Advanced monitoring of tailings dam performance using seismic noise and stress models

The global demand for minerals is increasing, driven by the transition to renewable energy sources. This, coupled with declining ore grades, leads to a rise in tailings—waste by-products of mining operations. Tailings dams, designed to retain these by-products, are among the world's largest engineered structures, with an estimated 8100 facilities globally. Despite being subject to similar regulations as conventional water storage dams in many industrialized nations, the likelihood of tailings dam failures is significantly higher, with the risk projected to increase. The catastrophic 2019 Brumadinho dam failure in Brazil, resulting in over 270 deaths and severe environmental damage, underscored the inadequacy of existing monitoring systems. Investigations revealed that installed instrumentation, including piezometers, inclinometers, and remote sensing data, failed to detect significant changes before the failure, highlighting a crucial need for improved monitoring techniques. A 2019 study revealed that the majority of tailings dam practitioners believed their monitoring practices were best in class, yet a prior review identified substantial gaps, including the reliance on outdated point-based sensors and insufficient sensor redundancy. To address these limitations, combining point-based sensors with broader area measurements is recommended. While remote sensing is increasingly used to detect surface changes, geophysical methods offer the potential to detect subsurface changes not readily apparent through remote sensing. Ambient noise interferometry (ANI) is a geophysical technique that uses naturally occurring seismic noise to reconstruct the impulse response of a wavefield. By monitoring changes in the coda waves (scattered waves), relative changes in seismic velocities (dv/v) can be measured. This approach has been applied to monitor volcanoes, landslides, and dams, but its application to tailings dam monitoring is not yet established. Shear wave velocities (Vs) are crucial for evaluating liquefaction susceptibility, and ANI's sensitivity to Vs makes it a potentially valuable tool for tailings dam monitoring by providing information on how Vs changes over time. This study aims to investigate the feasibility and effectiveness of using ANI, coupled with a stress model, to improve tailings dam monitoring.

Existing literature highlights the significant risks associated with tailings dam failures and the limitations of current monitoring practices. Studies such as Clarkson & Williams (2019) and Franks et al. (2021) have emphasized the need for improved monitoring methods, particularly concerning the inadequacy of point-based sensors and insufficient data redundancy. The Brumadinho dam failure (Silva Rotta et al., 2020; Robertson et al., 2019) served as a stark reminder of these limitations, showing that existing technologies failed to provide adequate warning. The use of remote sensing techniques like InSAR (Grebby et al., 2021; Lumbroso et al., 2019, 2021; Carlà et al., 2019) has shown promise in detecting surface deformations, but geophysical methods (Whiteley et al., 2019; Fan et al., 2021; Michalis & Sentenac, 2021; le Breton et al., 2021; Hamlyn & Bird, 2021) are necessary to understand subsurface changes. Previous research into ANI (Snieder, 2002; Campillo & Paul, 2003; Grêt et al., 2006; Sens-Schönfelder & Wegler, 2006) has demonstrated its efficacy in various applications, including volcano monitoring (Sens-Schönfelder & Wegler, 2006) and landslide monitoring (Planès et al., 2016; Mainsant et al., 2012; Bièvre et al., 2018; Olivier et al., 2017). The importance of shear wave velocity (Vs) in assessing liquefaction risk (Andrus & Stokoe, 2000; Youd et al., 2001) and the various methods for its measurement (Hussien & Karray, 2016) are also well-established. This study builds upon this existing body of knowledge by applying ANI to a tailings dam context, leveraging the sensitivity of Vs changes to assess dam performance and integrating this approach with an effective stress model.

The study utilized a geophone array installed at an active mine site in northern Canada. The array consisted of 25 5 Hz geophones in a T-shaped configuration, designed to optimize noise source detection and avoid construction noise. Data were acquired from June to early August 2020 during both active and inactive construction periods. Ambient noise interferometry (ANI) was employed to analyze the vertical component waveform data. Data acquired during periods of active construction were removed due to incoherencies in the cross-correlation waveforms, leaving three hours of data per day from inactive construction periods for processing. Standard ANI processing methodologies were applied to obtain seismic velocity changes (dv/v). These dv/v estimates were then compared with environmental site data, including tailings pond levels, rainfall, atmospheric pressure, and temperature. To validate the dv/v estimates, a power-law relationship between shear wave velocity (Vs) and effective vertical stress was implemented, drawing on the well-established relationships for granular materials. Site-specific parameters for this relationship were determined through power-law regression analyses with bootstrap sampling, using Vs and effective vertical stress data obtained from 52 seismic cone penetration tests (sCPTs). Daily effective vertical stresses were inferred from daily resampled pond data, and used alongside the site-specific parameters to calculate daily Vs values. Relative changes in daily Vs were then computed and compared to the dv/v results from ANI. A grid search and L1 norm minimization were used to determine the optimal depth sensitivity, with bootstrap and Monte Carlo sampling employed to estimate uncertainties. Detailed data processing steps involved detrending, tapering, filtering, and normalization of the raw seismic data. Cross-correlations were computed for each station pair, and stacked to obtain daily cross-correlation waveforms. A reference waveform was created by stacking across the entire data acquisition period. The ‘stretching’ method was then used to obtain the dv/v time series data and correlation coefficients.

The ANI analysis revealed a strong correlation between seismic velocity changes (dv/v) and water level fluctuations in the adjacent tailings pond. Three main trends were observed: (1) a dv/v increase of ~0.6% correlated with decreasing pond water levels; (2) a dv/v decrease of >0.5% followed the highest daily rainfall; and (3) a recovery to pre-rainfall levels in the final week. The effective stress model, calibrated with sCPT data and pond level recordings, successfully replicated these trends. The model indicated that changes in shear wave velocity (Vs) were the primary driver of the observed dv/v changes, mainly occurring at a depth of approximately 16 meters (with 95% confidence intervals from 14.1 m to 17.4 m). The close agreement between the modeled dV/Vs and the observed dv/v provided strong evidence for the effectiveness of the proposed monitoring approach. The α and β parameters obtained from the power-law regression analyses provided valuable information on the material properties of different stratigraphic units (compacted tailings, tailings, and glaciolacustrine clay). While some discrepancies existed between the model and observations, these were attributed to factors such as variations in pore pressure between the dam and the pond, the influence of suction in unsaturated zones, and construction-related noise impacting data coherency. The study found no significant correlations between dv/v and temperature or barometric pressure, possibly due to the relatively short data acquisition period and the limitations of the effective stress model in fully accounting for these factors.

The strong correlation between seismic velocity changes and tailings pond levels demonstrates the potential of ANI as a sensitive tool for monitoring tailings dam performance. The successful replication of observed trends using an effective stress model further validates the approach and supports the interpretation that changes in effective stress, primarily impacting shear wave velocities at a depth of approximately 16 meters, are the main drivers of the observed seismic velocity variations. The integration of sCPT data into the stress model allows for site-specific calibration, increasing the applicability of the method to different tailings dam environments. The ability to estimate the depth of sensitivity is also a significant advantage of this approach. However, the limitations of the model, such as the assumption of a direct relationship between pond level and pore pressure near the dam, should be considered when interpreting the results. The impact of suction in unsaturated soils and the influence of other environmental factors on seismic velocity warrant further investigation. The study's relatively short data acquisition period also limited the ability to fully assess the impact of seasonal changes and long-term trends.

This study successfully demonstrates a novel approach to tailings dam monitoring by combining ambient noise interferometry (ANI) with a site-specific stress model. The method provides highly sensitive (<1%) measurements of in-situ shear wave velocity changes, offering valuable insights into dam performance over time. The integration of sCPT data improves model accuracy and adaptability to different site conditions. This approach represents a significant advancement in tailings dam monitoring, offering a cost-effective and relatively simple method for continuous, real-time monitoring. However, integrating ANI with complementary monitoring technologies, such as piezometers and InSAR, is recommended to overcome limitations related to spatial resolution and the model’s assumptions. Future research should focus on extending the monitoring period to capture seasonal variations, improving the model to account for unsaturated soil behaviour and other environmental parameters, and further investigating the method's ability to detect precursors to failure.

The study's relatively short data acquisition period (approximately 41 days) limited the assessment of long-term trends and seasonal effects on seismic velocities. The model relied on the assumption that changes in pore pressure near the dam were directly related to water levels in the tailings pond, neglecting the influence of factors like hydraulic conductivity and beach width. The effects of suction in the unsaturated zone were also not fully considered. Construction-related noise affected data quality during some periods. Finally, the spatial resolution of the method is site-dependent, and the use of complementary monitoring techniques is recommended for a more comprehensive assessment.