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

The study addresses the need for improved monitoring of tailings dams, which have failure rates much higher than conventional water dams and have caused severe human and environmental impacts. Traditional monitoring systems relying on point sensors and limited redundancy have failed to provide adequate precursors in notable failures (e.g., Brumadinho, 2019). The research question is whether ambient noise interferometry (ANI) can provide sensitive, continuous measurements of in-situ seismic velocity (particularly shear-wave velocity, Vs) to track changes in dam performance and stress state. The context includes increasing mining activity, growing tailings volumes, and the limitations of remote sensing (surface changes) versus geophysical methods (subsurface sensitivity). The purpose is to demonstrate ANI at an active tailings dam, relate small dv/v changes to effective stress variations driven by pond level and rainfall, and validate with a stress–Vs model calibrated by seismic cone penetration tests (sCPT).

Prior work highlights gaps in tailings dam monitoring, including reliance on outdated, point-based sensors and the need to combine them with broad-area measurements. Remote sensing (e.g., InSAR) detects surface deformation and has identified precursory signals in some failures, but cannot probe subsurface changes. Geophysical monitoring can sense subsurface processes in dams and landslides. Ambient noise interferometry reconstructs impulse responses from noise and uses coda-wave sensitivity to detect small seismic velocity changes, applied to volcanoes, landslides, and dams, but not yet established for tailings dams. Vs is central to liquefaction assessment; standard methods (sCPT, MASW/SASW, refraction, downhole/crosshole, lab) provide point-in-time characterization and are costly to repeat. Empirical and theoretical work shows Vs depends on effective stress via power-law relationships for granular soils; coda waves in Poisson media are especially sensitive to Vs. These foundations motivate using ANI to monitor temporal Vs changes relevant to dam stability.

Site and array: An active tailings dam (~8 m high, upstream construction) in northern Canada adjacent to a tailings beach and a pond ~200 m from the array. Materials include compacted tailings (dam fill) over hydraulically placed tailings and glaciolacustrine clay; tailings are non-plastic fine sandy silt to silty sand. A T-shaped array of twenty-five 5 Hz vertical-component geophones was deployed: 19 along the dam crest and 6 extending into the tailings beach. Interstation distances ranged from 10 to 180 m. Sensors were buried 5–10 cm depth. Data acquisition: 12 hours/day recordings at 500 Hz over 35–41 days (June–early August 2020), comprising 9 hours during active construction and 3 hours during inactive construction (00:00–03:00 UTC). Because active construction produced incoherent cross-correlations, only the 3-hour inactive periods were used. Environmental data included pond levels, rainfall, barometric pressure, and temperature. Seismic processing: Standard ANI workflow was applied. Pre-processing included trimming to inactive periods, detrending, tapering, bandpass filtering from 5–15 Hz, one-bit time-domain normalization, and spectral whitening. Cross-correlations were computed for all station pairs using 20 s windows without overlap, stacked to daily cross-correlation functions (CCFs). A reference CCF was formed by stacking over the entire acquisition period for each pair. Relative seismic velocity changes (dv/v) were estimated via the stretching method applied to causal and acausal coda windows (±0.5 s to ±3.5 s), then averaged to improve SNR. Quality was tracked with correlation coefficients between daily and reference CCFs. Effective stress Vs model: Vs was linked to effective vertical stress σ' by Vs = α σ'^β, with α (m/s at σ' = 1 kPa) and β (stress sensitivity) as material constants. Layered stratigraphy and unit weights were used to compute daily σ'_v(z) from pond levels using σ'_v = σ_v − u, where pore pressure u = γ_w (z − d_w), and σ_v was obtained by summing layer unit weights multiplied by layer thickness, accounting for moist/saturated conditions relative to the inferred groundwater level (from pond elevations). Unit weights: compacted/coarse tailings 22.0 kN/m³ above and 24.0 kN/m³ below the pond level; tailings 20.5 kN/m³; glaciolacustrine clay 17.0 kN/m³. Site-specific α and β were estimated for three units (compacted/coarse tailings, tailings, clay) via power-law regression using Vs–σ' data from 52 sCPTs (2017–2018), with bootstrap resampling (150,000 samples) to quantify uncertainty. Resulting parameters: compacted/coarse tailings α≈280.5 m/s, β≈0.01 (n=175); tailings α≈55.9 m/s, β≈0.26 (n=415); clay α≈68.1 m/s, β≈0.21 (n=117). Daily Vs(z) profiles were computed for depths from near surface to ~39 m (bedrock). Relative changes dVs/Vs = (Vs − Vs_i)/Vs_i were obtained, where Vs_i denotes baseline. The modeled dV/Vs time series was fitted to observed dv/v using L1 norm across a grid of depth sensitivities z; uncertainty in z was quantified by Monte Carlo sampling (50,000 iterations) drawing α,β from bootstrap-derived distributions. Depth sensitivity was also cross-checked using wavelength considerations with average Vs ≈220 m/s over 5–15 Hz (depth sensitivity ~ one-third wavelength).

- ANI detected small relative seismic velocity changes (<1%) that tracked environmental forcing. Three main trends observed over ~41 days: (1) dv/v increase of ~0.6% over the first month coincident with decreasing pond levels; (2) dv/v decrease >0.5% in the five days following the highest daily rainfall; (3) recovery of dv/v to pre-rainfall levels in the final week. - A one-dimensional effective-stress Vs model using Vs = α σ'^β, driven by daily pond-level changes and calibrated by sCPT, reproduced the observed dv/v trends and magnitudes, indicating that dv/v predominantly reflects Vs changes due to effective stress variations. - The best-fit depth sensitivity for the dv/v signal was z ≈ 15.7 m, with 95% confidence intervals 14.1–17.4 m (from bootstrap plus Monte Carlo analysis). Independent wavelength-based estimates for 5–15 Hz and average Vs ~220 m/s suggested sensitivity depths of ~5–15 m, consistent with the model-derived ~16 m given structural heterogeneity. - The analysis confirms coda-wave sensitivity to Vs in this setting and demonstrates that dv/v can be interpreted as changes in soil stiffness relevant to dam performance. - Calibrated α, β values characterize stratigraphic units, with tailings showing stronger stress dependence (β ~0.26) and compacted/coarse tailings showing near stress-independence (β ~0.01), consistent with cementation effects.

The findings demonstrate that ANI provides a sensitive, continuous indicator of in-situ changes in shear-wave velocity, enabling interpretation in terms of effective stress and soil stiffness within the tailings dam. The close agreement between dv/v and the dV/Vs model indicates that pond-level driven effective stress variations are the dominant control on temporal seismic velocity changes during the study period. This addresses the monitoring challenge by furnishing a subsurface measure that complements surface-based remote sensing and traditional point sensors. The inferred depth sensitivity (~16 m) situates the method’s response within critical dam and foundation zones. The approach has practical implications: monitoring Vs can signal changes in stiffness that precede failure mechanisms (e.g., internal erosion, liquefaction), and thresholds could be established to trigger alerts based on site-specific seasonal variability. While barometric pressure and temperature effects were not significant over the short summer window, longer-term deployments could capture seasonal signatures. Integration with existing instrumentation (piezometers, tensiometers) would improve model fidelity and allow discrimination among hydrologic drivers (pore pressure vs. suction). Overall, ANI combined with an effective stress model offers an operationally feasible, quantitative addition to tailings dam monitoring systems, providing insight into the temporal evolution of internal conditions.

The study advances tailings dam monitoring by integrating ambient noise interferometry with an effective-stress-based Vs model calibrated using sCPT data and pond levels. The method captured sub-percent dv/v changes that correlate inversely with pond elevation, modeled as Vs variations with strong agreement, and localized the dominant sensitivity to ~16 m depth. Calibrated α and β parameters add geotechnical characterization of stratigraphic units. The workflow is computationally light and suitable for near-real-time operations (e.g., rolling three-day averages) and scalable by extending array coverage. Future work should: (i) co-locate geophones with pore pressure and suction instrumentation to refine modeling (including unsaturated mechanics and barometric effects); (ii) extend monitoring durations to resolve seasonal patterns; (iii) investigate sensitivity and warning times for liquefaction-related and erosion-driven failure modes; (iv) optimize array geometry for spatial resolution versus noise robustness; and (v) integrate with multi-sensor systems (InSAR, inclinometers, piezometers) for comprehensive risk management.

- Pore pressure near the array was not directly measured; the model assumes pore pressure changes follow pond level variations, potentially introducing discrepancies due to hydraulic gradients, beach width, and conductivities. - Unsaturated zone suction effects were neglected; capillarity can increase small-strain stiffness (higher Vs) above the water table, affecting dV/Vs. - Noise source variability (e.g., nearby construction) reduced waveform coherence during some periods, impacting dv/v stability and illustrating ANI’s sensitivity to changing noise fields. - Short monitoring duration (~41 days, summer) limited assessment of seasonal temperature and hydrologic effects; barometric pressure effects at depth were neglected due to complexity and likely damping/time delays. - Uncertainties in α and β, especially for compacted/coarse tailings and clay, stem from fewer sCPT samples and heterogeneity, affecting model precision. - Limited spatial coverage (25 geophones over ~180 m) yields site- and frequency-dependent spatial resolution; averaging across pairs may mask localized anomalies. - Data used only during inactive construction (3 hours/day), reducing temporal sampling and possibly missing transient events.