Coal bursts are violent, dynamic failures occurring in underground coal mines. Similar to rockbursts experienced by hardrock mines, coal bursts can disrupt mine operations, damage equipment, and potentially result in injuries or fatalities. Although notoriously difficult to predict, Mark (2016) identified several coal burst “risk factors” that include: significant depth of cover (>450 m), dipping stratigraphy, and the presence of stiff competent rocks in the roof and/or floor. It is difficult to estimate the global impact of coal bursts, because terminology and reporting criteria vary widely between countries, but Mark (2018) identified approximately 280 coal bursts between 1983 and 2017 in the U.S. mines, seven of which resulted in one or more fatalities. In China, 200 coal mines have reported coal bursts that resulted in at least 100 fatalities and 1000 injuries in the past decade (Rong et al., 2022). Likewise, the Upper Silesian Coal basin shared between Poland and the Czech Republic has experienced over 100 “significant events,” some of which resulted in injuries and fatalities (Mutke et al., 2015). Many other countries with coal mines have reported coal bursts, including Germany, Japan, Australia, and Russia.

Seismic monitoring is an important part of rockburst risk management in deep hardrock mines (Mendecki et al., 2010). Hardrock mines primarily employ networks of accelerometers or geophones installed in boreholes drilled from mine workings. Coal mining operations, though, have struggled to adopt this type of seismic monitoring (for more than short-term research projects) due to vast differences in mining rates and scales, regulations governing the operation and placement of electronics, geological complexities, and a variety of other challenges (Swanson et al., 2016). The greatest adoption rates are in the burst-prone Chinese coal mines where regulations mandate the implementation of “monitoring and forewarning” plans (Qi et al., 2015). The scarcity of voluntary (i.e., not government mandated) adoption among many burst-prone coal mines indicates that improvements to monitoring strategies and technologies are needed for coal mines to fully realize the benefits of seismic monitoring.

The most productive form of underground coal extraction is longwall mining. Longwalls mine a block of coal called a panel, for which access tunnels are developed beforehand via continuous mining machines. Typical longwall panel widths vary from 0.1 to 0.3 km, panel lengths can reach in excess of 1.5 km, and mining rates can exceed 10 m of face advance per day. The tunnels on either side of the panel are known as gateroads. Gateroads are subdivided into entries (usually between two and five) by coal pillars left for stability. The gateroad on the side of the panel heading deeper into the mine (e.g., adjacent to the next panel to be mined) is known as the headgate, and the other is known as the tailgate. On a high level, a longwall consists of hydraulic shields for supporting the roof, an armored conveyor belt for removing the coal, and a shearer traveling up and down the face cutting and knocking the coal onto the conveyor belt (Fig. 1). Once sufficient coal is removed, groups of shields advance and the roof collapses behind them forming a caved-out zone known as a “gob” or “goaf.”

Distributed acoustic sensing (DAS) is a relatively new technology that employs rapid laser pulses to sense subtle strains in fiber-optic cable (with fiber lengths of several kilometers or longer). Because DAS can use commodity fiber-optic cable approved for use anywhere in a coal mine, including the previously deployed communication fiber, DAS networks may enable entirely new or hybrid seismic monitoring approaches that are better suited for longwall mines than current monitoring methods. Although at any point along the sensing cable DAS has a much lower sensitivity than a geophone or accelerometer, its distributed nature and per-channel cost-effectiveness have made possible, or greatly improved, a wide range of geophysical sensing applications (Lindsey and Martin, 2021). DAS applications to the mining industry are rare, but a few recent studies have explored the topic. Zeng et al. (2021) temporarily deployed DAS cables in a few different configurations on the floor of a limestone mine to record signals from an accelerated weight-drop source and production blasts. Du Toit et al. (2022) monitored seismicity for two weeks with DAS at an active mine by deploying fiber along the metal mesh, which supports the mine walls, and fiber in a grouted borehole. They then compared the induced seismicity recordings to in-mine microseismic sensors. Luo and Duan (2022) used DAS with a trenched cable on the surface and a cable in a vertical borehole to monitor seismicity and caving associated with longwall coal mining.

Although the bulk of geophysical DAS research focuses on recording vibrations transmitted through solid materials, some DAS studies have recorded the transmission of mechanical waves through fluids. Zuo et al. (2021) characterized the signals transmitted via boat to lake-bed armored fiber cables as the first step in using ocean-bottom cables as receivers for underwater communication. Lior et al. (2021) used seafloor fiber cables for imaging an underwater basin using ambient noise tomography. Efforts to use fiber-optic technology to record sound waves date back at least to Cole et al. (1977), and several high-quality, fiber-based microphones have been commercially available for more than a decade (e.g., Bucaro et al., 2005). However, efforts to use DAS, which is also known as phase-coherent optical time domain reflectometry (OTDR), to construct a distributed array of acoustic sensors is relatively recent. Liu et al. (2021) demonstrated using a phase-based OTDR to create an array of 3D printed pucks wrapped with fiber for locating acoustic source located underwater and in air. Li et al. (2020) used a coherent OTDR to construct an array of acoustic sensors by wrapping fiber around 3D-printed hollow cylinders.

We present a new DAS-based monitoring approach, which consists of fiber-optic microphones and cable fastened to mining equipment, deployed on a longwall, for studying coal bursts occurring near the mining face. Because the seismoacoustic recordings are made in close proximity to burst damage, data of this type may provide insights about damage processes and associated hazards that are not provided by more distant recordings typical of other monitoring methods. Moreover, our approach eliminates or alleviates many of the operational challenges associated with traditional in-mine monitoring of longwall coal mines.

We begin by discussing the fabrication and testing of a DAS-based fiber-optic microphone based on the design of Li et al. (2020). We then detail the deployment of the seismoacoustic array on a longwall in an active underground coal mine. Next, we discuss a simple metric that shows potential for quantifying face damage and could offer insight into burst mechanics. Finally, we compare the advantages and disadvantages of this longwall DAS monitoring approach with traditional in-mine networks, discuss the potential of DAS for understanding burst damage in underground mines, and outline future work needed before routine use of fiber-optic seismoacoustic monitoring can be widely adopted by the mining industry.

The seismoacoustic array consists of fiber-optic cable attached to the longwall and the DAS microphones, which we designed, fabricated, and tested in the lab. The numerical modeling efforts of Li et al. (2020) indicate that microphone sensitivity increases with increasing cylinder radius, decreasing wall thickness, decreasing elastic modulus (stiffness), and increasing Poisson’s ratio. Rather than attempt to build a perfect microphone that optimizes all these parameters, we selected a commercially available polyethylene terephthalate glycol (PETG) cylinder typically used for packaging. PETG was preferred for its low stiffness, high Poisson’s ratio, and cost-effectiveness in relation to other common plastics. Each cylinder had an outer diameter of 104.3 mm, a wall thickness of 0.7 mm, a length of 305 mm, and cost less than a few USD. We used a hand drill and custom extension to wrap approximately 90 m of single-mode, tight-buffered fiber around each cylinder to construct the microphones, which were then interconnected via single-mode signal cable approved by the Mine Safety and Health Administration (MSHA) for use in coal mines. We recorded the data presented in this article using a Treble true-phase DAS interrogator from Terra15 Pty. Ltd. The Treble interrogator is somewhat unique in that it records velocity (i.e., along-fiber deformation rate) rather than strain rate. Table 1 shows the interrogator configuration, which was identical for the lab experiment and field deployment.

Because the performance of the DAS-microphone system depends on numerous factors, including the interrogator, cable type, cylinder geometry, and properties, we found it necessary to first evaluate the acoustic response of the system in a controlled setting. To that end, we constructed a cubic chamber, approximately 1.2 m in length, with one open side. The chamber was lined with jagged foam to dampen sound reflections. We then suspended the DAS microphone about mid-chamber and used a 12 W programmable speaker to play monotones (harmonics) ranging from 75 Hz to 2500 kHz into the chamber. We selected the lower end of this range because the speaker was incapable of producing lower frequencies, whereas the upper end represents a reasonable limit for corner frequencies of mining-induced events as small as moment magnitude −2.0 (Swanson et al., 2016). We tested a measurement microphone, the MiniDSP UMIK1, in an identical manner to compare the DAS microphone’s response to the UMIK1’s known response.

To aggregate the data channels along the DAS microphone into a single data stream, we calculated an average strain rate for each cylinder according to (1) ϵ˙(t)=X˙[t,x2]−X˙[t,x1]|x2−x1|, in which ϵ˙(t) is the along-fiber strain rate (1/s); X˙ is the along-fiber velocity (m/s, output format of interrogator unit); t is the time (s); and x₁ and x₂ are the start and end x positions of the fiber segment wrapping the cylinder. The physical meaning of the measurement is an averaged circumferential strain rate of the cylinder-fiber composite (assuming perfect fiber-cylinder coupling).

We compute a frequency-dependent signal-to-noise ratio, SNR(f), for each harmonic–microphone pair, which provides a means of comparing the frequency sensitivities between the DAS microphone and UMIK1. The SNR(f) is determined from the power spectrum composed of nonoverlapping short-time Fourier transform segments for 1.0 s of data for each microphone’s recording of the harmonics and a quiet period taken as representative of background noise levels. We then divide the sum of the power spectra around the signal source frequency by the sum of the background noise over the same frequencies: (2) SNR(f)=∑f=f1f2S(f)∑f=f1f2B(f), in which S(f) is the signal power spectrum; B(f) is the background power spectrum; f is frequency (Hz); and frequency limits f ₁ and f ₂ are f ± 5 Hz, respectively (Fig. 2a).

The power ratios between the microphones for each harmonic were calculated in a similar manner, but the output of the UMIK1 was first adjusted according to the amplitude calibration curve provided by the manufacturer: (3) P(f)=∑f=f1f2Pmic(f)∑f=f1f2PUMIK1′(f), in which P_mic(f) is the DAS-microphone power spectra; and PUMIK1′(f) is the adjusted UMIK1 microphone power spectra (Fig. 2b).

Unfortunately, due to the complex interactions with various operating system-level features and settings, the sensitivity factor needed to convert P_UMIK1 to absolute sound pressure level is scaled by some unknown constant (α). Moreover, the UMIK1 manufacturer provides no phase information for the microphone response. Therefore, Figure 2b shows a band-limited approximation of the power spectrum of the inverse of the transfer function scaled by α and cannot be used to convert from cylinder strain rate to absolute sound pressure level. However, Figure 2 is still useful in understanding general trends of the microphone response.

Compared to the measurement microphones, the DAS cylinder microphone records lower SNR(f) values for most frequencies except for some bands between 500 and 1200 Hz. The DAS-microphone performance is probably acceptable over the entire tested range but, considering the sharp low-frequency drop-off in both plots of Figure 2, may not be adequate to record frequencies much lower than 75 Hz. The limited band response of the microphone is a major shortcoming, because the anticipated corner frequencies of mine events with moment magnitudes between 0 and 3 range from low hundreds to only a few Hz (Swanson et al., 2016).

We also tested the effects of adding more cable between the interrogator and the microphone, which simulates longer cable runs and, because the microphones are just cable wrapped around plastic cylinders, more microphones in the array. We found that the background noise levels increased slightly, but there was no significant difference in the microphone response. Therefore, if the microphones are similar in their dimensions and length of wrapped cable, they will have very similar responses.

After characterizing the response of a single microphone, we developed an array of DAS microphones connected by signal cable (referred to as “lead cable”) with a grooved block in the center of the cylinders to relieve tension and protect splices (Fig. 3). The entire array consisted of two sections each with seven DAS microphones.

We conducted a field trial of our microphone array at an underground longwall coal mine in the western United States. The active panel was approximately 0.65 km deep. The interrogator was located at a power center (high-voltage transformer) approximately 0.4 km from the longwall. We fastened the fiber-optic signal cable to the monorail, which contains a suspended “accordion-like” structure that manages cables and hoses leading to the mining face by folding up as the longwall advances. The microphone array was connected to the signal cable, and one microphone deployed every ten shields—a span of approximately 17.2 m (Fig. 4). The lead cables were fastened to the longwall hydraulic hoses using zip ties. Because the hydraulic hoses must accommodate movement as shields advance, significantly more slack was needed than the width of a shield. Consequently, the length of the fiber-optic cable between each microphone was approximately 35 m. We performed a tap test to map each distance along the fiber to a physical location, and determined the distance ranges that constituted the microphones and lead cables. In addition, knowing the amount of fiber wrapped around the microphones and length of the lead cables provided bounds for these values. Although it is possible the cable fastened to the monorail recorded useful signals, it would be difficult to spatially orient the fiber because it gets folded up as the longwall advances. Consequently, we focused solely on the microphones and fiber deployed on the longwall.

The seismoacoustic array operated for 80 hr including three full 10 hr shifts and one partial shift. Around hour 80, one of the lead cables was pulled back into the gob, which disabled half of the array and ended data acquisition efforts. During the deployment, 35 seismic events were detected by a surface seismic network, ranging in magnitude from M_L 1.2 to 2.0. However, due to a small data gap, only 33 events were recorded by the DAS array. To manage the scope of this study, we restricted our efforts to examining signals from the 33 events; however, we are confident that the array could detect many more events, especially if the periodic noise associated with longwall operation can be suppressed with filtering.

Figure 5a shows waveforms of E10 with lead cable channels occurring in closely spaced pairs and microphone channels spaced out. The lead cables generally record much higher frequencies near the first arrival and lower frequencies in later parts of the waveform. We observed some cycle-skipping in the lead cable channels (Fig. 5c), which occurs when the strain exceeds the setting-specific dynamic range of the interrogator. The cycle-skipping occurs later in the waveform and will certainly bias σlead′, which we discuss later. We did not observe any cycle-skipping in the microphone channels. The microphone recordings have emergent signals with well-pronounced harmonics around 500 Hz (Fig. 5d). The microphone signals drop off quickly with distance. For event E10, even 50 m from the first arriving lead cable channel, the microphone signal has nearly reached background levels (Fig. 5e).

Figure 6 shows the moveout and σ′ for each event. All the 33 events are visible on both the microphones and the lead cable channels, despite high background noise levels during active mining. The lead cables show a clear moveout with the first arrival indicating the lateral location of the seismic event (see also Fig. 5). The first arrival picks made on the lead cable channels are generally smooth, but there are a few exceptions that have significant outliers or discontinuities (e.g., E23, E25, E30). The shapes of σmic′ (red lines) are highly variable. Several events display narrow spikes near the first arrival (E02, E20, E31, E33), whereas others are almost Gaussian shaped (E03, E10, E12, E19, E28). The narrow spikes could indicate a smaller area at the face that experiences enough damage to generate significant acoustic energy in the workings. The sharp drop-off over distance is due to the rapid attenuation of the relatively high frequencies recorded by the DAS microphones and lack of new noise generated near the microphones that recorded a lower σmic′ Thus, high amplitudes of σmic′ could correspond to face damage near the microphone, and sharp drop-offs could delineate the end of the damage zone. The observed σlead′ values generally follow the same trends as σmic′, but they tend to be rougher and fall off less quickly as a function of face distance (e.g., E03, E04 E11, E20, E27). The apparent difference in propagation characteristics is due to the two segment types effectively recording two different propagation modes; the lead cables record seismic energy as it propagates through the longwall, whereas the microphones record the acoustic energy propagating in the workings.

Figure 7 shows the smoothed amplitude spectra recorded by both microphone and lead cable segments for the short time window used to calculate σ′. The colors show channels that recorded the maximum, the median, and the minimum values of σ′ for each of the 33 events. The shaded regions delineate the first and the last decile, whereas the solid lines are the median of the frequency bins. Each microphone channel has a peak around 500 Hz (Fig. 7a), which corresponds to a spike in the frequency response (Fig. 2b). The microphone channels with the largest σmic′ recorded significant energy in the lower frequency bands (compared to the other microphone channels), which indicates that the microphones can record low frequencies and that, perhaps, these low frequencies correspond to gas liberation associated with the damage at the face. The lead cable channel spectra show little variability (apart from being scaled up), but the slope of the maximum spectra is slightly steeper (Fig. 7b). The high frequencies in lead cable spectra may be artificially inflated by cycle-skipping (Fig. 5c).

Although we currently lack a full transfer function, it is informative to compare the strain-rate measurements recorded by the microphones in the lab and mine settings. Figure 8 shows the relationship between σ′ (x axis) and the maximum absolute value of the short time window (y axis) for both types of recordings. For reference, the maximum values of max(μϵ˙) are 790 and 2620, and the maximum value of σ(μϵ˙) are 180 and 1830 for coal burst recordings and lab recordings, respectively. Although the lab recordings have higher values in both metrics by up to an order of magnitude, this does not necessarily mean that the absolute sound pressure was higher in the lab but could simply be an artifact of the strong frequency response of the microphones. The lab recordings have a linear relationship between σ(μϵ˙) and max(μϵ˙), because the signal is a simple sine wave.

Here, we present a DAS-based monitoring strategy that effectively turns an active longwall into a seismoacoustic array, thereby alleviating many of the shortcomings related to traditional in-mine burst monitoring. We demonstrate that the array can record both seismic and acoustic signals of mining-induced seismicity. We find that a modified standard deviation of the recordings, although the lead cable measures are biased by cycle-skipping, may be beneficial for understanding coal burst-related damage. This type of array is potentially useful for understanding damage related to dynamic failures in other mining contexts as well. However, significant work remains to improve microphone design, data interpretation and processing, and array robustness.

Data used in this study were collected in the manner described in the previous sections. They are considered proprietary as part of the collaborative agreement between the National Institute for Occupational Safety and Health (NIOSH) and the unnamed coal mine. However, the supplemental material, including several recordings from the distributed acoustic sensing (DAS) microphones and waveform plots of all the 33 events, may be found at https://derchambers.com/publications/bssa-2023 (last accessed March 2023).