- Research Letter
- Open Access
Multi-channel waveform clustering: a first look at microseismic multiplets from coalbed methane stimulation
© The Author(s) 2019
- Received: 20 January 2019
- Accepted: 22 April 2019
- Published: 11 May 2019
Interpreting microseismic events triggered by reservoir stimulation (especially hydraulic fracturing) has become a common practice to understand fracture dimension and geometry. In this area of study, the need for accuracy and resolution of microseismic data is relatively high since the object of investigation is relatively small compared to other seismological studies. Hence, a robust tool is necessary to assure the quality of microseismic event locations and support the interpretation. To achieve these primary objectives, we performed a waveform clustering workflow that analyzes all waveforms representing a microseismic event. Using this approach, we identified multiplets suggesting events that are closely located and originated from the same source mechanism. We tested the workflow on microseismic data from coalbed methane stimulation. The method increases our confident on using the dataset for interpretation especially since the monitoring survey is limited by a single borehole array with very minimal spatial coverage.
- Waveform clustering
Non-ideal survey configurations and limited knowledge of parameters in microseismic data processing manifest some degrees of uncertainty in microseismic hypocenter location. While the intention is to capture and sample seismic waves properly, most microseismic experiments are conducted in reasonably practicable operations. Determining microseismic source location is also factored by the lack of information about the medium in which seismic waves travel. Additionally, for most routine workflows that use travel-time inversion method, the resulted source locations are driven by the uncertainty in travel-time information extracted from seismograms.
In an attempt to obtain reliable results and meaningful interpretation from microseismic data, waveform clustering has been used for improving the solution of source information. This owes to earlier findings in seismology that repeating earthquake events originating from the same fault plane will have similar waveforms (Geller and Mueller 1980). In other words, those similar events have similar source mechanism and propagate through the medium with identical properties. A group of similar events is referred to a multiplet or specifically a doublet for two similar earthquake events (Poupinet et al. 1984). By performing waveform clustering for multiplet identification and analysis, we can assess the reliability of source properties (hypocenter location, moment tensor, etc.) and medium properties such as velocity, anisotropy, and attenuation. (Poupinet et al. 1984; Jones et al. 2014; Castellanos and Van der Baan 2015). Thus, multiplet analysis can be a robust Quality Control (QC) tool on key parameters in routine microseismic data processing (Castellanos and Van der Baan 2015).
Application of waveform clustering has been demonstrated in various seismological studies such as seismotectonics (Poupinet et al. 1984; Orozco-alzate 2007; Adelfio et al. 2012; Nakamura et al. 2016), induced seismicity in geothermal fields (Rowe et al. 2002; Moriya et al. 2003; Dyer et al. 2010), oil and gas fields (Arrowsmith and Eisner 2006; De Meersman et al. 2009; Fagan et al. 2013; Jones et al. 2014), and hydraulic fracturing treatment in tight reservoirs (Castellanos and Van der Baan 2015; Kumano and Tamagawa 2016). Those studies have shown the effectiveness of waveform clustering in deducing more accurate and higher resolution subsurface information.
In general, seismic waveform clustering means grouping seismic events based on waveform similarity. Some studies used a representative recording station (Buurman et al. 2013) or a representative receiver component (Moriya et al. 2003). The choice is commonly based on signal-to-noise ratio (S/N). In this study, we simultaneously used all channels and all components representing a seismic event so that we can compare relative changes between different components and different channels. We utilized waveforms in time domain aiming to capture characteristic of P and S waves (polarity, frequency content, phase), S to P time difference, S-to-P amplitude ratio, etc.
To demonstrate our waveform clustering workflow, we evaluated a microseismic dataset from coalbed methane (CBM) stimulation, a pilot project in Indonesia. It was collected from a single array of seismic recording tool (8 channels of 3 component receivers) in a nearly vertical borehole to understand fracture behaviors generated by hydraulic fracturing. We expected that waveform clustering can identify multiplets related to fracturing coal beds. We then checked consistency within each group especially in terms of hypocenter locations. Hence, a compelling basis has been established to ensure reliability of data interpretation.
Multi-channel waveform clustering (MWC) workflow
Microseismic event clustering begins with measuring similarity/dissimilarity of event pairs based on their waveforms. A common metric used in seismology is cross-correlation coefficient of two seismograms from two different events (Arrowsmith and Eisner 2006; De Meersman et al. 2009; Jones et al. 2014; Castellanos and Van der Baan 2015; Kumano and Tamagawa 2016). Cross-correlation coefficient is computed and normalized either in time or frequency domain. The similarity metric can also be determined based on the averaged coherency calculated from normalized cross-spectrum of event pairs (Poupinet et al. 1984; Moriya et al. 2003). Alternatively, one can also simply measure dissimilarity of two waveforms either in time or frequency domain. Using the measured dissimilarity based on normalized spectra of the waveforms has been demonstrated in some published studies (Orozco-Alzate 2007; Fagan et al. 2013).
Finding clusters based on waveform similarity metric can be performed by various workflows. Once similarity metric is determined, clustering can be done by applying a threshold. Event pairs that exceed the similarity threshold will form doublets (Aster and Scott 1993; Moriya et al. 2003; Arrowsmith and Eisner 2006; De Meersman et al. 2009; Jones et al. 2014; Castellanos and Van der Baan 2015). Arrowsmith and Eisner (2006) described a two-step process by first assigning events into doublets followed by grouping doublets into multiplets (chainlike clustering). Alternatively, a group of seismic events can only form multiplets if all pairs within the group meet the threshold criteria, and in fact are strongly correlated (Castellanos and Van der Baan 2015). No formulation exists on how a threshold is defined; the goal is to balance between clustering objectives and data quality. Low S/N data generally require low threshold, whereas a relatively high threshold is needed to differentiate multiplets (De Meersman et al. 2009). Castellanos and Van der Baan (2015) apply a cross-correlation threshold of 90% to be able to group strongly similar events for a QC purpose. One can also iterate the process by applying a relatively high threshold to identify strongly related multiplets and then a relatively low threshold to form multiplet groups (De Meersman et al. 2009). The approach of using a threshold is also used for another similarity metric other than cross-correlation. Moriya et al. (2003) defined multiplets as any group of seismic events having an average coherency above 0.68 within a certain frequency band.
In this study, we referred multi-channel waveform clustering (MWC) as the workflow we used to identify multiplets from a set of multi-component receivers. This consists of several steps: (1) precondition the waveforms, (2) concatenate the preconditioned waveforms from three components, (3) calculate a metric of waveform dissimilarity from pairs of seismic events for each channel, (4) obtain a multi-channel dissimilarity matrix, and (5) perform clustering based on a dissimilarity matrix.
Data preconditioning prior to waveform clustering can involve filtering, normalization, and global waveform cross-correlation. A band-pass filtering is commonly adequate to remove unwanted signals such as low- and high-frequency noise but retain some characters that are useful for clustering. Then, data normalization is performed to compensate variability of energy level from different events. In a multi-component experiment, each channel can be regarded as a vector. Its normalization is based on the highest hypotenuse value on the signal coda. Lastly, time alignment is useful to optimize dissimilarity metric calculation of time-series data. Very poor time alignment can lead to pitfall when comparing the waveforms.
The second step is simply to concatenate the waveforms from the three components (X–Y–Z) in each channel. By implementing this, we expect to have good event representations that contain all phases and directionality of seismic waves as recorded by three different components. We propose this approach as an alternative to averaging the similarity metric from each component (as in Arrowsmith and Eisner 2006).
Once the multi-channel dissimilarity matrix is formed, we perform clustering process using hierarchical agglomerative clustering (HAC). We choose this method for its illustrative advantages in understanding the nature of clustering from the data (Fagan et al. 2013). This method takes the dissimilarity metric to design data cluster in a hierarchical fashion. It starts by assigning each event as a cluster. Each cluster is then recursively merged with another cluster based on the dissimilarity metric. The mechanism of linkage is visualized by a hierarchical tree (i.e., a dendrogram) that allows us to design and choose criterion for data clustering.
Multiplet identification and analysis from CBM microseismic data
The field data we used to implement the multi-channel waveform clustering technique is microseismic data which recorded during stimulation phases of CBM reservoirs in Sanga–Sanga, Kalimantan. Halinda et al. (2013) presented the results and discussed the operational challenges regarding the monitoring program. The dataset was acquired by a single geophone array deployed in a nearly vertical well at around 100 m away from the treatment well. The array was 8 channels of 3-component receivers (one vertical and two horizontal sensors). Each channel was separated with a spacing of 30 m in measured depth along the wellbore. From the shallowest receiver (channel no 1), positioned at depth of 508 m, the array spread over 210 m along the wellbore to the deepest receiver (channel no 8). The recorded data were sampled at 4 kHz and stored as continuous digital signals.
It is highly expected that microseismic events can provide useful information to evaluate the outcome and performance of well treatment. Therefore, we used multiplet identification and analysis using the proposed waveform clustering workflow to assess the reliability of the estimated microseismic locations, so we can comfortably make an interpretation.
As described above, the first thing in the workflow was preconditioning the waveform from all events so they can be optimally used for clustering. We filtered the signals using band-pass filter with corner frequencies of 60 and 550 Hz to retain useful signals for clustering, and then normalized the filtered signals to the maximum energy in each channel (based on all 3 components). Therefore, signals from different events were equalized for clustering purpose. What important in this clustering workflow is the shape of the waveforms that may represent the source mechanism or ray path of the event.
Identified multiplet groups (excluding doublets), depths, back azimuth, and mean of geographical distances
Number of events
Mean ± std. dev.
Back Azimuth (°)
Mean ± std. dev
Averaged pairwise distances (m)
570 ± 2.1
231 ± 3.1
567 ± 1.5
230 ± 0.6
567 ± 0.6
229 ± 4.8
567 ± 0.8
232 ± 4.6
573 ± 0.6
234 ± 3.9
577 ± 0.8
208 ± 2.4
575 ± 0.6
208 ± 1.4
576 ± 1.4
210 ± 5.8
582 ± 1.2
212 ± 0.3
Figure 6 also demonstrates the multi-channel approach has performed effectively compensating poor variables in any sensor point (a combination between which channel and sensor component) with better variables from other sensor points. In case of multiplet groups in area A, channel 7 component Z has the least energy for S waves. Consequently, this sensor point is not a good variable for a standalone data clustering. Since the multi-channel approach was used, this is not an issue as all information from all channels and all waveforms are considered.
By looking at the hypocenter locations (Fig. 5) and the statistic results (Table 1) from each multiplet, we can notice reasonable consistency between the solution of hypocenter locations and waveform similarity. In overall, the hypocenter of all events within each multiplet are closely located and organized in a geologically sensible way. There are only a few potential mislocated events suggested by this study, one event in multiplet groups B3 and C1.
As we observed from Fig. 1a, the located microseismic events appear as a cloud of point sources with NW–SE major orientation. The main fracture system is oriented NW–SE. The dataset also suggests some fractures with nearly perpendicular orientation to the major orientation. Furthermore, the multiplet analysis in this study has identified few multiplets which are positioned parallel to each other following the major orientation of NW–SE (Fig. 5). Based on their hypocenter locations, time of occurrence, and waveforms characteristic that we demonstrated earlier (Fig. 6), we can be confident that those different multiplets originated from different fractures. It suggests that the microseismic cloud represents a fracture network that contains many small fractures. Looking at the located hypocentres, and time of occurrence of each multiplet as well as between different multiplets, we can infer how fractures have been created, and then propagated as the stimulation progress.
In this paper, a workflow to identify seismic multiplets based on waveform similarity has been presented. By incorporating all waveforms from all channels and all components of the recorded events, the method is proven as a robust tool in identifying multiplets from a microseismic cloud. The grouping criterion gives reasonably confident results as illustrated by the waveform plots. Since the input is essentially the recorded waveforms, we completely avoid pitfall from incorrect assumptions such as velocity model. Therefore, it is more effective than clustering based on hypocenter locations. In fact, this method helps to identify any questionable event locations that are not consistent with waveform similarity as discussed in this study. This study has also demonstrated that inferring the spread and trend of hypocenter locations is more prudent with the basis of multiplets because we can (1) identify mislocated events and (2) classify events based on similar source mechanism. This concept is very powerful for many studies that require relatively higher accuracy microseismic solution such as monitoring reservoir stimulation.
The MWC application to the Indonesian CBM stimulation dataset has identified several multiplets among the located microseismic events triggered by injection program. The first look and analysis at those multiplets in this study has gained more confident on the hypocenter locations. This study has also enriched the interpretation of fracture networks in the system. The spatial and temporal analysis of multiplets provides a meaningful interpretation of how fractures have been evolved in the context of injection activity. By implementing this method, we have demonstrated that we can extend microseismic data interpretation to evaluate CBM stimulation program, from looking at fault trend in general to a more detail analysis about the fracture networks.
The author would like to thank to Ardianto, a fellow student of Geophysical Engineering Study Program, Faculty of Mining and Petroleum Engineering, Institut Teknologi Bandung for many fruitful discussions during this study.
All authors read and approved the final manuscript.
The authors declare that they have no competing interests.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
- Adelfio G, Chiodi M, D’Alessandro A, Luzio D, D’Anna G, Mangano G (2012) Simultaneous seismic wave clustering and registration. Comput Geosci 44:60–69View ArticleGoogle Scholar
- Arrowsmith SJ, Eisner L (2006) A technique for identifying microseismic multiplets and application to the Valhall field, North Sea. Geophysics 71:V31–V40View ArticleGoogle Scholar
- Aster RC, Scott J (1993) Comprehensive characterization of waveform similarity in microearthquake data sets. Bull Seismol Soc Am 83:1307–1314Google Scholar
- Buurman H, West ME, Thompson G (2013) The seismicity of the 2009 Redoubt eruption. J Volcanol Geoth Res 259:16–30View ArticleGoogle Scholar
- Castellanos F, van der Baan M (2015) Waveform similarity for quality control of event locations, time picking, and moment tensor solutions. Geophysics 80:WC99–WC106View ArticleGoogle Scholar
- De Meersman K, Kendall J-M, van der Baan M (2009) The 1998 Valhall microseismic data set: an integrated study of relocated sources, seismic multiplets, and S-wave splitting. Geophysics 74:B183–B195View ArticleGoogle Scholar
- Dyer BC, Schanz U, Spillmann T, Ladner F, Ring MOH (2010) Application of microseismic multiplet analysis to the Basel geothermal reservoir stimulation events. Geophys Prospect 58:791–807View ArticleGoogle Scholar
- Fagan D, van Wijk K, Rutledge J (2013) Clustering revisited: A spectral analysis of microseismic events. Geophysics 78:KS41–KS49View ArticleGoogle Scholar
- Geller RJ, Mueller CS (1980) Four similar earthquakes in central California. Geophys Res Lett 7:821–824View ArticleGoogle Scholar
- Halinda D, Sesaro AW, Weatherall GD, Nugraha H, Collecott DJ (2013) The microseismic monitoring of hydraulic fracturing treatments of a CBM through the simultaneous use of downhole monitoring tools in the stimulated and an offset well. A case study—Sanga–Sanga CBM, Mahakam Delta, East Kalimantan, Indonesia. Society of Petroleum EngineersGoogle Scholar
- Jones GA, Kendall JM, Bastow ID, Raymer DG, Wuestefeld A (2014) Characterization of fractures and faults: a multi-component passive microseismic study from the Ekofisk reservoir. Geophys Prospect 62:779–796View ArticleGoogle Scholar
- Kumano Y, Tamagawa T (2016) Hydraulic fractures delineated by a relative microseismic multiplets hypocenter imaging technique: a case of single well monitoring at Eagle Ford shale field. Society of Petrophysicists and Well-Log AnalystsGoogle Scholar
- Moriya H, Niitsuma H, Baria R (2003) Multiplet-clustering analysis reveals structural details within the seismic cloud at the Soultz Geothermal Field, France. Bull Seismol Soc Am 93:1606–1620View ArticleGoogle Scholar
- Nakamura W, Uchida N, Matsuzawa T (2016) Spatial distribution of the faulting types of small earthquakes around the 2011 Tohoku-oki earthquake: a comprehensive search using template events. J Geophys Res 121:2591–2607View ArticleGoogle Scholar
- Orozco-alzate M (2007) Clustering on dissimilarity representations for detecting mislabelled seismic signals at Nevado del Ruiz volcano. Earth Sci Res J 11:131–138Google Scholar
- Poupinet G, Ellsworth WL, Frechet J (1984) Monitoring velocity variations in the crust using earthquake doublets: an application to Calaveras Fault, California. J Geophys Res 89:5719–5731View ArticleGoogle Scholar
- Rowe C, Aster R, Phillips W, Jones R, Borchers B, Fehler M (2002) Using automated, high-precision repicking to improve delineation of microseismic structures at the Soultz geothermal reservoir. Pure Appl Geophys 159:563–596View ArticleGoogle Scholar