σ₃ direction at a rate of 0.05 MPa/s until the sample failed, and the test was completed.

An acoustic emission (AE) monitoring system was used to continuously record AE waveforms during the entire vertical loading and lateral unloading process with drilling. The sampling frequency was set to 2 MSPS.

Based on the loading and lateral unloading scheme under drilling-induced disturbance shown in Figure 4, the estimated stress–time curve presents a continuous and coordinated variation trend. The vertical principal stress σ₁ keeps increasing at a constant loading rate and shows a steady linear upward trend, the lateral stress σ₂ is kept unchanged after being loaded to the preset value and forms a horizontal stable section, and the lateral stress σ₃ decreases continuously and linearly during the whole drilling process. The overall stress curve demonstrates the true triaxial stress path coupled with vertical loading, constant lateral confinement, continuous lateral unloading, and drilling-induced disturbance and such a combined stress condition accelerates the initiation and penetration of coal–rock fractures, providing a fundamental mechanical basis for the abnormal fluctuations of acoustic emission signals before instability.

2.3. Temporal Evolution Characteristics of Acoustic Emission Signals

2.3.1. Including Prefabricated Simulated Roadway Coal and Rock Fracture

This study focuses on the stress evolution characteristics of sandstone–raw coal–sandstone composite specimens containing prefabricated boreholes and divides the vertical loading and lateral unloading experiment into two sequential stages: the preloading stage and the formal loading stage. During the entire loading process, the collected acoustic emission signals are processed using event identification and extraction to determine the acoustic emission event count rate (events per second). The time-dependent variation in the applied stress and the corresponding evolution trend of the acoustic emission event count rate during the entire loading process of the composite specimens are presented in Figure 5. The experimental results indicate that the instability events of the sandstone–raw coal–sandstone composite specimens occur in the late loading stage, and both instability events lag behind the time node corresponding to the peak acoustic emission event count rate.

The instability point of the sandstone–raw coal–sandstone composite specimen arises during the stable compression stage of the formal loading phase. At this moment, the acoustic emission event waveform count per second has just exceeded its peak value, indicating that massive early-stage accumulated micro-fractures have further induced high-energy macro-fractures. During the preloading stage, the confining pressure increases gradually, and the acoustic emission event waveform count per second of the composite specimen exhibits frequent and small fluctuations. This phenomenon is caused by the slight structural adjustment of the initial microcracks and pores inside the composite specimen under low-stress conditions, thereby generating continuous and weak acoustic emission signals. Owing to the low stress level, the acoustic emission event count is relatively small and unstable, which reflects the preliminary response characteristics of the internal structure of the sandstone–coal–sandstone composite specimen under a low-stress environment.

In the initial stage of formal loading, with the gradual increase in confining pressure, the number of event waveforms per second of the sandstone–raw coal–sandstone composite increases synchronously, and the two present a significantly positive correlation. This phenomenon demonstrates that the increase in confining pressure strengthens the internal particle interaction of the composite, and the original microcracks gradually expand and connect with each other. Consequently, the occurrence frequency of acoustic emission events rises notably, suggesting that the internal structure of the sandstone–raw coal–sandstone composite has entered the stage of damage accumulation and progressive evolution under continuous stress.

In the middle stage of formal loading, as the confining pressure continues to increase and remains at a relatively constant level, the acoustic emission event count per second for the sandstone–raw coal–sandstone composite shows no continuous growth but fluctuates steadily within a certain range. This behavior can be attributed to the relatively stable development of internal fractures: new microcracks are initiated, and pre-existing cracks propagate continuously, forming a dynamic equilibrium state. Although internal damage still accumulates gradually, the overall number of acoustic emission events remains relatively stable.

In the later stage of formal loading, as the confining pressure is rapidly elevated to a higher level, the acoustic emission event count per second of the sandstone–raw coal–sandstone composite increases sharply, accompanied by intensive clustering of acoustic emission events. This indicates that internal fractures inside the composite expand and penetrate rapidly, the structural damage intensifies continuously, and the specimen approaches the critical state of instability and fracture. The intensive acoustic emission activity directly reflects the rapid release of internal strain energy, which implies that macroscopic fracture is imminent.

Figure 6 illustrates the dynamic evolution of the average time–frequency characteristics of acoustic emission event waveforms during the fracture process of the sandstone–raw coal–sandstone composite under vertical loading and lateral unloading conditions. The experimental results demonstrate that during the critical period of 150 s prior to the instability of the composite specimen, the mean-square amplitude energy and the rectified average value per second show obvious abnormal fluctuations. This phenomenon reflects the continuous adjustment of the internal stress field and the enhanced fracture activity inside the specimen. Meanwhile, the average values of the dominant frequency and secondary frequency oscillate frequently across different frequency bands, accompanied by a significant increase in high-frequency components. These time–frequency indicators maintain relatively high magnitudes and exhibit persistent abnormal fluctuations. Such behavior indicates that the initiation, propagation, and interaction of multi-scale cracks are significantly intensified, which suggests a sharp increase in the risk of coal–rock fracture and instability.

2.3.2. Coal Rock Fracture Under Drilling Conditions

Based on the stress loading evolution characteristics, the vertical loading and lateral unloading test of the sandstone–raw coal–sandstone composite under drilling conditions is divided into three sequential stages: preloading, formal loading, and continuous drilling. Through identification and quantitative analysis of the acquired acoustic emission signals, the acoustic emission event count per second, the applied stress on the specimen, and the temporal evolution trend of the acoustic emission event count are obtained, as illustrated in Figure 7. The experimental results reveal that the acoustic emission event count per second exhibits pronounced fluctuations prior to the instability of the composite specimen. This behavior indicates that the specimen is in a mechanically unstable state, implying that a high-energy fracture event is imminent. From the perspective of energy evolution, such fluctuations represent the external response to the imbalance between energy accumulation and energy release inside the coal–rock mass. During the fluctuation process, the elastic strain energy within the specimen continuously accumulates, while local crack propagation induces intermittent energy release. When the elastic strain energy accumulates to a critical threshold and cracks propagate to form a through-going fracture surface, a sudden high-energy fracture of the specimen is triggered.

Figure 8 illustrates the evolutionary law of the average time–frequency characteristics of acoustic emission event waveforms during the fracture process of the sandstone–raw coal–sandstone composite under vertical loading and lateral unloading conditions during drilling. The study found that the maximum average mean-square amplitude energy per second of the acoustic emission signals generated during composite fracture occurs in the drilling stage. Analysis indicates that neither vertical nor lateral loading alone induced significant deformation of the composite, and the frictional effect between the drill rod and the specimen did not increase substantially. Instead, the combined stress process of vertical loading and lateral unloading significantly increased the fracture susceptibility of the composite specimen.

During the critical 250-s period prior to the instability of the composite specimen during drilling, its time–frequency characteristics are marked by fluctuations in the mean-square amplitude energy and the average rectified value per second, along with an overall state of low energy. This indicates that the rock is in an energy accumulation stage, a process consistent with the principle of energy conservation. Although microcracks have already started to propagate inside the coal–rock mass, most of the energy released by crack propagation is absorbed by the formation of new crack surfaces, frictional energy dissipation, and elastic deformation of the specimen. As a result, no large-scale energy release has yet occurred. The acoustic emission events during this stage are characterized by low energy and high frequency, which fully reflects that the specimen is in the early stage of instability, on the verge of a high-energy fracture event.

Through the aforementioned laboratory tests, the precursor characteristics and evolution laws of acoustic emission signals during coal–rock fracture under true triaxial loading and drilling-induced disturbance have been clarified. On this basis, field microseismic monitoring was carried out to verify and apply the above-mentioned laws under actual engineering conditions in order to extract effective precursor indicators suitable for field roof disaster early warning.

3. Field Monitoring Process and Microseismic Precursor Characteristics of Roadway Roof Disasters

3.1. Microseismic Monitoring Layout and Data Acquisition Process

The experimental mine is a high-gas mine situated in Shanxi Province, mainly focusing on the mining of 3^# and 15^# coal seams. Currently, the mine is in the mining stage of 3^# coal seam. Gas emission monitoring results show that the absolute gas emission of the mine reaches 241.76 m³/min, and the relative gas emission is 19.46 m³/t; in addition, the 3^# coal seam is classified as a Class III coal seam with no coal dust explosion hazard and no spontaneous combustion tendency.

Figure 9 illustrates the overall layout of the second panel area of the mine. This area is equipped with a total of five intake and return air roadways, specifically: Zone 2101 Conveyor Belt West Roadway, Zone 2102 Return Air West Roadway, Zone 2103 Auxiliary Transport West Roadway, Zone 2104 Return Air West Roadway, and Zone 2105 Auxiliary Transport West Roadway. Among them, the 2101 Conveyor Belt West Roadway serves as an air intake roadway and also undertakes the coal transportation function; the 2102 Return Air West Roadway is a dedicated return air roadway; the 2103 Auxiliary Transport West Roadway is an air intake roadway and also serves as a material transportation roadway; the 2104 Return Air West Roadway is a dedicated return air roadway; and the 2105 Auxiliary Transport West Roadway is an air intake roadway and is used for material transportation.

Combined with the overall layout of the mining project and the arrangement of downhill roadway excavation, the Return Air West Roadway of Zone 2102 and the Auxiliary Transport Connection Roadway are selected as the monitoring roadways (coal roadway excavation) for this project. Meanwhile, microseismic sensors are installed in the adjacent 2105 Auxiliary Transport West Roadway to realize real-time monitoring of the excavation process and internal structural changes of the roadway.

Figure 10 and Figure 11 present the top view and cross-sectional view of the five intake and return air roadways in the second panel area, respectively. All of these roadways are arranged in a westward and downhill layout during the excavation process.

The coordinates of microseismic sensors in a certain mine in Shanxi are shown in

Table 2. Microseismic sensor coordinates.

Sensor Number	Coordinates/m
	X	Y	Z
N1	3,987,396.29	527,760.97	128.14
N2	3,987,392.12	527,790.71	126
N3	3,987,387.96	527,820.42	124.29
N4	3,987,383.8	527,850.1	123.04
N5	3,987,379.63	527,879.81	121
N6	3,987,375.49	527,909.51	117.93
N7	3,987,371.3	527,939.22	114.48
N8	3,987,367.13	527,968.93	113
N9	3,987,362.56	528,001.61	111.2
N10	3,987,358.39	528,031.32	108.77
N11	3,987,354.22	528,061.03	107.19
N12	3,987,350.06	528,090.74	103.31

.

As a crucial defensive line for ensuring safe production in coal mines, the microseismic (MS) monitoring system relies heavily on the connection stability of its internal components and the efficiency of its data transmission process. Multiple uniaxial microseismic sensors are deployed throughout the underground coal mine, acting as highly sensitive detectors that continuously track subtle vibration changes in the surrounding rock mass. A well-functioning MS monitoring system ensures the effective capture of weak microseismic signals, which lays a solid foundation for the early identification of surrounding rock instability hazards.

The uniaxial MS sensors adopted in this study have a sensitivity of 100 V/m/s, enabling them to respond sharply to extremely weak microseismic signals and accurately capture even the vibrations caused by minor rock mass deformation. The operating frequency range of these sensors is set between 20 Hz and 4000 Hz, which fully covers the frequency band of most microseismic signals generated during the fracturing process of roadway surrounding rock. Furthermore, the minimum detectable seismic source energy of the sensors is 10² J, allowing for the accurate monitoring of low-energy microseismic events that are difficult to identify through conventional monitoring methods.

As a core component of the entire microseismic monitoring system, the microseismic monitoring host serves as a pivotal connection hub, linking the data processing server through both underground and ground-based photoelectric switches. The underground photoelectric switch plays a critical role in ensuring the efficient and stable transmission of signals collected by microseismic sensors to the ground surface, even in the complex and harsh underground mining environment. The underground working environment is confronted with various adverse interference factors, including electromagnetic interference and high humidity, which easily lead to signal attenuation or waveform distortion in traditional signal transmission methods. In contrast, photoelectric switches adopt optical signal transmission technology, which boasts excellent anti-interference performance and high-speed data transmission efficiency, thereby effectively ensuring the overall quality of signal transmission. Additionally, the ground-based photoelectric switch accurately transmits the signals uploaded from the underground to the data processing server, realizing seamless connection and coordinated transmission of underground and ground data (Figure 12).

3.2. Analysis of Precursor Response Characteristics for Early Warning

Based on the 120 indicators listed in

Table 3. Calculation formulas for different waveform indicators.

Num	Recognition Indicators	Calculation Methods	Num	Recognition Indicators	Calculation Methods
1	Mean	$y={\displaystyle \frac{x_1+x_2+\dots +x_n}{n}}$	11	Decay Time	$Dc=En-M_t$
2	Maximum	$M=\max (x)$	12	Dominant Frequency	$F$
3	Minimum	$m=\min (x)$	13	Kurtosis	$S_{\mathrm{kur}}={\displaystyle \frac{{\displaystyle \sum _{i=1}^n{(x_i-\overline{x})}^4}}{(n-1){\sigma }^4}}$
4	Peak to Peak	$Mm=\max (x)-\min (x)$	14	Signal Factor	$S_{\mathrm{sf}}={\displaystyle \frac{\sqrt{{\displaystyle \sum _{i=1}^n{x}_i^2/n}}}{\overline{x}}}$
5	Standard Deviation	$\sigma ={\left({\displaystyle \frac{1}{n-1}}{\displaystyle \sum _{i=1}^n{(x_i-\overline{x})}^2}\right)}^{{\displaystyle \frac{1}{2}}}$	15	Peak Factor	$S_{\mathrm{cf}}={\displaystyle \frac{\max \left(x\right)-\min \left(x\right)}{\sqrt{{\displaystyle \sum _{i=1}^n{x}_i^2/n}}}}$
6	Root Mean Square	$Q_n=\sqrt{{\displaystyle \frac{x_1^2+x_2^2+\dots +x_n^2}{n}}}$	16	Impulse Factor	$S_{\mathrm{if}}={\displaystyle \frac{\max \left(x\right)-\min \left(x\right)}{\overline{x}}}$
7	Mean of Positive Values	$M_{\mathrm{p}}={\displaystyle \frac{\left|x_1\right|+\left|x_2\right|+\dots +\left|x_n\right|}{n}}$	17	Clearance Factor	$S_{\mathrm{cif}}={\displaystyle \frac{\max \left(x\right)-\min \left(x\right)}{({\displaystyle \sum _{i=1}^n\sqrt{\left|x_i\right|}})/n}}$
8	Energy	$E=x_1^2+x_2^2+\dots +x_{\mathrm{n}}^2$	18	Frequence Difference	$MF={\displaystyle \frac{(F-f_1)+(F-f_2)+\dots +(F-f_1)}{l}}$
9	Duration	$Da=En-Ari$	19	Secondary Frequency CR	$FS={\left({\displaystyle \frac{1}{101}}{\displaystyle \sum _{l=1}^{101}{(f_l-\overline{f})}^2}\right)}^{{\displaystyle \frac{1}{2}}}$
10	Rise Time	$Db=M_t-Ari$	20–120	Secondary Frequencies	$f_l$

, the maximum, mean, minimum, and cumulative values corresponding to each microseismic event waveform index were calculated at hourly intervals for each monitoring period. These derived statistical parameters were then employed as candidate indices to construct the early warning indicator system for this study. Let x denote a valid acoustic emission signal consisting of n sampling points, where x_i (i = 1,2,…,n) represents the amplitude at the i-th sampling point, and y stands for the mean amplitude of the signal. Let Ari be the arrival time, En be the end time, M_t be the time corresponding to the peak amplitude, F be the dominant frequency, and f represent the secondary frequencies. A total of 101 secondary frequencies are considered, with the l-th secondary frequency expressed as f_l (l = 1,2,…,101).

This study utilized 720 consecutive hours of microseismic monitoring data (from 7:05 on 22 March 2025 to 7:05 on 21 April 2025) collected from a coal mine in Shanxi Province to systematically analyze the evolutionary pattern of the hourly time–frequency index (maximum value) of microseismic event waveforms during the mine excavation process. The monitoring data indicate that a roof disaster occurred at the 278th hour of monitoring, with a released energy of 1.1561 × 10⁴ J.

This study adopts 720 h of on-site microseismic data for model establishment and verification. The only roof disaster during the monitoring period occurred at the 278th hour. To distinctly highlight the precursor response characteristics prior to the disaster, only the time–frequency index variation curves of the first 400 h are presented in Figure 13, Figure 14 and Figure 15. No geological disasters occurred from 400 to 720 h, and the microseismic signals gradually became stable; hence, these data are not illustrated in the figures. From the perspective of response characteristics, the time–frequency indicators of the waveforms exhibit obvious abnormal fluctuations prior to the roof disaster. Specifically, before the roof disaster occurred, the peak-to-peak value, standard deviation, rectified average value, and mean-square amplitude energy all showed a state of intensified fluctuation with relatively large fluctuation ranges.

In contrast, after the disaster, these indicators no longer displayed obvious abnormal fluctuations, indicating their high sensitivity to disaster precursors. Meanwhile, indicators including rise time, kurtosis, waveform factor, skewness, dominant frequency, secondary frequency, the average difference between dominant and secondary frequencies, and secondary frequency concentration all exhibited abnormal fluctuations before and after the roof disaster. This phenomenon suggests that these indicators have low sensitivity to disaster precursors and are not suitable for early disaster warning.

3.3. Optimization Process of the Early Warning Indicator System

Figure 16 presents the results of dimensionality reduction analysis based on kernel principal component analysis (KPCA) applied to microseismic data, which were collected during roadway excavation-induced rock fracture in Experimental Mine-2 (a coal mine located in Shanxi Province).

Figure 16a illustrates the information feature proportion after dimensionality reduction, where the horizontal axis represents the feature indices and the vertical axis represents the information proportion. As shown in the figure, the highest information proportion reaches 79%, indicating that key features concentrate the vast majority of valid information from the original microseismic data. This confirms that a small number of key features can effectively represent the core information of the original data, ensuring that the main characteristics of the microseismic signals are retained after dimensionality reduction.

Figure 16b is a scatter plot of the feature data after dimensionality reduction, with the horizontal axis representing the first principal component after KPCA dimensionality reduction and the vertical axis representing the second principal component. The scatter distribution clearly shows that the data points exhibit a certain clustering trend in the two-dimensional space, rather than being randomly distributed. Although the data points are not concentrated at a single point, the obvious clustering trend reflects the inherent correlation between different microseismic features, which is consistent with the internal structural changes of the rock mass during the fracture process.

Figure 17a illustrates the weight distribution of the maximum value, mean value, minimum value, total value, and growth difference of each hourly time–frequency indicator. Figure 17b presents a curve depicting the decreasing trend of these weights. It can be observed that there are still significant differences in the maximum value, mean value, minimum value, total value, and weight of the event time–frequency indicators within each hour.

On this basis, to optimize the input parameters of the intelligent early warning model and improve the accuracy and effectiveness of the early warning system, we have established screening criteria to select warning indicators with a weight value of more than 0.08 (a total of 49 indicators) as the input parameters of the intelligent early warning model. The threshold for indicator screening (weight > 0.08) is comprehensively determined based on the inflection point characteristics of the KPCA weight distribution curve and multiple groups of comparative experiments. An obvious boundary appears at the weight value of 0.08 on the weight curve. Indicators with weights higher than this value possess high contribution degree and abundant information to disaster precursors. Meanwhile, comparative verification proves that this threshold can greatly reduce input dimensions and improve model efficiency and early warning accuracy. Accordingly, a total of 49 indicators with weights greater than 0.08 are finally selected as the model inputs.

The top 20 warning indicators with the highest weight are presented in

Table 4. The top 20 warning indicators with the highest weight.

Weight Ranking	Early Warning Indicator	Weight Ranking	Early Warning Indicator
1	Maximum hourly mean-square amplitude energy	11	Maximum growth difference of waveform factor per hour
2	Maximum hourly kurtosis	12	Maximum hourly standard deviation growth difference
3	Maximum growth difference of hourly rise time	13	Maximum frequency per hour
9	Maximum value of pulse factor per hour	19	Maximum hourly margin factor
10	Maximum frequency concentration per hour	20	Maximum root mean square value per hour

. Among these, the top 10 warning indicators with the highest weight are concentrated in the maximum value of the hourly time–frequency indicators and their time-dependent change rate of the microseismic waveform. This concentration trend indicates that the maximum value of the hourly time–frequency index and its temporal variation rate are the key signal characteristics reflecting the potential disaster risk of the deep excavation roadway roof, which is of great significance for early warning of roof instability.

Based on the field microseismic data, a variety of time–frequency indicators were obtained and further optimized using kernel principal component analysis and weight sorting. By screening out high-contribution precursor indicators, the key input parameters for early warning model were determined. Accordingly, an integrated multi-algorithm fusion early warning process was established to realize intelligent and accurate early warning of roadway roof disaster.

4. Integrated Microseismic Monitoring and Early Warning Process for Roadway Roof Disaster

4.1. Proposed Early Warning Method and Working Procedure

The proposed method follows a complete, sequential process flow, including data preprocessing, feature optimization, multi-algorithm clustering, and voting-based decision making. Clustering analysis, as a core approach for feature-level data fusion, mainly realizes data classification based on the internal correlation and inherent logical connections within the raw datasets. Unlike traditional methods that rely on training models to assign weights to early warning indicators, this proposed technique involves significantly lower subjective interference and has been widely applied in various research fields, including mechanical fault diagnosis.

In the context of roadway excavation engineering, the excavation face, along with its matched supporting facilities, excavation equipment, and monitoring devices, can be integrated into a complete coal–rock dynamic system. Under conventional excavation speed conditions, the stress state of the coal–rock mass presents a periodic cyclic process of “stress concentration–stress release–stress re-concentration”, and the monitoring data collected by various monitoring systems fluctuate stably within a reasonable and fixed range. However, when a roadway roof disaster is imminent, the vibration field and stress field in the coal–rock medium will undergo drastic disturbances, which further leads to obvious “deviation” characteristics in the early warning indicator data monitored by the system.

Establishing an automatic identification algorithm to detect such abnormal data deviation characteristics can support the intelligent early warning of roadway roof disaster through microseismic monitoring, a technical route that is highly consistent with the unsupervised learning attribute of clustering analysis. The proposed method fully leverages the dynamic evolution rules of real-time monitoring data, explores the topological distribution characteristics of abnormal monitoring data, and thus provides a data-driven intelligent technical solution for the early identification and early warning of potential roof disaster.

Before conducting spatial–temporal early warning analysis, data normalization processing must be performed in advance to ensure the accuracy and comparability of subsequent analysis. For non-negative early warning indicators, the corresponding normalization calculation is carried out in strict accordance with Equation (1); for negative early warning indicators, the normalization formula applied is specified in Equation (2).

$$W_{\mathrm{j}}={\displaystyle \frac{x_{ij}-\min (x_{ij})}{\max (x_{ij})-\min (x_{ij})}},$$

(1)

$$W_{\mathrm{j}}={\displaystyle \frac{\max (x_{ij})-x_{ij}}{\max (x_{ij})-\min (x_{ij})}},$$

(2)

In this context, W_j stands for the j-th normalized early warning indicator, while x_i refers to the j-th early warning index corresponding to the i-th sample. The subscript i ranges from 1 to mt, and j ranges from 1 to tn, where mt represents the total number of samples, and nt denotes the total count of early warning indicators.

In the research on early warning of roadway roof disaster, microseismic events with an energy release exceeding 10⁴ J are classified as high-energy events. Due to the intense energy release feature of such events, they pose a significant threat to the stability and structural integrity of the roadway roof. In contrast, the operational state where no high-energy microseismic events are detected is defined as a low-risk state. Under this state, although there may still be potential safety hazards in the roadway roof, the overall risk remains at a relatively low level, which is manageable and controllable.

High-energy microseismic events can be further subdivided into two distinct categories based on their corresponding engineering consequences. Specifically, a high-energy event that does not trigger actual roof disaster is defined as a non-hazardous high-energy event. Although this type of event is accompanied by obvious energy concentration and release, it does not lead to catastrophic damages such as roadway collapse. This outcome is jointly influenced by multiple factors, including the spatial location of energy release, the mechanical properties of the roof rock mass, and the supporting effectiveness of the roadway support structure.

On the other hand, a high-energy event that directly triggers typical roadway disasters (such as roof disaster and surrounding rock collapse) is classified as a hazard-type event. This classification framework is established based on the coupling relationship between the intensity of energy release and the actual impact of disasters. It not only takes full account of the physical characteristics of microseismic energy release but also integrates the occurrence and evolution mechanisms of disasters in practical engineering scenarios. This framework provides a scientific and standardized criterion for the accurate classification of roadway roof risk levels, and lays a solid foundation for building a hierarchical, targeted roof disaster early warning system.

The detailed flow chart of the proposed early warning method is illustrated in Figure 18. The entire early warning process consists of four sequential stages: data preprocessing, feature optimization, multi-algorithm clustering, and voting-based decision making.

At the data input stage, the statistical parameters corresponding to all early warning indicators in each time interval are treated as independent samples and synchronously imported into five typical unsupervised clustering algorithms, namely K-means clustering, K-medoids clustering, hierarchical clustering, spectral clustering, and self-organizing mapping. Each algorithm performs sample clustering in accordance with its inherent clustering mechanism, with the cluster number set to 3. Subsequently, decision variables (Ha, Hb, Hc, Hd, and He) are obtained by identifying whether each sample is classified into the cluster with the smallest sample size. The cluster number k = 3 is determined according to the physical classification of microseismic events and engineering risk levels, corresponding to the low-risk state, non-disaster high-energy event, and roof disaster, respectively. It conforms to the physical mechanism of coal–rock fracture and is verified to be the optimal number for clustering effect.

Ha, Hb, Hc, Hd, and He represent the judgment results of K-means, K-medoids, hierarchical clustering, spectral clustering, and self-organizing mapping algorithms, respectively. When an algorithm classifies a sample into the cluster with the smallest sample size, the corresponding variable is set to 1; otherwise, it is set to 0. Within a continuous 24 h period, G denotes the number of algorithms that identify at least one abnormal sample, and F denotes the number of algorithms that identify at least three abnormal samples. If G < 3, the state is low risk. If G ≥ 3 and F ≤ 2, it is a non-disaster high-energy event. If G ≥ 3 and F > 2, a roof disaster is identified, and an early warning is issued.

The setting of three clusters is designed to correspond to three typical operational scenarios: low-risk state, non-hazardous high-energy events, and actual disaster events, respectively. During the time-domain early warning phase, the risk level of roadway roof disaster is determined by calculating the sum of the five aforementioned decision variables and comparing the total value with the preset threshold. The specific determination criteria are as follows: within a continuous 24 h period, if fewer than three algorithms classify at least one sample in the corresponding time interval into the smallest cluster, the roof disaster is identified as low. Conversely, if no fewer than three algorithms assign at least one sample to the smallest cluster within 24 h, further verification is required to determine whether the number of algorithms that classify at least three samples into the smallest cluster exceeds two. If the number exceeds two, a formal disaster warning will be issued; otherwise, the scenario will be classified as a non-hazardous high-energy event.

This proposed method can be directly applied to the intelligent monitoring process of coal mine roadway roof stability, providing a standardized, process-oriented technical solution for disaster early warning.

The criteria for evaluating the effectiveness of the time-domain early warning process are clearly defined as follows: an early warning signal is deemed accurate if a roadway roof disaster occurs within 24 h after the warning is issued. Meanwhile, specific rules for counting the number of early warnings corresponding to different early warning methods are specified in detail below.

For a single machine learning algorithm, when the algorithm classifies a sample into the cluster with the smallest sample size and a roof disaster occurs within the subsequent 24 h window, this result is recorded as one valid judgment; otherwise, it is counted as one invalid judgment. For the integrated method combining unsupervised machine learning and group decision making, a complete judgment process is considered completed only when no fewer than three out of the five unsupervised machine learning algorithms assign at least one sample to the smallest cluster within a 24 h period.

In addition, the final total number of judgments for the aforementioned integrated method is determined by the third-highest value among the judgment counts obtained from the five individual unsupervised machine learning algorithms. This structured and standardized process effectively enhances the robustness and accuracy of microseismic early warning, making it more applicable to practical engineering scenarios.

All five clustering algorithms are set with the number of clusters k = 3, corresponding to low-risk state, non-disaster high-energy event, and roof disaster, respectively. Other parameters are set as default or optimal values recommended in the related literature to ensure clustering stability and effect (

Table 5. Parameters of the five clustering algorithms used in this study.

Clustering Algorithm	Key Parameters	Value/Setting
K-means clustering	Number of clusters	3
	Initialization method	K-means++
	Maximum iterations	300
K-medoids clustering	Number of clusters	3
	Distance metric	Euclidean distance
	Algorithm	PAM
Hierarchical clustering	Number of clusters	3
	Linkage method	Ward
	Distance metric	Euclidean distance
Spectral clustering	Number of clusters	3
	Nearest neighbors	10
	Kernel function	RBF kernel
Self-organizing map (SOM)	Grid size	10 × 10
	Neighborhood function	Gaussian
	Learning rate	0.01

).

4.2. Field Application Process and Validation Results

Before constructing a roadway disaster early warning model based on time-domain analysis, it is necessary to normalize all indicators in the warning indicator set for a specific period in the target mine. This step is designed to eliminate the negative impact of scale differences among various warning indicators on the model’s prediction accuracy, a key prerequisite for ensuring the reliability of the early warning model.

The normalization process maps raw monitoring data to a unified numerical interval, which not only enables effective comparison among different warning indicators in subsequent analysis but also optimizes the stability of the model and improves its prediction accuracy. This standardized normalization process is an essential link in the early warning workflow, laying a solid foundation for the subsequent construction of the time-domain early warning model and ensuring its adaptability to practical engineering scenarios.

This study takes the excavation roadway of a coal mine in Shanxi Province as the engineering background and selects microseismic monitoring data collected from 7:05 am on 22 March 2025 to 7:05 am on 21 April 2025 (with a total duration of 720 h) to conduct in-depth research on intelligent early warning of roadway roof disaster.

To explore the applicability of different clustering algorithms in the dimensionality reduction and clustering of microseismic data, a weight-based screening strategy was adopted: specifically, selecting warning indicators with a weight value greater than 0.08 as input variables. On this basis, five typical clustering algorithms were compared and analyzed, including K-means clustering, K-center clustering, hierarchical clustering, spectral clustering, and self-organizing map (SOM) clustering. The relevant analysis results are presented in Figure 19, Figure 20 and Figure 21.

In these figures, the first and second dimensions obtained from dimensionality reduction are used as the horizontal and vertical axes, respectively, and different colors are employed to distinguish different clustering categories to visually present the clustering characteristics of each algorithm through data visualization.

The K-means clustering results show that the clustering categories are distributed in a scattered elliptical shape, with obvious overlaps in some regions. This phenomenon indicates that the K-means algorithm is highly sensitive to the selection of initial clustering centers, and its clustering results are easily affected by the density of data distribution, which leads to a limitation of ambiguous classification when dealing with data with non-uniform distribution.

In contrast, the K-center clustering algorithm demonstrates significant advantages in anti-noise performance by taking actual data points as cluster centers. As shown in the figures, the red clustering category presents a compact and concentrated distribution, while the number of discrete points at the edge of the green clustering category is significantly reduced. Compared with the K-means algorithm, the K-center algorithm has stronger robustness in processing noisy data, though it still has the defect of unclear definition of cluster boundaries.

The hierarchical clustering results present a tree-like hierarchical structure through two-dimensional mapping, where the purple clustering category forms independent circular clusters. This structure intuitively reflects the nested similarity characteristics among data points and is particularly suitable for revealing the inherent hierarchical relationships within the data.

Spectral clustering relies on similarity matrices for graph segmentation, and its cluster boundaries exhibit distinct separation characteristics (e.g., the precise partitioning of red and blue clusters along the x-axis). It has remarkable advantages in processing data with nonlinear manifold structures and can effectively capture the complex distribution patterns of microseismic data.

The self-organizing map (SOM) clustering maintains the topological relationship of the original data through a two-dimensional grid structure; the blue clustering category in the visualization presents a continuous band-like distribution, which clearly reflects the dynamic evolution law of microseismic signal characteristics.

Table 6. Information on high-energy events.

Serial Number	Date	What Hour is It /h	X/m	Y/m	Z/m	Magnitude /Mw	Energy/J	Form of Expression
1	2 April 2025 21:05:55	278	527,992.4	3,987,402.6	116.9	−0.03	1.1561 × 104	roof disaster

documents the details of high-energy microseismic events (with event energy exceeding 1 × 10⁴ J) recorded during the roadway excavation process of a coal mine in Shanxi Province, covering the period from 7:05 on March 22, 2025, to 7:05 on April 21, 2025 (a total duration of 720 h). During this monitoring period, only one high-energy event was detected, and this event directly led to a roof disaster in the mining area.

As shown in Figure 22, the daily variation in the maximum gas concentration monitored by the T1 sensor during the period from 22 March 2025 to 21 April 2025 (a total monitoring period of 30 days) is clearly presented. From the perspective of the overall variation trend, during the 30-day monitoring process, although the daily maximum gas concentration exhibited certain fluctuations, all fluctuations were within the normal range, and there were no abnormal phenomena that exceeded the standard fluctuation threshold.

The daily gas concentration values fluctuated between 0.07% and 0.29% throughout the monitoring period, without any signs of sustained increase, sudden sharp jumps, or other abnormal variation trends that could indicate potential safety risks in the mining process. This indicates that the gas concentration in the mining area remained stable during the monitoring period, providing a safe operational environment for the roadway excavation process.

Based on Figure 23 and Figure 24, we conduct a comprehensive analysis of the early warning results for the warning indicators with input weights greater than 0.08. The analysis results show that at the 278th hour of the monitoring period (starting from 7:05 on 22 March 2025), a roof disaster occurred. Within 24 h prior to the disaster, the number of decision variables (Ha, Hb, Hc, Hd, and He) with values exceeding 2 reached three, which fully meets the preset early warning criteria for roof disasters.

This result indicates that the selected warning indicators can effectively capture the potential risk signals of roof disaster before the disaster occurs, verifying the feasibility and effectiveness of the early warning method in practical engineering scenarios. However, during other monitoring periods, individual unsupervised machine learning algorithms exhibit certain limitations. Specifically, when processing complex and variable microseismic data, these single algorithms often misclassify weak dangerous states as actual dangerous states, leading to frequent misjudgments.

The root cause of this phenomenon lies in the fact that a single unsupervised learning algorithm is difficult to accurately identify the subtle differences between weak dangerous states and true dangerous states when faced with complex and variable microseismic data, which ultimately results in inaccurate judgment of disaster risks.

Based on the early warning results listed in

Table 7. Statistics of the early warning results of different early warning methods based on the early warning indicators whose weights are above 0.08.

Evaluation Parameters	K-Means Clustering Algorithm	K-Center Clustering Algorithm	Hierarchical Clustering Algorithm	Spectral Clustering Algorithm	Self-Organizing Feature Mapping Algorithm	Unsupervised Machine Learning and Group Decision Making
Tp	1	1	0	1	0	1
Fp	24	25	1	17	22	0
Pe	0.04	0.04	0	0.06	0	1
Re	1	1	0	1	0	1

, obtained using warning indicators with input weights greater than 0.08, it can be observed that single clustering algorithms exhibit a significantly higher misjudgment rate. For the K-means and K-medoids algorithms, the overwhelming proportion of normal samples in the dataset causes the algorithms to frequently misclassify numerous normal signals as hazardous categories (Ha = 1/Hb = 1), which reveals a serious deficiency in identifying minority-class samples.

In the case of hierarchical clustering and self-organizing map algorithms, no disaster samples were successfully identified at all (Tp = 0). This issue arises because these methods rely heavily on sample density or topological structure, making it difficult for individual abnormal samples to form effective clustering centers and thus leading to missed detections.

In contrast, the ensemble decision-making mechanism issues a formal disaster warning when at least three of the five unsupervised machine learning algorithms assign no fewer than three samples to the cluster with the smallest sample size within the 24 h period preceding the disaster. The results demonstrate that the ensemble decision strategy achieved one true positive (Tp = 1) and zero false positives (Fp = 0), which validates the remarkable superiority of multi-algorithm collaborative decision making in enhancing the accuracy and reliability of microseismic early warning.

A total of 162 warning indicators, derived from the microseismic signals generated by the surrounding rock fracture during the mine excavation process in Shanxi, were simultaneously input into five unsupervised machine learning algorithms. On this basis, the unsupervised machine learning ensemble decision-making process was carried out.

The early warning results, as detailed in

Table 8. Statistics of early warning results of different clustering algorithms based on all early warning indicators.

Evaluation Parameters	K-Means Clustering Algorithm	K-Center Clustering Algorithm	Hierarchical Clustering Algorithm	Spectral Clustering Algorithm	Self-Organizing Feature Mapping Algorithm	Unsupervised Machine Learning and Group Decision Making
Tp	1	1	0	1	0	1
Fp	23	26	3	15	25	0
Pe	0.04	0.04	0	0.06	0	1
Re	1	1	0	1	0	1

, showed no significant fluctuations or effective early warning signals, which indicates the presence of redundant warning indicators in the initial indicator set. To address this issue, we adopted a weight-based indicator screening process, which successfully filtered out the indicators with weak contribution to the characterization of disaster risks. This screening process not only reduced the dimensionality of the model input but also effectively lowered the complexity of the early warning model, laying a solid foundation for the efficient operation of the subsequent early warning process.

5. Discussion

Roof disasters in underground mines are low-frequency, sudden events. High-quality labeled datasets that can accurately mark disaster precursors, risk levels, and fracture types are extremely scarce, making it difficult to obtain sufficient and reliable samples for training supervised learning models. Meanwhile, the distribution of microseismic precursor signals varies significantly under different mining disturbances and in situ stress environments, resulting in poor generalization ability of supervised models trained with fixed labels. In contrast, unsupervised learning can automatically classify states based on the inherent distribution of data without manual labeling in advance and can independently identify risk states from real-time monitoring data. This makes it more suitable for complex and variable on-site monitoring scenarios and better meets the practical demands of online early warning and rapid decision making in mines. Since coal–rock instability and fracture are progressive evolutionary processes, monitoring data present natural clustering characteristics, allowing unsupervised algorithms to effectively mine the inherent aggregation patterns of data and accurately distinguish stable, transitional, and disaster states.

This study adopts unsupervised learning methods for all analysis and modeling without introducing supervised algorithms as benchmarks. Therefore, direct quantitative comparison between the proposed method and mainstream supervised early warning models cannot be conducted. In future research, sufficient labeled samples will be collected, and various supervised algorithms will be introduced for comparative experiments to further verify the comprehensive performance and applicable boundaries of the proposed fusion model. In the experimental aspect, abnormal acoustic emission signals were observed at 250 s before coal–rock fracture in this study, which is basically consistent with the precursor interval of 100–300 s before rock sample fracture reported in the existing literature. However, the experimental conditions in this study are more consistent with on-site drilling-induced disturbance, thus yielding better engineering representativeness.

Based on 720 h of field microseismic data, the proposed method achieved zero false positive warnings, which is significantly superior to most traditional single-algorithm models (with a common false positive rate of 15–40%). This proves that multi-algorithm fusion decision making can effectively suppress noise and misjudgment. Mechanically, multi-feature input is more consistent with the progressive fracture law of coal and rock; statistically, the fusion algorithm reduces the bias and variance of a single model, leading to a remarkable improvement in overall early warning performance.

Nevertheless, the zero false positive rate achieved in this study was verified based on only one single high-energy roof disaster during the monitoring period. The limited sample size cannot provide sufficient statistical support, so the results do not have extensive universal statistical reliability. The experimental result of zero false positives for the ensemble algorithm in Table 7 only reflects the model performance under a single disaster event and cannot be directly generalized to various mines and multiple disaster scenarios. Before large-scale on-site deployment and engineering application, it is still necessary to collect monitoring data of multiple roof disasters in different mining areas and under various geological conditions for repeated verification, expand the number of disaster samples, further calibrate model parameters, continuously optimize the integrated early warning algorithm, and gradually improve the generalization ability and statistical reliability of the model.

6. Conclusions and Process Implications

This study addresses the limitations of poor pertinence and high misjudgment rate in the existing microseismic early warning system for roof disasters in coal mine excavating roadways. Through systematic laboratory experiments, on-site monitoring, and algorithm optimization processes, the key conclusions are summarized as follows:

(1)

True triaxial loading and drilling tests clarified the spatiotemporal AE characteristics of coal and rock fracture. For composite specimens with prefabricated simulated roadways, the instability of the roadway occurs after the peak of the acoustic emission event waveform count, and the abnormal fluctuations of time–frequency indicators 150 s before instability serve as effective precursors for roof disaster. Under drilling conditions, the maximum mean-square amplitude energy appears during the drilling process, and the fluctuations of relevant indicators 250 s before instability can reflect the accumulation of energy, which provides a reliable experimental basis for the identification of disaster precursors.

(2)

The KPCA and weight-based indicator optimization effectively extracted high-sensitivity precursor features. Based on 720 h of on-site microseismic monitoring data, the maximum value and growth difference of time–frequency indicators are identified as high-weight core features. By selecting 49 indicators with a weight greater than 0.08, redundant indicators are effectively eliminated (verified by consistent results with the original 162 full indicators) while retaining the key hazard-sensitive features. This optimization not only reduces the complexity of the early warning model but also ensures the accuracy of the warning system, laying a solid foundation for the practical application of the model.

(3)

The multi-algorithm ensemble warning model achieved zero false positives and significantly improved reliability. Individual algorithms often have shortcomings such as misjudgment or missed detection in practical application; in contrast, the voting-based group decision-making model can accurately classify three typical scenarios: low-risk states, non-disaster high-energy events, and disaster events. Field application results show that the proposed model achieves a false positive rate of zero on the 720 h in-situ microseismic monitoring dataset, which significantly improves the reliability and stability of the microseismic early warning system.

(4)

The established early warning criteria and technical system have strong engineering applicability. By normalizing the warning indicators and defining three clustering categories for risk classification, the model realizes the quantitative judgment of roof disaster within 24 h. This technical route is consistent with the characteristics of unsupervised learning, avoiding reliance on labeled data, and providing a data-driven intelligent solution for roof disaster early warning. Moreover, it offers a complete and operable process reference for the microseismic early warning of roof disaster in underground coal mines, which is highly compatible with engineering practice. This work presents a standardized process for microseismic early warning.