Trend term and noise removal method for blasting vibration signals based on Fourier decomposition Authors: Yan Zhao, Lujun Yin, Wenzhe Zhang, Jiaxing Tao, He Wang, Ziyi Yun Published: 2026 ABSTRACT AbstractThe trend term and noise in blasting vibration signals severely affect the accuracy of analysis. However, existing adaptive decomposition methods still have limitations in removing these trend and noise components. This study develops a Fourier decomposition–based (FDM) preprocessing approach for blasting vibration signals by optimizing key parameters, including the cross-correlation threshold, the ultra-low-frequency energy threshold, and the decomposition search path. A comparison between the High-to-Low Frequency Searching (HTL-FS) and Low-to-High Frequency Searching (LTH-FS) search paths indicates that LTH-FS achieves superior decomposition performance when applied to blasting vibration signals. The sensitivity of FDM to different ultra-low-frequency energy thresholds is systematically examined, and a criterion for identifying valid components is established. The results demonstrate that FDM can effectively separate low-frequency trend components, high-frequency noise, and meaningful signal information in blasting vibration records. Compared with Empirical Mode Decomposition (EMD), Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), and Variational Mode Decomposition (VMD), the pure signal obtained via FDM exhibits the smoothest time-history curve, the highest signal-to-noise ratio (23.2406), and the lowest root-mean-square error (0.0246). At the same time, this paper further confirms the effectiveness of FDM as a high-precision preprocessing method for blasting signals by analyzing the blasting vibration signals in two actual cases: the Chongli Tunnel blasting project and the Caomao Mountain Tunnel blasting project.Keywords: Blasting signal processing, Fourier decomposition, Trend term, Noise removalSubject terms: Energy science and technology, Engineering Introduction During blasting operations, the resulting vibration signals inevitably induce dynamic responses in the surrounding environment and nearby structures. Therefore, controlling vibration hazards generated by blasting is particularly important. The analysis and processing of measured blasting vibration signals is one of the key approaches for investigating the characteristics of blasting-induced vibration responses. As a short-duration, impact-type non-stationary nonlinear signal, a blasting vibration record may contain a trend term due to factors such as instrument installation and environmental conditions. Meanwhile, because of the drilling-and-blasting construction process, the original signal may be contaminated by high-frequency noise. In subsequent signal processing, the presence of trend terms, noise, and other interference components will directly affect the accuracy of extracting detailed signal features. Therefore, removing trend terms and noise interference from blasting vibration signals has become a research hotspot in this field. Li et al.1 processed measured blasting signals using EEMD and wavelet analysis to remove trend terms and noise. However, the accuracy of wavelet analysis is influenced by subjective choices—such as the basis function type, thresholding criterion, and number of decomposition levels—leading to limited processing accuracy. Based on Fourier decomposition2,3, reconstructed the dominant signal components and suppressed high-frequency content in blasting signals using wavelet packet analysis. Fu et al.4 combined CEEMD with sparse baseline estimation to effectively address baseline drift in blasting signals. However, CEEMD does not fundamentally resolve mode mixing during signal decomposition, and their study did not elaborate on the different manifestations of blasting signals under the LTH-FS and HTL-FS conditions. Fei et al.5 proposed a denoising method that combines EEMD with wavelet thresholding. By decomposing the signal into multiscale modal components and then distinguishing noise characteristics via wavelet thresholding, the denoising accuracy is improved. However, the aforementioned methods have not fundamentally solved the modal aliasing problem, and they lack a systematic exploration of the manifestations of blasting signals under different decomposition paths, making it difficult to achieve precise separation of trend components and noise. Fourier decomposition is a signal processing method based on frequency-domain adaptive filtering, first proposed by Singh et al.6. It was developed to address the limitations of the traditional Fourier transform in handling nonlinear and non-stationary signals, and it has been widely applied in areas such as blasting vibration analysis and mechanical fault diagnosis. Building on this theory, Brincker et al.7 introduced FDM into engineering vibration analysis and proposed a modal-parameter identification method based on the power spectrum matrix. Wang et al.8 systematically compared the performance of three adaptive decomposition methods (FDM, VMD, and EMD) in blasting vibration signal processing and established an evaluation framework for this application. Chen et al.9 combined energy entropy to identify the dominant frequency of blasting signals and used an FDM–energy entropy threshold method to reduce the identification error to within 5%. Zhang et al.10 applied dynamic-bandwidth FDM to blasting signal processing and achieved strong analytical results. Overall, FDM can reduce processing errors and effectively lower the computational cost of blasting signal analysis. Based on the above analysis, this paper proposes an FDM-based preprocessing method for blasting vibration signals, aiming to remove interference information from measured blasting signals that include trend terms and high-frequency noise. Compared with the time–frequency localization trade-off and the mode-mixing problem in traditional blasting signal preprocessing methods, FDM-based preprocessing offers several advantages. First, FDM provides stronger frequency-domain adaptive decomposition capability: it can dynamically adjust the frequency-band range and decompose the signal into multiple narrowband mode components, thereby enabling accurate band separation. By comparison, EMD also has adaptive decomposition capability, but its bandwidth is essentially determined by the spacing between signal extrema and cannot be dynamically adjusted according to signal characteristics. VMD requires predefining an appropriate number of modes K and the bandwidth-constrained parameter a11. Wavelet transforms rely on fixed basis functions and a predefined framework for band partitioning, and thus cannot adapt dynamically to changes in signal characteristics. Second, under intermittent signals or noisy backgrounds, the physical meaning of EMD decomposition results can be ambiguous, and severe mode mixing may occur12. EEMD cannot eliminate noise-induced mixing simply by ensemble averaging with added noise and may lead to substantial computational cost. VMD can effectively suppress mode mixing only when its parameters are appropriately selected13. In contrast, FDM achieves mode separation via SVD, thereby avoiding mode mixing common in traditional methods14. With frequency-band isolation, it prevents different frequency components from being erroneously merged and significantly improves component purity. Moreover, the core operation of FDM is based on the fast Fourier transform (FFT), which reduces the computational complexity and overall computation to some extent, making it more suitable for blasting signal preprocessing. Finally, FDM also exhibits strong robustness in noise suppression. It can adaptively separate noise from effective components using the intrinsic time–frequency structure of the signal, showing greater stability and effectiveness under different noise types. Compared with EMD, FDM can fundamentally avoid noise propagation caused by mode mixing. Compared with VMD, FDM is more tolerant to parameter mismatch when noise characteristics are unknown, because its “decompose first, then determine” workflow is inherently more robust to such mismatches. FDM time–frequency analysis In order to improve the mode-separation accuracy of FDM in environments with strong noise and complex signal composition, singular value decomposition (SVD) (Gao et al.32 ) is introduced as a preprocessing step to decompose the time–frequency distribution matrix into a series of components with clear physical meaning. First, as shown in Eq. (1), a frequency-slice wavelet transform is applied to the original signal to obtain an initial time–frequency matrix, which is then decomposed using SVD: where TFR denotes the initial time–frequency matrix, and U and V represent the left and right singular vectors, respectively. Secondly, based on the decomposition results, mode separation is achieved by truncating SVD. As shown in Eq. (2), the operation can be performed by setting a threshold k (determined according to the abrupt change point of the spectrum line) to retain the first k dominant singular triples (σi, ui, vi) to reconstruct the purified time–frequency matrix15. In SVD preprocessing, the key is to determine the number of singular values k to retain. By calculating the differences di of the singular values σi and plotting the difference curve. The first differential di that meets the following two conditions at the same time is identified as the inflection point. where α is the proportional coefficient, and β is the decay proportion threshold, used to infer whether the difference di decreases significantly. m is the length of the smoothing window, used to determine whether a smooth trend is exhibited over several consecutive points. According to practical engineering experience, the parameter values are α = 0.35, β = 0.25, and m = 3. Finally, SVD is used to obtain the enhanced time–frequency component TFRsvd, which provides a clearer and better-separated time–frequency basis for subsequent frequency-slice extraction in FDM. This enhanced representation is then fed into the FDM procedure, ultimately enabling FDM to more accurately separate the trend term, meaningful signal information, and noise components. As shown in Eq. (6), FDM can adaptively decompose a signal into a finite number of analytical Fourier eigenband functions (AFIBFs) and a residual component (Zhong et al.37): where n(t) is the residual component, and yi(t) ε C∞[a,b], which represents the analytical Fourier eigenband function. AFIBFs mainly include two physical quantities: the amplitude ai(t) and the instantaneous frequency yi(t). The amplitude ai(t) represents the process of blasting energy release and propagation, mainly reflecting the intensity of blasting vibration. The instantaneous frequency yi(t) reflects the time-varying evolution of the spectrum. The corresponding yi(t) and ai(t) can be expressed as follows. where ai(t) represents the vibration energy carried by the i-th AFIBF component at time t, yi(t) describes the local dominant frequency of the i-th AFIBF component at time t. Ci(t) is the signal component, and Ĉi(t) represents the Hilbert transform result of the component Ci(t). Φi(t) is the instantaneous phase of the signal at time t, indicating the position of the signal component Ci(t) at time t. By representing a signal as a superposition of multiple analytical fourier eigenband functions, FDM provides deeper insight into signal analysis. In particular, in the fields of blasting vibration analysis and communication systems, FDM can more accurately extract key information from measured signals. Moreover, its adaptive characteristics can further improve time–frequency analysis accuracy. Meanwhile, FDM can rapidly separate meaningful components, trend terms, and noise from the original signal, thereby effectively reducing mode mixing. The study (Zhong et al.38) indicates that, when searching for AFIBFs using FDM, either the HTL-FS algorithm or the LTH-FS algorithm can be adopted. Among them, LTH-FS performs better in trend extraction and preservation of effective signal components. Specifically, LTH-FS first identifies and separates the ultra-low-frequency components of the signal and then, on this basis, progressively constructs and isolates higher-frequency oscillatory modes. This strategy effectively avoids the drawback of HTL-FS, in which high-frequency errors are propagated and amplified into low-frequency trend terms, thereby yielding a purer and more reliable trend-term estimate. Therefore, compared with HTL-FS, LTH-FS provides a stronger theoretical basis and more reliable results. The study2,3 indicates that, based on energy-distribution characteristics16,17, empirical statistical patterns, and noise-suppression requirements, a modal component is typically regarded as noise or a trend term when its cross-correlation coefficient with the original signal is less than 0.1. From the perspective of energy distribution, blasting signals are characterized by abrupt high energy, and their energy amplitude is often significantly higher than that of background noise. In addition, numerous studies (Cao et al.30) have shown that the normalized amplitudes of the principal components of blasting signals are mostly distributed within [0.1, 1.0]. Based on the practical needs of noise suppression in blasting signal processing, the cross-correlation coefficient threshold is set to 0.1, which can effectively distinguish the effective components (> 0.1) and the noise components (< 0.1) contained in the signal. The following provides a detailed explanation of the rationale for choosing 0.1 as the treatment threshold. First, experimental analyses of a large number of blasting signals from different engineering scenarios (e.g., mine blasting, demolition blasting, and military blasting)18 show that 89.3% of the effective signal components have normalized amplitudes greater than 0.1, whereas only 2.1% of the noise components exceed 0.1. Under a threshold of 0.1, the fidelity of signal reconstruction can exceed 92.7%. Secondly, from the perspective of the physical meaning of the threshold, a threshold of 0.1 often corresponds to a noise suppression level of − 20 dB and significant events other than 3σ (the symbol σ represents the standard deviation) (assuming Gaussian noise). In addition, engineering practice shows that a threshold of 0.1 is often the effective detection limit of most blasting monitoring equipment (GB 672233). Meanwhile, to demonstrate the rationality of the 0.1 threshold in practical blasting signal decomposition, we selected five commonly used threshold values (0.05, 0.10, 0.15, 0.20, and 0.25) for comparative analysis. The suitability of the 0.10 threshold was evaluated using four criteria: signal-to-noise ratio (SNR), root-mean-square error (RMSE), signal reconstruction rate, and computational efficiency. As shown in Table 1, the signal-to-noise ratio (SNR) reaches its peak within the threshold range of [0.1, 0.15]. If the threshold is set too low, noise in the effective signal cannot be completely separated; if it is set too high, the effective signal will be lost. Table 1The trend of SNR variation under different threshold conditions.ThresholdAverage SNR (dB)Rate of change0.0518.2—0.122.7 + 24.7%0.1524.1 + 6.2%0.223.8− 1.2%0.2522.1− 7.1% The trend of SNR variation under different threshold conditions. As shown in Table 2, the RMSE shows the smallest change rate within the range of [0.10, 0.15], which further supports the rationality of selecting a threshold of 0.1. In summary, setting the cross-correlation coefficient threshold between modal components and the original signal to 0.10 can effectively increase the SNR and reduce the RMSE in signal processing for most practical blasting scenarios. Table 2The trend of RMSE variation under different threshold conditions.ThresholdAverage RMSERate of change0.050.142—0.10.087− 38.7%0.150.079− 9.2%0.20.085 + 7.6%0.250.0103 + 21.2% The trend of RMSE variation under different threshold conditions. As shown in Table 3, in different practical applications, a threshold of 0.1 also demonstrates excellent performance in signal reconstruction rate and computational efficiency, balancing the extraction of effective signals from the original signal with the suppression of noise signals. Therefore, using 0.1 as the threshold for the cross-correlation coefficient is reasonable. Table 3Comprehensive performance evaluation.Evaluationindicators0.050.10.150.20.25SNRPoorOptimalOptimalMediumPoorRMSEPoorOptimalOptimalMediumPoorReconstructionrateOptimalOptimalMediumPoorPoorComputationalefficiencyMediumOptimalOptimalOptimalOptimal Comprehensive performance evaluation. Reconstruction rate Computational efficiency If the trend-term amplitude is large, it may produce an artificially high correlation coefficient, thereby affecting the accuracy of the classification. Fortunately, the trend term can also be identified based on the proportion of energy in the ultra-low-frequency band19. Using measured vibration signals as the research object, FDM is employed to obtain the correlation coefficient R(i) between each component and the original signal, as well as the corresponding proportion of energy in the ultra-low-frequency band. These indicators are then used to distinguish trend components from noise components. Finally, the FDM iteration is terminated once the following conditions are satisfied. (1) Artificially setting the residual energy threshold By calculating the ratio of the energy of the current residual signal to the initial energy of the original signal, a percentage threshold ε (Boudraa et al. 2007) is set artificially. When the energy ratio is less than or equal to the set threshold ε, the FDM iteration is stopped. This ensures that after passing through FDM, the residual energy of the original signal reaches a level that can be ignored, and it can be considered that the original signal has fully extracted the main components of the signal through FDM. (2) Set the maximum number of components Before applying FDM to the original signal, it is necessary to preset the maximum number of components allowed in the iterative decomposition. The iteration is terminated when the number of decomposed components reaches this preset maximum. (3) Set the maximum number of iterations By manually specifying a maximum number of iterations, the validity and controllability of the algorithm’s final results can be ensured. During FDM of the original signal, once the preset maximum number of iterations is reached, the iteration is forcibly terminated to prevent the algorithm from entering an infinite loop when other stopping conditions are not satisfied. As shown in Fig. 1, the FDM framework completes the entire workflow from the original signal to final component separation. First, the original signal x(t) is subjected to a frequency-slice wavelet transform to obtain a two-dimensional time–frequency distribution matrix TFR(t,f). Next, SVD is applied as a preprocessing step to denoise and enhance the time–frequency matrix, thereby improving the clarity of mode separation. Subsequently, following the LTH-FS search strategy, each frequency slice is reconstructed into AFIBFs along the selected search path. The resulting AFIBFs components are then screened based on their physical characteristics to distinguish valid signal components from noise and trend terms. Finally, the selected valid AFIBFs are used to reconstruct the signal. FDM flowchart. Numerical simulation experiment Numerical simulation signal creation The key characteristics of numerically simulated signals are directly determined by their construction formulas. Their dominant frequency is determined by the preset sine component frequency, reflecting the multi-frequency component characteristics of blasting vibration signals. The trend term is defined by low-frequency components, simulating non-physical interferences such as sensor bias. The noise component is determined by the added non-Gaussian noise. At the same time, the physical meaning of each component of the simulated signal should correspond one-to-one with the actual characteristics of blasting vibration signals (multi-frequency components, trend interference, random noise, time-varying energy, etc.). By verifying through simulated signals, it is possible to quantify and systematically evaluate the identification accuracy and noise reduction performance of signal processing methods in key areas such as frequency recognition, trend separation, and noise suppression under known true values (), providing a reliable theoretical basis and algorithm verification support for subsequent analysis of measured blasting signals20,21. By superposing sine and cosine functions, waveforms with different frequencies, amplitudes, and phases are combined to generate signals with complex periodic fluctuations, thereby simulating blasting vibration signals. The robustness of the numerical signals is evaluated by adding different forms of artificial noise and trend terms. The numerical simulation vibration signal is constructed as follows: where x is the original pure signal, α(t) is non-Gaussian noise, β(t) is the non-stationary trend term, and z is the original signal. As shown in Fig. 2, the simulation signal containing noise and trend terms has been severely affected , making it difficult to extract useful information. Therefore, it is necessary to preprocess the signal before conducting signal analysis to remove the interference of trend terms and noise. Simulated vibration signal waveform. FDM-based simulation signal trend term and noise removal Based on FDM, the simulation signal is adaptively decomposed into a series of AFIBFs. Among them, we can choose LTH-FS or HTL-FS as the search algorithm. To compare the decomposition effects of the two search algorithms, the center frequency of the first component is selected, Energy concentration of the component and residual signal characteristics as evaluation parameters to evaluate the computational performance of the two search algorithms. As shown in Table 4, for this numerical simulation signal, when the center frequency of the first component is used as the evaluation criterion, the HTL-FS algorithm may yield decomposed effective components that still contain unrecognized noise signals or high-frequency resonance. In addition, the central frequencies of the first components extracted by the HTL-FS and LTH-FS decomposition paths are relatively close, mainly due to two factors: firstly, the trend component in the simulated signal has relatively weak energy, so when the LTH-FS path prioritizes extracting the lowest frequency components, it first separates the 2 Hz component. Secondly, the signal does not contain effective high-frequency components above 10 Hz, so the highest frequency component extracted by the HTL-FS path is locked at 10 Hz. When Energy concentration of the component is used as the evaluation criterion, the components decomposed by HTL-FS exhibit a larger spread; compared with the results obtained by LTH-FS, the energy is less concentrated and the decomposition efficiency is lower. When the signal characteristics of the residual are used as the evaluation criterion, low-frequency fluctuations still remain in the HTL-FS residual, whereas LTH-FS can completely absorb the residual components. Overall, LTH-FS shows better search stability. Therefore, this study selects the LTH-FS algorithm to perform the FDM search for AFIBFs. Table 4Comparative experiments of LTH-FS and HTL-FS decomposition.Evaluation indicatorsLTH-FS resultsHTL-FS resultsComparison conclusionFirst component centerfrequency2 Hz10 HzLTH-FS is more consistentwith physical meaningEnergy concentration ofthe componentRequires 3 components tocover 85% of energyRequires 5 components tocover 80% of the energyLTH-FS is more efficientResidual signal characteristicsThe residual is stationarywhite noiseThe residual contains residuallow-frequency fluctuationsLTH-FS decomposition is morethorough Comparative experiments of LTH-FS and HTL-FS decomposition. First component center frequency LTH-FS is more consistent with physical meaning Energy concentration of the component Requires 3 components to cover 85% of energy Requires 5 components to cover 80% of the energy The residual is stationary white noise The residual contains residual low-frequency fluctuations LTH-FS decomposition is more thorough As shown in Table 5, the decomposition focus of FDM mainly includes the frequency slicing function, the slicing step size, and the setting of termination conditions. The frequency slicing function uses a Gaussian window function with a standard deviation σ of 0.3. If the standard deviation is too small, it will lead to insufficient frequency resolution; if it is too large, it will cause blurred time localization, affecting the precise capture of blasting impact signals. The frequency slicing step size is determined based on the main frequency range of the blasting signal, and a step size of Δf = 2 Hz can achieve a balance between accuracy and efficiency12. If the step size is too small, it will increase the computational burden, while a step size that is too large may miss critical frequencies. Finally, the decomposition termination condition is set so that the residual energy is less than 0.1% of the total energy of the original signal. This threshold ensures that the trend component is fully extracted while avoiding excessive decomposition that produces redundant components. Table 5Key elements of FDM and the basis for its setting.ElementsRecommended values/methodsSet basisExplosion signal adjustmentsuggestionsFrequency slicing functionGaussian window functionAccording to the conventionaltime–frequency resolutionRectangular windows are usedfor signals with strong impulsesto enhance impulse response,while Hanning windows areused for smooth signalsGaussian function standarddeviation σσ = 0.3Time–frequency resolutionbalance pointReduce to 0.2 when high-frequency components areabundant, and increase to 0.4Frequency slicing stepΔfΔf = 2 HzSet based on the mainfrequency range of the signalWhen precise analysis offrequency changes is needed,reduce to 1 Hz; for rapidanalysis, it can be increasedto 5 HzDecomposition terminationconditionResidual energy < 0.1% of thetotal energy of the originalsignalEnsure completeness ofdecompositionTighten to 0.05% when highprecision is required, and relaxto 0.5% for rapid analysisMaximum decompositionlevelNmax = 15Provide sufficientdecomposition limitWhen the signal componentsare complex, it can beincreased to 20; when thecomponents are simple, it canbe reduced to 10 Key elements of FDM and the basis for its setting. According to the conventional time–frequency resolution Rectangular windows are used for signals with strong impulses to enhance impulse response, while Hanning windows are used for smooth signals Gaussian function standard deviation σ Time–frequency resolution balance point Reduce to 0.2 when high- frequency components are abundant, and increase to 0.4 Frequency slicing step Δf Set based on the main frequency range of the signal When precise analysis of frequency changes is needed, reduce to 1 Hz; for rapid analysis, it can be increased to 5 Hz Decomposition termination condition Residual energy < 0.1% of the total energy of the original signal Ensure completeness of decomposition Tighten to 0.05% when high precision is required, and relax to 0.5% for rapid analysis Maximum decomposition level Provide sufficient decomposition limit When the signal components are complex, it can be increased to 20; when the components are simple, it can be reduced to 10 Due to space limitations, as shown in Fig. 3, this article only displays 5 low-frequency components and 5 high-frequency components. FDM decomposition results. In order to accurately identify and separate trend terms and noise in blasting vibration signals, this study uses time domain characteristics and correlation verification as standards to screen the components obtained after FDM. Time-domain characteristics and correlation analysis indicate22 that a waveform initially identified as the signal trend term should vary slowly over time and exhibit a high correlation coefficient with the original signal. In contrast, a component identified as noise should appear as random fluctuations with a zero mean, and its autocorrelation function should decay rapidly. As shown in Fig. 3, compared with other component waveforms, the curve waveform of C1 obviously shows a slowly changing trend. According to Table 6, it can be seen that the correlation coefficient between the C1 component and the original signal is 0.8863. As mentioned before, the C1 component can be preliminarily determined to be the trend component. Table 6Correlation coefficients between AFIBFs and original signals.AFIBFsC2C3C4C5C6C7C8C9 R 0.20150.36710.21260.01440.01650.02150.02660.0221AFIBFsC10C11C12C13C14C15C16C17 R 0.02350.02810.01960.02890.03220.03130.03080.0301AFIBFsC18C19C20C21C22C23C24C25 R 0.02680.02720.02960.03070.02940.02310.01990.0218AFIBFsC26C27C28C29C30C31C32C33 R 0.01870.01960.01710.01690.01310.01280.01090.0118AFIBFsC34C35C36C37C38C39C40 r R 0.00930.00620.00650.00760.00490.00360.00060.0006 Correlation coefficients between AFIBFs and original signals. As shown in Table 6, components C2–C4 exhibit good correlation with the original signal and can therefore be regarded as the effective components of the original signal. In contrast, the correlation coefficients between the remaining components and the original signal are all less than 0.1, and their waveforms display random fluctuations with a zero mean; thus, they can be considered noise components. Moreover, the cross-correlation coefficients of the effective components and the noise components differ markedly from each other, indicating that FDM decomposition effectively separates the useful information from the noise interference contained in the original signal. Subsequently, the C2 ~ C4 components are further reconstructed to obtain a pure signal as shown in Fig. 4. Waveform diagram of pure signal. In addition to identifying trend terms in the original signal through cross-correlation coefficients, the energy proportion of ultra-low-frequency bands can also be used for screening and verification. (1) Signal decomposition Through FDM, the original signal x(t) is decomposed into multiple AFIBFs and a residual r(t), and its expression is as follows23: where Ci(t) represents the i-th AFIBFs component, and r(t) is the residual. The trend term often corresponds to the lower frequency band in the signal, therefore, only the first four low-frequency components and the residual are selected here, as shown in Table 7. Table 7Statistical results of FDM.WeightMeanStandard deviationMaximum amplitudeMinimum amplitudeC10.15390.87722.9674-2.1869C20.12880.90141.2567-1.2540C30.13760.39960.6131-0.6319C40.16530.51100.0698-0.0718Residual0.11580.12230.3522-0.2011 Statistical results of FDM. (2) Calculate the ultra-low-frequency band energy proportion of each component After applying a Fourier transform to each decomposed component, the corresponding spectral content is obtained, and the energy proportion of each component within the ultra-low-frequency band is then calculated (Dragomiretskiy et al.31). What needs to be explained here is that the energy proportion of the ultra-low-frequency band refers to the percentage of the energy in the ultra-low-frequency band of this component to its own total energy. The calculation formula is: where R’ci is the energy proportion of component Ci in the ultra-low-frequency band, Ci low is the energy of component Ci in the 0–1 Hz frequency band (GB/T 1344134), and Ci total is the total energy of component Ci. The calculation results are shown in Table 8. Table 8The energy proportion in the ultra-low frequency band.WeightMain frequency (Hz)Ultra-low-frequencyband energy proportionC1218.8682.03%C2162.150.46%C3148.300.84%C41.250.01%Residual749.801.94% The energy proportion in the ultra-low frequency band. Blasting vibration signals usually have strong impact and transient characteristics, causing the low-frequency components of the signal to usually concentrate in the very low-frequency band. The trend components in the signal typically appear as relatively smooth, long-term varying elements, and their frequency characteristics are often close to 0 Hz. In this case, selecting a 0–1 Hz frequency band can effectively distinguish and extract the trend components in the signal while avoiding the inadvertent removal of low-frequency vibration components. In addition, to avoid mistakenly identifying valid signals with a main frequency below 0.5 Hz as trend components, the range of 0–1 Hz is set as the ultra-low frequency band. Studies have shown that the main energy of blasting vibration signals is concentrated between 5 and 100 Hz, while components with frequencies below 5 Hz are mostly noise or baseline drift. Therefore, setting the 0–1 Hz band as the ultra-low frequency band can effectively separate noise and baseline drift while retaining the true trend components in the signal. In summary, it can be observed that, from a physical definition, the trend component is the slowest-changing part of a signal, with a frequency infinitely close to 0 Hz. At the same time, to avoid mistakenly removing potentially physically meaningful ultra-low-frequency vibrations (such as 2–5 Hz), the 0–1 Hz frequency band is therefore chosen as the very low-frequency band. (3) Set a very low energy proportion threshold to filter trend items After clarifying the ultra-low-frequency band energy proportion of each component, a threshold needs to be set as a criterion to distinguish trend items. Generally, the threshold is 80–90%24. This study sets the threshold for the ultra-low-frequency energy proportion to 80% to identify trend terms. As shown in Table 9, compared with other threshold conditions, when the threshold is set at 80%, both the trend component energy capture rate and the effective signal deletion rate can meet the preprocessing requirements of blasting signals. Table 9Comparison of extraction results at different thresholds.ThresholdTrend item energycapture rateEffective signal falsedeletion rate70%98.5%12.3%80%96.2%1.8%90%89.1%0.2% Comparison of extraction results at different thresholds. As shown in Table 10, compared with other threshold values, setting the threshold to 80% can effectively distinguish the trend term from oscillatory modes, making it easier to extract the signal trend term. Table 10Comparison of advantages and disadvantages of different thresholds.ThresholdAdvantagesShortcomingEvaluateBelow 80%Can capture almost all trenditemsThe specificity is too low, andit is easy to misjudge valid low-frequency oscillations (such asvibrations below 5 Hz) as trend itemsToo aggressive and high risk.May result in loss of effectivesignal80%It has good balance, canreliably capture trend items,and at the same time caneffectively let go of thosecomponents whose mainfrequency is low but isindeed oscillatingWeak oscillatory informationmixed in the trend term maybe eliminated altogetherThe optimal balance pointavoids misjudgments to thegreatest extent while ensuringaccuracyMore than 80%The specificity is extremelyhigh, and it can almostcompletely guarantee thatthe filtered components arepure trend itemsInsufficient sensitivity maymiss some composite trendterms with relatively dispersedenergy, resulting in incompletedetrendingToo conservative and maynot be able to completelyremove all trend components Comparison of advantages and disadvantages of different thresholds. Can capture almost all trend items The specificity is too low, and it is easy to misjudge valid low- frequency oscillations (such as vibrations below 5 Hz) as trend items Too aggressive and high risk. May result in loss of effective signal It has good balance, can reliably capture trend items, and at the same time can effectively let go of those components whose main frequency is low but is indeed oscillating Weak oscillatory information mixed in the trend term may be eliminated altogether The optimal balance point avoids misjudgments to the greatest extent while ensuring accuracy The specificity is extremely high, and it can almost completely guarantee that the filtered components are pure trend items Insufficient sensitivity may miss some composite trend terms with relatively dispersed energy, resulting in incomplete detrending Too conservative and may not be able to completely remove all trend components (4) Reconstruction and removal of trend items. Add all the trend term components to get the complete trend term present in the original signal, and then subtract the trend term from the original signal. Trend term and noise removal effect evaluation In order to compare and analyze the effect of this method on removing trend terms and noise, we use EMD, CEEMDAN, and VMD to process the same signal respectively. As shown in Fig. 5, all of the above methods can effectively remove the trend term from the original signal, but their denoising performance differs markedly. The purified signal reconstructed by EMD exhibits evident residual spiky noise, whereas CEEMDAN leaves only slight spiky noise residue. Compared with these two methods, although the purified signal obtained by VMD is smoother, its peak amplitudes are noticeably lower than those of the original signal, indicating that effective information is removed during trend and noise suppression, which is unfavorable for subsequent signal analysis. Comparison of trend item and noise removal effect. Subsequently, the signal-to-noise ratio (SNR) and root mean square error (RMSE) are further used to quantitatively analyze the decomposed effective signal (Huang et al.35). As shown in Table 11, although EMD, CEEMDAN, and VMD can all separate effective information from noise, the signal-to-noise ratio obtained by FDM is higher than that of these three methods, indicating that FDM can more accurately identify noise in the original signal. In addition, compared with EMD, CEEMDAN, and VMD, FDM yields the smallest root-mean-square error, suggesting that the denoised signal is closest to the original purified signal and that FDM achieves better denoising performance. Table 11Denoising effect comparison.MethodEMDCEEMDANVMDFDMSNR16.954218.259614.740225.1537RMSE0.06800.04780.06240.0273 Denoising effect comparison. In order to make the comparison results more convincing, we show here the steps and key parameters of several other decomposition methods. As can be seen from Table 12, EMD decomposition mainly involves envelope interpolation, the IMF stopping criterion, endpoint treatment, and the setting of termination conditions for the final residual. Typically, the threshold for screening intrinsic mode components is set between 0.2 and 0.3. If the threshold is lower than 0.2, over-screening may occur during component selection; if it is higher than 0.3, mode impurity is more likely to arise. In addition, because blasting signals exhibit abrupt shock-wave characteristics, the threshold must be strictly controlled to avoid peak clipping caused by over-filtering. Meanwhile, given the presence of geological energy-induced vibrations, an appropriate residual threshold should be retained during the parameter-setting process. Table 12Key elements of EMD and the basis for its setting.ElementsRecommended values/methodsSet basisExplosion signal adjustmentsuggestionsEnvelope interpolationmethodcubic spline interpolationBalancing smoothness andcomputational efficiencykeep defaultIMF stopping criteriastandard deviation threshold(SD = 0.2 ~ 0.3)Filter by SD valueSD = 0.25(Controlling shockwave shape distortion)Endpoint handlingmirror symmetry continuationSuppress boundary flyingwing phenomenonIncrease extension lengthby 10%Residual terminationconditionenergy ratio ε ≦ 5%Blasting residuals containlow-frequency geologicalvibrationsRelax to ε = 8% Key elements of EMD and the basis for its setting. Envelope interpolation method Balancing smoothness and computational efficiency standard deviation threshold (SD = 0.2 ~ 0.3) SD = 0.25(Controlling shock wave shape distortion) Suppress boundary flying wing phenomenon Increase extension length by 10% Residual termination condition Blasting residuals contain low-frequency geological vibrations As shown in Table 13, in CEEMDAN, the primary difficulty lies in the noise-injection strategy, which involves selecting the noise amplitude coefficient, the number of ensembles, the noise injection mode, and the corresponding difference-energy threshold. Based on the Monte Carlo principle, when the number of ensembles exceeds 100, the noise-cancellation effect tends to stabilize. In addition, given the impact characteristics of blasting signals, the noise amplitude coefficient typically varies between 0.15 and 0.25. If the coefficient is too small, mode mixing cannot be effectively suppressed during decomposition; if it is too large, the main components of the original signal may be contaminated. Therefore, blasting spectrum analysis should be taken into account to improve the accuracy of the decomposition results. Table 13Key elements of CEEMDAN and the basis for its setting.ElementsRecommended values/methodsSet basisExplosion signal adjustmentsuggestionsNoise amplitude coefficientβ = 0.2*std(s(t))Trade off signal-to-noise ratio:β is too small—insufficientmodal aliasing suppressionβ is too large—signal distortionβ = 0.15—0.25( Dynamicadjustment based on signal-to-noise ratio)Number of sets II = 100–500Law of Large Numbers: Theincrease in the number of setsI will lead to an increase in thenoise cancellation effect andcalculation costI = 200 (Engineering accuracybalance point)Noise injection modeAdaptive staged injectionHigh frequency modes requireless noiseKeep defaultResidual energy thresholdε = 1%After ensemble averaging, theresiduals are more stableε = 0.8% Key elements of CEEMDAN and the basis for its setting. Trade off signal-to-noise ratio: β is too small—insufficient modal aliasing suppression β is too large—signal distortion β = 0.15—0.25( Dynamic adjustment based on signal-to- noise ratio) Law of Large Numbers: The increase in the number of sets I will lead to an increase in the noise cancellation effect and calculation cost I = 200 (Engineering accuracy balance point) High frequency modes require less noise After ensemble averaging, the residuals are more stable As shown in Table 14, compared with EMD and CEEMDAN, the main control parameters of VMD25 include the number of modes, the penalty factor, the convergence tolerance, and the initialization of center frequencies. The number of modes is typically determined in conjunction with the number of dominant frequency bands in the original blasting signal, including components such as shock waves, vibrations of the surrounding rock mass, and equipment noise. The penalty factor is set based on the Gabor uncertainty principle to avoid band overlap while accommodating the primary blasting frequency range of 0–500 Hz. Table 14Key elements of VMD and the basis for its setting.ElementsRecommended values/methodsSet basisExplosion signal adjustmentsuggestionsMode number KNumber of main peaks in thespectrum + 1The main components ofblasting: shock wave + rockmass vibration + equipmentnoise (usually K = 3–5)K = 4 (measured energy ratio > 95%)Penalty factor α α = 2000–5000Bandwidth control: Penaltyfactor α↑, will result in modalbandwidth ↓Explosion main frequencybandwidth = 50 Hzα = 3200 (Strictly control frequencyband overlap)Convergence tolerancetol = 1e-6Ensure modal stability (highaccuracy is required forblasting sudden changes intime domain)tol = 5e-7(Only variable segmentconvergence)Initialize center frequencyε = 1%Avoid local optimum (blastingsignal band separation isobvious)Initialization based on a priorispectrum Key elements of VMD and the basis for its setting. Number of main peaks in the spectrum + 1 The main components of blasting: shock wave + rock mass vibration + equipment noise (usually K = 3–5) Penalty factor α Bandwidth control: Penalty factor α↑, will result in modal bandwidth ↓ Explosion main frequency bandwidth = 50 Hz α = 3200 (Strictly control frequency band overlap) Ensure modal stability (high accuracy is required for blasting sudden changes in time domain) tol = 5e-7(Only variable segment convergence) Avoid local optimum (blasting signal band separation is obvious) Initialization based on a priori spectrum As a comparison of the effects of FDM, Table 15 shows the specific parameter settings of EMD, CEEMDAN, and VMD in this experiment. Table 15Specific parameter settings for each decomposition method.MethodKey parametersParameter valueIllustrateEMDStopping criterionSD = 0.25Based on controlled shockwave morphological distortionEndpoint handlingMirror extensionResidual termination conditionε = 8%Consider low-frequencygeological vibrationsCEEMDANNoise amplitude0.2Adjust according to signal-to-noise ratioNumber of collections200According to engineeringprecision balance pointresidual energy thresholdε = 0.8%VMDMode number4penalty factor3200Impose strict bandwidthconstraintsConvergence tolerance5e-7Ensure that the algorithmfully converges Specific parameter settings for each decomposition method. Based on controlled shock wave morphological distortion Consider low-frequency geological vibrations Adjust according to signal-to- noise ratio According to engineering precision balance point Impose strict bandwidth constraints Ensure that the algorithm fully converges By comparing the denoising performance indicators of the four methods, it can be found that the purified signal produced by VMD is smoother than that obtained by CEEMDAN, but it yields a lower signal-to-noise ratio and a higher root-mean-square difference. This indicates that useful information is not effectively retained after VMD processing, which may negatively affect subsequent signal analysis. In contrast, the purified signal obtained using FDM exhibits the smoothest time-history curve, the highest signal-to-noise ratio, and the smallest root-mean-square difference, indicating that the useful information in the original signal is effectively preserved and the mixed noise is effectively removed, thereby facilitating accurate subsequent analysis. Processing of measured blasting vibration signals The blasting vibration signals in this study come from the Chongli Tunnel blasting project and the Caomao Mountain Tunnel blasting project of the Beijing–Zhangjiakou High-Speed Railway. The Chongli Tunnel (Case I) is related to the 2022 Beijing Winter Olympics projects, located in Xiwanzi Town, Chongli District, Zhangjiakou City, Hebei Province. It is a single-tube tunnel, situated mainly in low to medium mountainous areas, with long stretches where the mountain terrain is steep. The tunnel exit and sections with large mileage have gentler slopes, and ravines are well developed. Overall, the terrain is higher in the middle and lower on both sides. Figure 6 shows the construction site of the No. 3 inclined shaft of the Chongli Tunnel. Chong-li tunnel 3# inclined shaft work area. Caomao Mountain Tunnel (Case II) is located along terrain with significant undulations, situated on the main ridge of Caomao Mountain. The strata include Upper Quaternary New Loess, silty clay, and fully weathered tuff. Some of the rock and soil masses have moderate expansibility, posing considerable safety risks for tunnel blasting construction. Figure 7 shows the construction site at the entrance work area of Caomao Mountain Tunnel. Caomao mountain tunnel entrance area. As shown in Fig. 8(a), the blasting vibration signal in Case I contains obvious trend components and high-frequency noise, making it necessary to perform preprocessing to address the interference components. From Fig. 8(b), it can be seen that the blasting vibration signal corresponding to Case II is affected by the trend components, exhibiting a significant zero-drift phenomenon, although the characteristics are not very obvious. Blasting vibration signal waveform. Trend term and noise removal based on FDM The measured blasting vibration signal was adaptively decomposed using FDM, with LTH-FS selected as the search algorithm. The 43 AFIBFs and one residual component r were obtained in Case I, the 14 AFIBFs and one residual component r were obtained in Case II. For brevity, only the five low-frequency modal components and five high-frequency modal components are presented here, as shown in Fig. 9. Following the procedure described above, the cross-correlation coefficients between each AFIBF and the original signal were calculated using FDM, as reported in Tables 16 and 17. FDM results. Table 16AFIBFs parameters of Case I.AFIBFsC1C2C3C4C5C6C7C8C9 R 0.93360.19860.46850.45730.60240.29560.14000.21650.1144Main frequency/Hz2.06.213.417.030.644.869.279.487.6AFIBFsC10C11C12C13C14C15C16C17C18 R 0.01310.00730.00690.00700.00690.00760.00810.01100.0125Main frequency/Hz126.0150.4179.4190.8223.6256.0288.2316.6351.0AFIBFsC19C20C21C22C23C24C25C26C27 R 0.01820.01520.01070.00980.00890.00870.00970.00890.0084Main frequency/Hz423.6431.2537.2586.0652.0683.6733.4783.8842.4AFIBFsC28C29C30C31C32C33C34C35C36 R 0.00680.00570.00650.00490.00380.00470.00470.00290.0036Main frequency/Hz852.0886.4908.8935.4945.6947.0964.4971.0978.8AFIBFsC37C38C39C40C41C42C43 r R 0.00270.00230.00210.00070.00140.00080.00050.0005Main frequency/Hz988.2990.4996.6997.0998.4999.0999.80 AFIBFs parameters of Case I. Main frequency/ Hz Table 17AFIBFs parameters of Case II.AFIBFsC1C2C3C4C5C6C7C8C9 R 0.143 70.102 80.099 80.076 90.049 30.689 20.768 30.553 70.253 5Main frequency/Hz1.32.13.23.94.757.396.8156.2275.6AFIBFsC10C11C12C13C14 r R 0.150 90.025.90.001.30.002.70.000 10.000 1Main frequency/Hz316.1576.8876.2920.4975.60 AFIBFs parameters of Case II. Main frequency/ Hz Main frequency/ Hz The blasting vibration meter used in this study was a TC-4850N, with a minimum measurable frequency of 5 Hz. In Case I, although the dominant frequency of component C1 was 2 Hz—well below the instrument’s minimum measurable frequency—this alone is not sufficient to determine whether it is a trend term. This is because the dominant frequency of blasting vibrations is influenced by factors such as the charge quantity, charging method, distance from the blast center, and the properties of the rock–soil medium26,27. As a result, the dominant frequency varies dynamically and may indeed fall below 5 Hz. In Case I, although C1 has a cross-correlation coefficient of 0.9336 with the original signal, indicating strong correlation, it is still necessary to determine whether a component is a trend term based on its physical nature. As shown in Fig. 7, the time-domain waveform of C1 is non-oscillatory and does not exhibit clear periodic fluctuations, which is the key physical basis for identifying a trend term. The high cross-correlation coefficient (0.9336) in this case instead reflects that the trend term has a large amplitude and varies slowly, representing substantial interference rather than evidence that it is an effective vibration component. In summary, C1 can be preliminarily regarded as a trend term that should be removed. In Case II, the cross-correlation coefficient of C1 is 0.1437, which does not indicate a good correlation. From this, we can infer that C1 contains interfering components. In summary, it can be preliminarily considered that C1 is a trend term that needs to be removed. In addition, the proportion of energy in the ultra-low-frequency band can also be used to identify trend terms, so only five low-frequency components are listed here. Since a trend term is physically defined as the slowest-varying component with the longest period in a signal, its energy in the frequency domain should be highly concentrated in the ultra-low-frequency band. Numerous studies28 have shown that, under an energy-threshold–based criterion for automatic trend-term identification, if the proportion of a component’s energy within the ultra-low-frequency band exceeds a specified threshold, the component’s energy can be considered highly concentrated in the band where the trend term resides, and the component can therefore be identified as a trend-term component. As mentioned earlier, we calculated the energy proportion in the ultra-low-frequency band for each component. The results of case I are listed in Table 18. It can be observed that, with the ultra-low-frequency energy threshold set to 80%, only C1 reaches 80.93%, exceeding the preset threshold. This indicates that 80.93% of C1’s energy is concentrated in the ultra-low-frequency band, i.e., its energy is highly concentrated in the ultra-low-frequency range. Therefore, C1 can be identified not as an analytic Fourier eigenband function with pronounced oscillatory characteristics, but as a distinct trend component that reflects baseline drift or a long-term trend in the signal. Accordingly, based on this criterion, C1 is removed from the original signal as a trend component, whereas C2, C3, C4, and C5 are retained as effective components. The results of case II are listed in Table 19. It can be observed that under the condition where the energy proportion threshold in the ultra-low frequency band is set to 80%, only C1 has an energy proportion of 83.76% in the ultra-low frequency band, exceeding the preset threshold. This indicates that 83.76% of C1’s energy is concentrated in the ultra-low frequency band, meaning its energy is highly concentrated in the ultra-low frequency range. Therefore, it can be demonstrated that C1 does not belong to a Fourier intrinsic band function with a clear oscillatory characteristic, but is instead a distinct trend component representing signal baseline drift or long-term trends. Based on this criterion, C1 should be removed from the original signal as a trend component. Table 18Statistical results of the energy proportion of case I in the ultra-low frequency band corresponding to C1 ~ C5.WeightMain frequency(Hz)Ultra-low-frequencyband energy ratioC12.080.93%C26.20.31%C313.40.58%C417.00.66%C530.60.78% Statistical results of the energy proportion of case I in the ultra-low frequency band corresponding to C1 ~ C5. Table 19Statistical results of the energy proportion of case II in the ultra-low frequency band corresponding to C1 ~ C5.WeightMain frequency(Hz)Ultra-low-frequencyband energy ratioC11.383.76%C22.10.18%C33.20.35%C43.90.67%C54.70.83% Statistical results of the energy proportion of case II in the ultra-low frequency band corresponding to C1 ~ C5. In summary, based on the frequency information and the cross-correlation coefficients, it can be confirmed that in Case I, C1 can be identified as the trend component, while C10 ~ C43 are classified as high-frequency noise components. Therefore, these components should be removed from the original signal, and C2 ~ C9 should be selected as the effective information to reconstruct the purified signal. The final signal preprocessing result is shown in Fig. 10(a). In Case II, C1 can be identified as the trend component, and C10 ~ C14 as high-frequency noise components. Therefore, they should be removed from the original signal, and C6 ~ C9 should be selected as effective information to reconstruct the clean signal. The final signal preprocessing result is shown in Fig. 10(b). The preprocessed clean signal. Evaluation of trend items and noise removal effectiveness To compare the analytical effects of different preprocessing methods, we performed signal processing using EMD, CEEMDAN, and VMD on the same vibration signal20,21. The specific processing results are shown in Fig. 11. Comparison of denoising effects of different methods. It can be observed that, although all three methods can effectively remove the trend term from the original signal, the purified signal obtained by EMD contains more residual spiky noise, and the peak particle velocity is significantly reduced. While CEEMD and VMD achieve favorable denoising performance, they also reduce the peak particle velocity to some extent. Therefore, a comparative analysis of the three methods is conducted from three aspects: signal-to-noise ratio, root-mean-square error, and peak particle velocity.. As shown above, the Case I of peak particle velocity of the original signal is 0.767 3 cm s–1, the Case II of peak particle velocity of the original signal is 0.789 4 cm s–1. However, as reported in Tables 20 and 21, although the signal processed by VMD exhibits a smooth waveform, it yields the lowest signal-to-noise ratio and a pronounced reduction in peak particle velocity. This indicates that, when the original signal contains noise, VMD may remove key signal information, which is unfavorable for subsequent accurate analysis. Compared with EMD and VMD, CEEMDAN significantly improves both the signal-to-noise ratio (SNR) and peak particle velocity (PVO) of the processed signal. In contrast, FDM achieves the highest SNR, the lowest RMSE, and a PVO closest to the true value. In summary, FDM effectively removes trend terms and noise interference while preserving the most informative components of the original signal, thereby facilitating more accurate subsequent signal analysis. Table 20Comparison of signal preprocessing effects of Case I.MethodSNRRMSEPeak vibrationspeed (cm/s)EMD6.33790.02290.5364CEEMDAN19.22460.00520.7409VMD6.11000.02350.5735FDM28.37360.00180.7658 Comparison of signal preprocessing effects of Case I. Table 21Comparison of signal preprocessing effects of Case II.MethodSNRRMSEPeak vibrationspeed (cm/s)EMD5.98460.02360.5731CEEMDAN21.48750.00630.7364VMD6.20620.02980.5967FDM29.21090.00190.7782 Comparison of signal preprocessing effects of Case II. To further verify the effect of FDM on the preprocessing of blasting vibration signals, a spectral analysis of the blasting signals before and after preprocessing was conducted. As shown in Fig. 12(a), the original blasting vibration signal exhibits a markedly amplified ultra-low-frequency component near 2 Hz in the spectrum. This component primarily arises from trends such as sensor baseline drift and gradual system response, rather than the dominant frequency of structural vibration induced by blasting. Moreover, this trend increases the energy level in the ultra-low-frequency band. If the original spectrum is interpreted directly, the 2 Hz component could easily be mistaken for the dominant frequency of the blasting vibration, which contradicts the generally accepted engineering frequency range of 5–100 Hz. As shown in Fig. 12(b), after FDM preprocessing, the dominant-frequency energy of the purified signal is concentrated within the 5–100 Hz range. The peak location and bandwidth are consistent with the primary frequency range of tunnel blasting vibrations reported in the literature23,22), while the energy in the ultra-low-frequency band is significantly reduced. Spectrum diagrams of blasting vibration signals before and after FDM processing. In Case II, as shown in Fig. 12(c), the original blasting vibration signal exhibits a significant ultra-high low-frequency amplitude around 0–5 Hz in the frequency spectrum. This component mainly originates from trend terms such as zero drift in the measured blasting vibration signal, rather than the structural vibration fundamental frequency under blasting force. As shown in Fig. 12(d), after FDM preprocessing, the main frequency energy of the clean signal is concentrated around 62 Hz, which corresponds to the main frequency range of the tunnel blasting signal. These results indicate that FDM preprocessing can effectively remove the ultra-low-frequency trend term and, while preserving the effective blasting-vibration components, corrects the misleading impression that the dominant frequency is pulled down by the trend term. It is worth noting that, although the nominal lower limit of the effective frequency band of the vibration acquisition instrument used in this study is 5 Hz, components in the 0–5 Hz range can still be observed in the spectrum shown in Fig. 12. This phenomenon can mainly be attributed to two factors. First, inherent system noise and residual responses cause sub-5 Hz components to appear in the spectrum. Specifically, the electronic circuitry of the acquisition system introduces ultra-low-frequency interference, and although the sensor response below the 5 Hz cutoff decays sharply, it does not vanish completely. These residual signals, together with system noise, constitute the ultra-low-frequency content in the spectrum. Second, adaptive signal decomposition algorithms such as FDM inevitably exhibit endpoint effects when processing strong transient impulse signals, which can reconstruct part of the energy into the ultra-low-frequency region, forming components that represent the overall trend of the signal; this can be regarded as an algorithm-induced reconstruction effect. The spectra of the trend terms extracted by the algorithm may therefore extend over the original frequency band. Therefore, the 0–5 Hz spectral components in Fig. 12 mainly reflect the noise floor of the system, the residual response of the sensor, and the inherent characteristics of the signal processing process, which are shown in the spectrum of the blasting vibration signal. Conclusion Based on the characteristics of blasting vibration signals, this paper introduces a signal preprocessing method based on FDM. The applicability of the method is verified through numerical simulation, and it is applied in practice based on measured tunnel blasting vibration signals. The following conclusions are drawn: After FDM decomposition and selection of dominant components, the reconstructed signal eliminates invalid information while effectively preserving the characteristics of the original signal, thereby maintaining an unchanged peak particle velocity and consistent frequency characteristics. After FDM decomposition and selection of dominant components, the reconstructed signal eliminates invalid information while effectively preserving the characteristics of the original signal, thereby maintaining an unchanged peak particle velocity and consistent frequency characteristics. During blasting signal preprocessing, setting the cross-correlation coefficient threshold between modal components and the original signal to 0.10 can effectively increase the SNR and reduce the RMSE in practical engineering applications. When the threshold for the ultra-low-frequency energy proportion is set to 80%, the trend term can be effectively identified, thereby minimizing misclassification while ensuring accuracy. During blasting signal preprocessing, setting the cross-correlation coefficient threshold between modal components and the original signal to 0.10 can effectively increase the SNR and reduce the RMSE in practical engineering applications. When the threshold for the ultra-low-frequency energy proportion is set to 80%, the trend term can be effectively identified, thereby minimizing misclassification while ensuring accuracy. Compared to signal processing methods such as EMD, CEEMDAN, and VMD, FDM achieves better preprocessing results for blasting signals. It has the highest signal-to-noise ratio (28.3736), the smallest root mean square error (0.0018), and the peak particle velocity is closest to the actual value. Compared to signal processing methods such as EMD, CEEMDAN, and VMD, FDM achieves better preprocessing results for blasting signals. It has the highest signal-to-noise ratio (28.3736), the smallest root mean square error (0.0018), and the peak particle velocity is closest to the actual value. Prospect Although FDM demonstrates significant advantages in blasting vibration signal processing and analysis, its effectiveness and accuracy still depend on specific application conditions and have inherent limitations. The FDM decomposition method is essentially designed to handle non-stationary signals. Because blasting vibration signals have rapidly changed transient characteristics, they are highly compatible with the core processing of FDM. Conversely, when dealing with stable vibration signals, compared to the FDM method, the Fourier transform is more efficient in processing. At the same time, in order to accurately capture the entire process of blasting vibration and ensure frequency domain resolution, the FDM method requires a sufficient sampling rate and recording length, that is, it must include the complete stages of vibration from initiation, main vibration, to decay, and its computational complexity is relatively high. The threshold setting in the current FDM method, such as an 80% energy ratio, has undergone sensitivity validation but still has a certain empirical aspect. Future research could explore combining FDM with machine learning, such as support vector machines, random forests, or deep learning, such as convolutional neural networks and long short-term memory networks, using large-sample training to achieve intelligent optimization of thresholds and decomposition paths, further enhancing the method’s adaptability and generalization performance. Compared with other signal decomposition methods, although FDM has a certain robustness to noise, its mode separation effect will deteriorate in low signal-to-noise ratio environments. When the background noise intensity is comparable to weak blast signals, the AFIBFs obtained by its decomposition may be severely contaminated by noise, making feature extraction difficult. In such extreme cases, it is necessary to use it in combination with pre-denoising techniques such as SVD. In practical applications, FDM is still subject to many limitations. First, in the process of processing ultra-long signals, compared with traditional filtering methods, FDM has a higher computational cost, which is not conducive to real-time processing. At the same time, when faced with incomplete data at both ends of the signal, FDM may produce end effects, causing distortion in the components at the beginning and end of the signal. When dealing with transient impacts of short-duration high energy generated by precise blasting, FDM may disperse their energy into multiple adjacent AFIBFs instead of concentrating it in a single component, resulting in modal aliasing. Abbreviations Time The i-th singular value The i-th left singular vector The i-th right singular vector Non-Gaussian noise Non-stationary trend term Original signal The i-th AFIBFs component Residual The proportion of energy of component Ci in the ultra-low frequency band The energy of component Ci in the ultra-low frequency band Total energy of component Ci Author contributions Zhao Yan was responsible for writing the paper. Yin Lujun was responsible for the typesetting and review of the paper. Zhang Wenzhe was responsible for the theoretical derivation of the paper content. Tao Jiaxing was responsible for the on-site monitoring part of the paper content. Wang He was responsible for the data verification part of the paper content. Yun Ziyi was responsible for the data processing part of the paper content. Funding The work described in this paper was supported by the National Natural Science Foundation of China (number: 52304108), Central guide local science and technology development fund project (regional science and technology innovation system project) (number: 254Z3801G), Basic research projects of Shijiazhuang universities in Hebei province (number: 241790797A), Zhangjiakou Basic Research and Talent Training Program Project (number: 2511012A), Graduate Innovation Fund Project of Hebei University of Architecture (number: XY2026024; number: XY2026075), Sports Science and Technology Project of Hebei Provincial Sports Bureau(number: 2026QT17), Hebei Province Institutions Of Higher Education Scientific Research Project (number: QN2026625). Data availability If necessary, the raw data in the manuscript can be obtained by contacting the first author or corresponding author. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement If necessary, the raw data in the manuscript can be obtained by contacting the first author or corresponding author.