Adaptive denoising of mine microseismic signals by integrating spectral kurtosis estimation and fourier synchrosqueezing transform

Mine microseismic signals are characterized by short duration, strong non-stationarity, and pronounced transient behavior. The principal event wave train and broadband random noise often overlap in both the time and frequency domains, making it difficult for conventional fixed-band filtering or single-threshold methods to simultaneously suppress noise and preserve waveform structure. This study proposes an adaptive denoising method integrating spectral-kurtosis–energy band identification, the Fourier synchrosqueezing transform dominant-ridge tracking, and frequency-row-wise Stein’s unbiased risk estimate (SURE) thresholding. The field example is an eight-channel transient record acquired by an underground ARAMIS M/E microseismic network during a controlled destressing blast; it is used as a reproducible field-source waveform rather than as an independent population of spontaneous microseismic events. For channel T3, the dominant frequency remains 63.96 Hz after denoising, background RMS decreases by 29.54%, the event-window correlation coefficient is 0.9937, and the energy-retention ratio is 98.10%. A corrected five-variant ablation experiment is then performed on 100 field-noise-driven resampled signals at input SNRs of −10, −5, 0, 5, and 10 dB. The proposed SK-FSST-SURE variant yields an average output SNR of 10.32 ± 7.40 dB, root-mean-square error of 0.184 ± 0.165, correlation coefficient of 0.895 ± 0.134, and relative energy error of 28.06% ± 56.09%. The SURE-based variants are numerically close, showing that spectral-kurtosis and ridge constraints do not provide a universal gain in global SNR; instead, they modify the trade-off among noise suppression, onset protection, and energy preservation. Independent analytic-reference tests and parameter perturbations further show that performance is stable over a practical range of ridge and time-frequency parameters, while overlap ratio produces a measurable SNR–energy-preservation trade-off. The conclusions are therefore restricted to the tested waveform families, noise conditions, and acquisition environment, and cross-event or cross-mine generalization requires independent field validation.