Xiao Zhuowei, Wang Lifeng, Chen Jianye, Chen Shi. 2026: Detection and location of laboratory acoustic emission events using the DiTing seismic foundation model. Acta Seismologica Sinica, 48(6): 1131-1143. DOI: 10.11939/jass.20260106
Citation: Xiao Zhuowei, Wang Lifeng, Chen Jianye, Chen Shi. 2026: Detection and location of laboratory acoustic emission events using the DiTing seismic foundation model. Acta Seismologica Sinica, 48(6): 1131-1143. DOI: 10.11939/jass.20260106

Detection and location of laboratory acoustic emission events using the DiTing seismic foundation model

  • Laboratory fault experiments allow the progressive development of fracturing to be tracked from initial loading to repeated stick-slip failure under controlled conditions, providing a link between small-scale rupture processes and earthquake preparation. Acoustic-emission (AE) experiments, however, generate large volumes of continuous, high-sampling-rate data. Manual inspection is inefficient, while conventional threshold-based detectors can be sensitive to noise level, sensor response, and experimental configuration. Here we study whether a seismic waveform foundation model pretrained mainly on natural-earthquake data can be applied directly to laboratory AE records and used to construct an event catalog for subsequent analysis.
    The data were acquired during a biaxial loading experiment at the State Key Laboratory of Earthquake Dynamics, Institute of Geology, China Earthquake Administration. The specimen consisted of two granodiorite blocks separated by a mechanically polished, precut interface. Pressure-sensitive film measurements at 5 MPa showed heterogeneous contact stress, with two major maxima at along-fault distances of approximately 63 and 398 mm, referred to as Asperities 1 and 2. Load, displacement, and AE signals were recorded synchronously by 16 sensors at 3 MHz. The two loading axes were first increased until the confining pressure reached 5 MPa. The second principal load was then held constant while the first was driven under displacement control at 0.5, 0.1, and 1 μm/s during different stages, producing an initial loading phase followed by repeated stick-slip cycles.
    Event detection and P-wave picking were performed with the 100-million-parameter DiTing seismic waveform foundation model. Continuous single-channel waveforms were divided into 10 000-sample windows with a 5 000-sample step, corresponding to 3.33 ms per window. The model produced event-detection and P-wave probability curves. Event windows were triggered at a detection probability of 0.2, and P candidates were identified from local maxima above 0.1. When several P candidates occurred within one detected event, the first was retained. This procedure yielded 558 348 candidate P picks across the 16 channels.
    Candidate picks were then associated across sensors using a 0.000 25 s time window. If more than one pick from the same sensor fell within an association window, only the one with the highest P probability was retained. Events required at least five contributing sensors, and picks used for location had to satisfy a travel-time residual tolerance of 0.000 12 s. Locations were obtained by grid search in a homogeneous P-wave velocity model of 5 250 m/s with source depth fixed at 0 m. For each candidate source position, theoretical travel times and origin time were estimated, and the solution with the minimum root-mean-square travel-time residual was selected.
    Two grid configurations were compared. The first searched the full 0−0.30 m by 0−0.30 m plane with a nominal 0.002 m spacing and 22 801 candidate points. The second restricted candidate positions to the known diagonal fault trace, together with the rectangular boundary of the experimental region. Both configurations located 40 952 events. The full-plane solutions had a median of 10 picks and a median RMS residual of 10.094 μs, whereas the constrained solutions had the same median pick count and a median residual of 10.125 μs. Despite these similar residuals, the unconstrained solutions commonly formed a narrow trend near the sensor line rather than on the known fault and also scattered through the plane and along its boundaries. The constrained search, by incorporating the known interface geometry, placed most solutions on the precut fault while retaining boundary solutions for events not well represented by on-fault candidates. Further spatial filtering produced two catalogs with different levels of geometric confidence. Removing rectangular-boundary solutions except for the two fault endpoints left 35 682 events, with a median of 10 picks and a median residual of 10.385 μs. Restricting the catalog to the interior of the diagonal fault yielded 10 511 events and a lower median residual of 7.596 μs. Representative waveform sections showed systematic P-wave moveout with source-to-sensor distance, providing an independent check on the temporal coherence of the picks used for association and relative location.
    The resulting catalog resolves clear changes in AE activity through the experiment. During the initial loading stage, AE activity was abundant and relatively continuous. Several dense groups occurred at approximately 35–50 min, followed by a pronounced increase at 70–85 min. After about 90 min, the overall event rate decreased and activity became sparser and more quasiperiodic. Although event counts differed among sensors, the main temporal peaks were broadly consistent across the array. Projection of the located events along the fault further reveals the spatial evolution of microfracturing. Between about 3 000 and 8 000 s, the system transitioned from initial stress accumulation to established stick-slip behavior. After approximately 3 000 s, relatively continuous microfracturing developed along the fault. From 4 000 to 5 500 s, events occupied multiple along-fault positions, indicating that activity expanded from localized regions to a larger part of the interface. Between about 6 500 and 7 600 s, event density increased markedly and became concentrated in the middle-to-lower part of the fault near Asperity 2. After approximately 8 000 s, repeated gradual loading and rapid unloading were accompanied by fewer AE events overall, which appeared as discrete near-vertical bands in time-distance space corresponding to individual loading and stress-drop cycles. Activity remained relatively pronounced near Asperity 1 and between the two asperities, whereas the vicinity of Asperity 2 became comparatively quiet, forming an AE gap during the later stick-slip stage.
    Despite large differences in frequency band, sampling rate, source scale, and sensor response, waveform representations learned mainly from natural earthquakes transferred effectively to the laboratory AE data. The DiTing model produced coherent detections and P-wave picks that could be associated across sensors, located with simple travel-time constraints, and used to resolve the temporal and spatial evolution of microfracturing. The resulting catalog links changes in AE activity to independently measured fault heterogeneity and to the transition from distributed loading-stage damage to localized, cycle-related stick-slip behavior, providing a basis for using large laboratory waveform data sets to investigate fault evolution and earthquake process.
  • loading

Catalog

    Turn off MathJax
    Article Contents

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return