A system for implementing InSAR deformation time series analysis and fault creep event detection

Abstract

The application discloses a system for realizing InSAR deformation time series analysis and fault creep event detection, and relates to the technical field of geophysical monitoring.The system comprises a comprehensive monitoring platform, and the comprehensive monitoring platform is in communication connection with the following modules: a time series noise filtering module, which is used for extracting a creep deformation signal related to a fault sliding mechanism from original InSAR deformation data of a target detection area by adopting a space-time adaptive filtering algorithm in combination with wavelet transform and independent component analysis, filtering out the interference of seasonal hydrology, thermal expansion and non-tectonic noise, and obtaining a filtered time series component sequence.The application can effectively separate and inhibit seasonal hydrology, thermal elastic deformation and other non-tectonic noise by adopting the combination of wavelet transform and independent component analysis, retain time series components highly related to fault creep activities, and significantly improve the extraction precision of tectonic signals in InSAR deformation data.

Description

Technical Field

[0001] This invention relates to the field of geophysical monitoring technology, specifically to a system for realizing InSAR deformation time series analysis and fault creep event detection. Background Technology

[0002] Synthetic Aperture Radar Interferometry (InSAR) is a high-precision remote sensing technology that uses radar signals to acquire information on surface deformation. It is widely used in ground deformation monitoring and geological disaster research. With the acceleration of urbanization and the frequent occurrence of natural disasters, the demand for real-time monitoring of surface deformation is increasing. InSAR deformation time series analysis can effectively extract the cumulative displacement and dynamic change trend of the surface within a specific time period, providing a detailed description of the surface deformation characteristics and providing scientific basis for fields such as geology, engineering monitoring, and environmental management. In earthquake and fault activity research, the detection of fault creep is particularly important. InSAR technology can identify and analyze small-amplitude, continuous deformation, which is of great significance for early warning, disaster assessment, and risk management.

[0003] For example, Chinese Patent Publication No. CN116579068A discloses a method and apparatus for determining the creep deformation of an active fault. By using a pre-constructed target stiffness formula for the target active fault, the stiffness of the target active fault in the target time period is determined, so that the determined stiffness of the target active fault in the target time period is a value that varies with time, thereby making the creep deformation of the active fault determined based on the stiffness that varies with time more accurate.

[0004] In existing technologies, when faced with long-term InSAR deformation observation data, it is difficult to effectively separate the true deformation signal caused by fault creep activity because it is often overwhelmed by non-tectonic noise such as strong surface seasonal changes, hydrological loads, and thermoelastic deformation. Secondly, after successfully extracting the creep-dominated deformation time series signal, existing technologies lack the ability to finely characterize the dynamic process of creep activity. They cannot automatically and accurately detect the start, duration, and termination time of a single creep event from continuous time series, nor can they quantitatively assess its slip amount and sliding mode. This restricts the in-depth understanding of the mechanism of fault creep behavior and the risk assessment of related disasters. To address these issues, a system for InSAR deformation time series analysis and fault creep event detection is proposed. Summary of the Invention

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a system for realizing InSAR deformation time series analysis and fault creep event detection, comprising an integrated monitoring platform, wherein the integrated monitoring platform is communicatively connected to the following modules:

[0006] The temporal noise filtering module is used to extract creep deformation signals related to fault slip mechanism from the original InSAR deformation data of the target detection area by employing a spatiotemporal adaptive filtering algorithm, combined with wavelet transform and independent component analysis (ICA). At the same time, it filters out the interference of non-tectonic noise such as seasonal hydrology and thermal expansion, obtains the filtered temporal component sequence, and improves the fidelity of the tectonic signal of the data.

[0007] The creep signal enhancement module is used to further process the filtered time sequence components to obtain clean time sequence components by using time sequence reconstruction and signal enhancement techniques, based on filtering non-structural noise. This improves the continuity and signal-to-noise ratio of the creep-dominated deformation time sequence, providing high-fidelity and high-resolution time sequence input for subsequent event detection.

[0008] The modal decomposition analysis module is used to decompose the temporal components in the clean temporal component sequence into multiple intrinsic mode functions using the variational mode decomposition (VMD) algorithm, extract creep dynamic features at different time scales, form a creep dynamic feature sequence, extract multi-scale creep dynamic features, and provide structured, multi-modal temporal input for event state recognition.

[0009] The event state recognition module establishes a creep event state transition model based on the Hidden Markov Model (HMM) algorithm. It takes the extracted creep dynamic features as input, automatically identifies the start, duration and termination states of creep events through probabilistic inference, and outputs the event time sequence. This enables intelligent segmentation and state labeling of the event time sequence, and improves the intelligence and accuracy of event boundary detection.

[0010] The slip quantization and classification module is used to accurately calculate the slip amount of each creep event based on event state recognition, combined with temporal deformation variables and event duration. It also establishes a creep event classification model based on slip rate, duration and deformation curve characteristics, automatically identifying and classifying steady-state creep and intermittent accelerated creep, improving the depth of understanding of fault slip behavior, and realizing accurate calculation and intelligent classification of creep event slip amount.

[0011] Preferably, the temporal noise filtering module includes a multi-scale temporal decomposition unit and an independent component separation unit;

[0012] The multi-scale temporal decomposition unit is used to receive the original InSAR deformation data of the target detection area and use wavelet transform to perform multi-scale decomposition on the original InSAR deformation data, separate different frequency components, initially enhance the identification ability of the construction-related signal, realize the initial separation of the original data in multiple frequency bands, and significantly enhance the spatiotemporal resolution of the fault-related signal.

[0013] The independent component separation unit is used to further separate independent signal sources in InSAR deformation data based on the multi-scale decomposition of wavelet transform, combined with spatiotemporal adaptive filtering algorithm and independent component analysis (ICA), to extract time-series components highly correlated with fault creep activity, effectively filter out non-tectonic noise interference to deformation analysis, integrate the extracted time-series components to obtain a filtered time-series component sequence, significantly suppress non-tectonic noise, extract high-fidelity creep time-series components, and improve the signal-to-noise ratio of tectonic signals.

[0014] Preferably, the multi-scale temporal decomposition unit performs the following steps:

[0015] The raw InSAR deformation data of the target detection area is received, and the raw InSAR deformation data is decomposed into multiple scales using wavelet transform to obtain a set of time-series components containing different frequency components, thereby achieving effective separation of the signal in the time-frequency domain and preserving multi-scale structural activity information.

[0016] Frequency-spatial domain analysis is performed on each time series component in the time series component set to screen out high-frequency and low-frequency time series components related to fault slip mechanism, preliminarily separate non-tectonic noise interference, enhance the identifiability of tectonic related signals, and suppress seasonal and other noise interference.

[0017] The selected high-frequency and low-frequency time-series components are normalized to output the preliminary decomposed time-series component sequence, providing input for subsequent independent component analysis, unifying data dimensions, and improving the stability and convergence efficiency of subsequent processing.

[0018] Preferably, the independent component separation unit performs the following steps:

[0019] The temporal component sequence is smoothed in the spatial domain and filtered in the temporal domain based on the spatiotemporal adaptive filtering algorithm, which suppresses local noise interference, significantly improves the continuity of the signal in space and time, and provides a clean temporal basis for subsequent analysis.

[0020] Independent component analysis was performed on the filtered time series components to extract multiple statistically independent signal sources. Independent time series components highly correlated with fault creep activity were identified, effectively separating tectonic-related signals, eliminating non-tectonic interferences such as seasonality and hydrology, and improving the accuracy of signal source identification.

[0021] The extracted independent time-series components are integrated and reconstructed to form a filtered time-series component sequence after filtering out non-structural noise, generating a high-fidelity creep time-series signal that significantly enhances the spatiotemporal consistency of fault activity characteristics.

[0022] Preferably, the creep signal enhancement module includes a timing reconstruction optimization unit and a signal-to-noise ratio enhancement unit;

[0023] The time-series reconstruction optimization unit is used to selectively reconstruct the time-series components extracted from the filtered time-series component sequence, remove irrelevant components, highlight the time-series features related to the fault slip mechanism, continuously enhance the intensity of the creep signal in the time-series dimension, obtain the time-series component reconstruction sequence, enhance the time-series continuity and identifiability of the creep signal, and improve the accuracy of event detection.

[0024] The signal-to-noise ratio enhancement unit is used to further suppress residual noise in the time-series component reconstruction sequence by employing adaptive smoothing and trend correction algorithms, improve the temporal consistency and resolvability of the creep signal, and output a clean time-series component sequence to provide high-quality input for subsequent mode decomposition and event recognition.

[0025] Preferably, the timing reconstruction optimization unit performs the following steps:

[0026] The filtered time series component sequence is received, and the time series component is segmented and analyzed based on a sliding time window. The characteristic components related to the fault slip mechanism in each segment are identified. The local time features of fault activity are accurately extracted through local analysis, thereby improving the signal's identifiability within a short time window.

[0027] By removing time-series components unrelated to fault slip, and performing weighted superposition and time-series alignment on the remaining feature components, a reconstructed time-series component sequence highlighting the intensity of creep signals is generated. This effectively suppresses non-tectonic interference, significantly enhances the continuity and spatiotemporal consistency of fault-related deformation signals, and performs trend fitting and baseline correction on the reconstructed time-series component sequence. This enhances the continuity and structure of the time-series signal in the time dimension, eliminates long-term trend drift and system offset, and improves the structural stability of the time-series signal and the reliability of event detection.

[0028] Preferably, the signal-to-noise ratio enhancement unit performs the following steps:

[0029] An adaptive smoothing algorithm is used to suppress local noise in the reconstructed sequence of the time components, maintain the temporal abrupt change characteristics of the creep signal, significantly suppress random noise, improve the local stationarity of the signal, and preserve the temporal abrupt change details of the fault creep.

[0030] The trend correction algorithm is used to fit the overall trend of the smoothed sequence, remove long-term drift and seasonal residual components, effectively eliminate background drift and seasonal residue, highlight the deformation trend dominated by the structure, improve the temporal consistency, and output a clean time series component sequence after noise suppression and trend correction. This provides high-fidelity and high-resolution time series input for subsequent mode decomposition analysis. The output signal has a high signal-to-noise ratio and high time series fidelity, providing stable and reliable input data for mode decomposition.

[0031] Preferably, the modal decomposition analysis module performs the following steps:

[0032] The clean time-series component sequence is received and decomposed into multiple intrinsic mode functions using a variational mode decomposition algorithm. This enables adaptive extraction of structural signal components at different time scales, enhancing the analytical capability of creep dynamic features.

[0033] Based on the center frequency and energy distribution of each intrinsic mode function, modal components related to the creep dynamic process are screened out, effectively eliminating noise and seasonal interference, retaining the modal subset that is significantly related to fault activity, and improving the feature targeting.

[0034] The selected modal components are recombined to form a multi-timescale creep dynamic feature sequence, constructing multi-scale temporal features with clear physical meaning, providing structured and high-resolution input for subsequent event recognition.

[0035] Preferably, the event state recognition module performs the following steps:

[0036] A state transition model for creep events is constructed based on a hidden Markov model. The model defines three states of creep events: start, duration, and termination. This enables physical modeling of event states, provides a probabilistic basis for automatic segmentation, and improves the structurality of detection.

[0037] The creep dynamic feature sequence is input into the constructed creep event state transition model. The event state at each time point is identified through probability inference and Viterbi decoding algorithm, and the most likely state at each time point is automatically output, so as to realize the temporal intelligent identification and boundary extraction of creep events.

[0038] Output an event time sequence with status labels, marking the start time, duration and end time of each creep event, and generating a structured event table that can be directly used for slip calculation, improving the efficiency and consistency of subsequent analysis.

[0039] Preferably, the slip quantization classification module performs the following steps:

[0040] Based on the event time sequence, the temporal deformation and event duration corresponding to each creep event are extracted to accurately obtain the deformation data and duration within the event segment, providing basic data support for slip calculation and event classification, and improving the accuracy of subsequent analysis;

[0041] By combining the fault slip mechanism and deformation curve characteristics, a slip rate calculation model is established to quantify the slip amount and average slip rate of each creep event, thereby achieving physical quantification of slip amount and rate and enhancing the ability to objectively describe the fault slip process and identify its dynamic characteristics.

[0042] Based on the average slip rate, duration, and deformation curve morphology, a creep event classification model is constructed to automatically identify and classify steady-state creep and intermittent accelerated creep, and to automatically complete the event type discrimination, thereby improving the understanding of fault slip behavior patterns and the effectiveness of geological disaster risk classification management.

[0043] This invention provides a system for realizing InSAR deformation time series analysis and fault creep event detection. It has the following beneficial effects:

[0044] (i) The system for realizing InSAR deformation time series analysis and fault creep event detection, by combining wavelet transform and independent component analysis, can effectively separate and suppress non-tectonic noise such as seasonal hydrology and thermoelastic deformation, retain the time series components that are highly correlated with fault creep activity, and significantly improve the extraction accuracy of tectonic signals in InSAR deformation data.

[0045] (II) This system for realizing InSAR deformation time series analysis and fault creep event detection can effectively suppress residual noise and enhance the spatiotemporal continuity of creep signals through time series reconstruction optimization and adaptive smoothing, and output high-fidelity, high-resolution clean time series component sequences. It realizes refined analysis and feature enhancement of the dynamic process of fault creep, and provides a time series input with clear structure and superior signal-to-noise ratio for multi-scale feature extraction and event detection.

[0046] (III) The system for realizing InSAR deformation time series analysis and fault creep event detection, based on the state transition model established by the Hidden Markov Model, can automatically identify the start, duration and termination states of creep events, realize automatic segmentation and state labeling of long time series, and greatly improve the automation level of event detection and the refinement of time series analysis.

[0047] (iv) This system for realizing InSAR deformation time series analysis and fault creep event detection accurately calculates the slip amount of each event based on piecewise linear fitting, and combines multi-dimensional features such as slip rate variation coefficient and deformation curve kurtosis. It uses a support vector machine model to realize intelligent classification of steady-state creep and intermittent accelerated creep, which enhances the understanding and analysis of fault slip behavior patterns and realizes accurate quantification and multi-feature classification of creep event slip amount. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the workflow of a system for realizing InSAR deformation time series analysis and fault creep event detection according to the present invention;

[0049] Figure 2 This is a data flow diagram of a system for realizing InSAR deformation time series analysis and fault creep event detection according to the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a system for realizing InSAR deformation time series analysis and fault creep event detection, including an integrated monitoring platform, which is communicatively connected to the following modules:

[0052] The temporal noise filtering module is used to extract creep deformation signals related to fault slip mechanism from the original InSAR deformation data of the target detection area by employing a spatiotemporal adaptive filtering algorithm combined with wavelet transform and independent component analysis (ICA). At the same time, it filters out interference from non-tectonic noise such as seasonal hydrology and thermal expansion, obtains the filtered temporal component sequence, and improves the fidelity of the tectonic signal of the data. The temporal noise filtering module includes a multi-scale temporal decomposition unit and an independent component separation unit.

[0053] The multi-scale temporal decomposition unit receives the raw InSAR deformation data of the target detection area and performs multi-scale decomposition on the raw InSAR deformation data using wavelet transform to separate different frequency components, thereby initially enhancing the identification ability of structure-related signals and realizing the initial separation of raw data in multiple frequency bands. This significantly enhances the spatiotemporal resolution of fault-related signals. The unit receives the raw InSAR deformation data of the target detection area and performs multi-scale decomposition on the raw InSAR deformation data using wavelet transform to obtain a set of time-series components containing different frequency components, achieving effective separation of signals in the time-frequency domain and preserving multi-scale tectonic activity information. Frequency-spatial domain analysis is performed on each time-series component in the time-series component set to screen out high-frequency and low-frequency time-series components related to fault slip mechanisms, initially separating non-tectonic noise interference, enhancing the identifiability of structure-related signals, and suppressing seasonal and other noise interference. The screened high-frequency and low-frequency time-series components are normalized, and the initially decomposed time-series component sequence is output, providing input for subsequent independent component analysis, unifying data dimensions, and improving the stability and convergence efficiency of subsequent processing.

[0054] The specific work involves: after receiving the raw InSAR deformation data of the target detection area, using wavelet transform to perform multi-scale decomposition on the raw temporal deformation data to extract temporal components of different frequencies. Specifically, the Morlet wavelet basis function is selected, and the scale parameter range is set from 2...0 Up to 2 5 (i.e., 1 to 32 scale levels), corresponding to a frequency range covering 0.01Hz to 1.0Hz, to meet the frequency domain characteristic requirements of fault slip-related signals. After decomposition, a set of time-series components containing different frequency bands is obtained. Each component contains both temporal and spatial dimension information. While preserving the original spatiotemporal structure of the data, high-frequency short-term fluctuations and low-frequency long-term trends are effectively separated. For the decomposed time-series component set, further frequency-spatial domain joint analysis is carried out to screen signal components related to the fault slip mechanism. In the frequency domain analysis, threshold conditions are set: high-frequency components with a frequency range ≥ 0.1Hz correspond to instantaneous slip or short-term accelerated creep events; low-frequency components with a frequency range ≤ 0.01Hz correspond to long-term steady-state creep processes. In the spatial domain analysis, combined with the fault strike and the geometric parameters of the slip surface, the following analysis is conducted: Spatial correlation assessment was performed on each component, and components with significant spatial coherence near the fault zone were selected. By analyzing the time series components related to non-tectonic noise such as seasonal hydrological changes and thermoelastic deformation, the identification accuracy of tectonic signals was improved. After signal screening, the selected high-frequency and low-frequency time series components were normalized to eliminate the dimensional differences between different components and improve the stability of subsequent analysis. The normalization method adopted Z-score standardization. The mean and standard deviation of the time series data of each component were calculated and linearly transformed to make the distribution of each time series component data conform to the standard normal distribution with a mean of 0 and a standard deviation of 1. The processed time series components were integrated into the preliminary decomposed time series component sequence in chronological order, preserving the spatiotemporal characteristics of fault slip correlation signals, while significantly suppressing non-tectonic interference.

[0055] The Independent Component Separation Unit (ICA) is used to further separate independent signal sources in InSAR deformation data based on multi-scale decomposition using wavelet transform, combined with spatiotemporal adaptive filtering algorithms and Independent Component Analysis (ICA). It extracts time-series components highly correlated with fault creep activity, effectively filters out non-tectonic noise interference in deformation analysis, integrates the extracted time-series components to obtain a filtered time-series component sequence, significantly suppresses non-tectonic noise, extracts high-fidelity creep time-series components, and improves the signal-to-noise ratio of tectonic signals. Based on the spatiotemporal adaptive filtering algorithm, it performs spatial domain smoothing and temporal filtering on the initially decomposed time-series component sequence to suppress local noise. The filtering process significantly improves the spatial and temporal continuity of the signal, providing a clean temporal basis for subsequent analysis. Independent component analysis is performed on the filtered time series to extract multiple statistically independent signal sources. Independent time series components highly correlated with fault creep activity are identified, effectively separating tectonic-related signals and eliminating non-tectonic interferences such as seasonality and hydrology, thus improving the accuracy of signal source identification. The extracted independent time series components are integrated and reconstructed to form a filtered time series sequence after filtering non-tectonic noise, generating a high-fidelity creep time series signal and significantly enhancing the spatiotemporal consistency of fault activity characteristics.

[0056] The specific work involves: based on the initially decomposed temporal component sequences, spatial domain smoothing and temporal filtering are performed using a spatiotemporal adaptive filtering algorithm to effectively suppress local noise interference. In the spatial domain smoothing, an adaptive window technique is employed, dynamically adjusting the filtering window size according to the coherence of the surrounding region of each pixel, with the window radius set to 3 to 7 pixels. For regions with coherence higher than 0.6, a small window (3 pixels) is used to preserve detailed features, while for regions with coherence lower than 0.4, the window is expanded to 7 pixels to enhance the smoothing effect. In the temporal filtering stage, Savitzky-Golay filtering is combined with... The filter performs local polynomial fitting on the time series of each pixel, with a window length of 9 time points and a polynomial order of 2, to smooth short-term fluctuations while maintaining the integrity of the temporal trend. Stable convergence is achieved through 2 to 3 iterations, effectively eliminating spatial random noise and high-frequency temporal oscillations, while ensuring that the structure of the fault-related signal in the spatiotemporal dimension is not destroyed. After completing the spatiotemporal adaptive filtering, independent component analysis is performed on the filtered time series to extract multiple statistically independent signal sources. The FastICA algorithm is used as the core processing method, with a convergence tolerance of 1×10⁻. 5 The maximum number of iterations is 500 to ensure stable convergence and computational efficiency. Input data is preprocessed by whitening for decorrelation and standardization. The top 10 to 15 components with a cumulative contribution rate exceeding 95% are retained as the analysis basis. During independent component decomposition, each independent component is estimated using the negative entropy maximization criterion. Combined with kurtosis as an independence metric, independent time-series components highly correlated with fault creep activity are identified. The specific criteria are: selecting components with significant average amplitude (more than twice the standard deviation of background noise) within the fault zone spatial range and exhibiting a continuous or intermittent trend in time series; and excluding components with obvious seasonal cycles or high correlation with environmental factors such as temperature and precipitation (correlation coefficient > 0.7) to ensure that the selected components possess typical characteristics of tectonic activity. The extracted components highly correlated with fault creep activity are then analyzed. Independent temporal components related to degree are integrated and reconstructed to form a filtered temporal component sequence after filtering out non-structural noise. During the integration process, a weighted superposition method is used, and weights are assigned based on the mean spatial coherence of each independent component within the fault zone (threshold > 0.5) and the temporal signal-to-noise ratio (threshold > 3). The weight coefficients are normalized and summed to 1. The original spatiotemporal dimension structure is maintained during reconstruction, the temporal resolution is consistent with the input sequence, and the spatial range covers the target fault zone and its affected area. After generating the filtered temporal component sequence, the root mean square error (RMSE < 0.5 mm) of the residuals between the filtered sequence and the original sequence within the fault zone is calculated to verify the filtering effectiveness and ensure the fidelity of the structural signal. The filtered sequence shows that non-structural noise is significantly suppressed, and the fault creep signal is highlighted, which can be directly used for subsequent temporal reconstruction optimization and event detection analysis.

[0057] The creep signal enhancement module is used to further process the filtered time-series component sequence to obtain a clean time-series component sequence through time-series reconstruction and signal enhancement techniques, based on filtering non-structural noise. This improves the continuity and signal-to-noise ratio of the creep-dominated deformation time sequence, providing high-fidelity and high-resolution time-series input for subsequent event detection. The creep signal enhancement module includes a time-series reconstruction optimization unit and a signal-to-noise ratio enhancement unit.

[0058] The temporal reconstruction optimization unit is used to selectively reconstruct the temporal components extracted from the filtered temporal component sequence, remove irrelevant components, highlight the temporal features related to the fault slip mechanism, continuously enhance the intensity of the creep signal in the temporal dimension, obtain the temporal component reconstruction sequence, enhance the temporal continuity and identifiability of the creep signal, and improve the accuracy of event detection. It receives the filtered temporal component sequence, performs segmented analysis on the temporal components based on the slip time window, identifies the feature components related to the fault slip mechanism in each segment, and accurately extracts the temporal components of fault activity through local analysis. By analyzing local features, the signal's recognizability within a short time window is improved. Temporal components unrelated to fault slip are removed, and the remaining feature components are weighted, superimposed, and time-aligned to reconstruct a temporal component reconstruction sequence that highlights the intensity of the creep signal. This effectively suppresses non-tectonic interference, significantly enhances the continuity and spatiotemporal consistency of fault-related deformation signals, and performs trend fitting and baseline correction on the reconstructed temporal component sequence. This enhances the continuity and structure of the time-series signal in the time dimension, eliminates long-term trend drift and system offset, and improves the structural stability of the time-series signal and the reliability of event detection.

[0059] The specific work involves: after receiving the time series components filtered for non-structural noise, firstly, segmenting the sequence based on a preset sliding time window. The length of the sliding time window is set according to the timescale characteristics of the target fault creep activity, with a window length of 30 to 90 observation periods selected. The sliding step size is set to 1 / 3 to 1 / 2 of the window length to balance the temporal resolution and continuity of the analysis. Within each sliding time window, the statistical characteristics of each time series component are calculated, including but not limited to local mean, variance, trend slope, and spatial covariance with fault strike. By setting threshold conditions, features significantly correlated with the fault slip mechanism are automatically identified. Feature components are identified to achieve localized, refined analysis of long-term time-series data, effectively capturing the phased characteristics of creep events in the temporal dimension. After identifying the feature components within each window, irrelevant components are removed. The removal criteria are based on the spatial-temporal correlation index between the component and the fault slip mechanism: spatially, the average amplitude of the component within the main fault zone (based on the fault trace buffer, the width is usually set to 1-2 km) should not be less than 1.5 times the standard deviation of the regional background noise; temporally, its autocorrelation coefficient should remain above 0.5 within a lag of 1-3 periods to ensure the temporal continuity of the signal. For feature components that meet the retention criteria, a weighted superposition method is used for reconstruction. The signal-to-noise ratio (SNR≥4) and mean spatial coherence (≥0.6) of each component within its corresponding window are dynamically calculated and normalized to ensure a total weight of 1 for each window. The reconstruction process employs a temporal alignment algorithm to ensure smooth signal transitions between windows, ultimately generating a temporal component reconstruction sequence that highlights the intensity of the creep signal. This sequence significantly enhances the continuity and identifiability of fault-related deformation signals in the time dimension. Trend fitting and baseline correction are then performed on the temporal component reconstruction sequence to further improve its structure and stability. Trend fitting uses a locally weighted regression method, with the fitting window length set to 15%-20% of the total temporal length and a polynomial order of 2. To smooth local fluctuations and preserve medium- to long-term trends, the fitted trend term is used to characterize the background deformation of creep activity. Baseline correction is then performed: the fitted trend term is subtracted from the original reconstructed sequence to obtain the detrended residual sequence. The residual sequence is then zero-meaned, i.e., its global average value in the time dimension is subtracted to eliminate systematic shifts. During the correction process, the effectiveness and fidelity of the correction operation are verified by calculating that the root mean square of the residuals (RMSE) of the sequences before and after correction within the main fault zone is less than 0.3 mm and the coefficient of determination (R²) between the trend term and the original sequence is greater than 0.7. The final output time series signal has stronger temporal continuity and structure.

[0060] The signal-to-noise ratio (SNR) enhancement unit employs adaptive smoothing and trend correction algorithms to further suppress residual noise in the reconstructed temporal component sequence, improve the temporal consistency and resolvability of the creep signal, and output a clean temporal component sequence. This provides high-quality input for subsequent mode decomposition and event recognition. The adaptive smoothing algorithm performs local noise suppression on the reconstructed temporal component sequence, maintaining the temporal abrupt change characteristics of the creep signal, significantly suppressing random noise, improving the local stationarity of the signal, and preserving the temporal abrupt change details of fault creep. Based on the trend correction algorithm, the smoothed sequence is fitted with an overall trend, removing long-term drift and seasonal residual components, effectively eliminating background drift and seasonal residues, highlighting the deformation trend dominated by the structure, improving temporal consistency, and outputting a clean temporal component sequence after noise suppression and trend correction. This provides high-fidelity and high-resolution temporal input for subsequent mode decomposition analysis. The output signal has a high SNR and high temporal fidelity, providing stable and reliable input data for mode decomposition.

[0061] The specific work involves: after receiving the reconstructed sequence of time components, an adaptive smoothing algorithm is used to suppress local noise in the sequence to preserve the temporal abrupt change characteristics in the creeping signal. The smoothing intensity is dynamically adjusted based on the local signal-to-noise ratio (SNR): for segments with an SNR higher than 4, a smaller smoothing window (window length of 5 time points, using a Gaussian kernel function with a standard deviation of 1.5) is used to avoid excessive smoothing leading to signal detail loss; for segments with an SNR lower than 2, the smoothing window is increased to 9 time points, and Savitzky-Golay filtering (polynomial order 2) is used to effectively suppress high-frequency noise in space. In the domain, each pixel is weighted based on its spatial coherence within the main fault zone (threshold set to 0.6) to ensure spatial structure consistency. During smoothing, iterative calculations (maximum 3 times) are performed to converge the root mean square error of local residuals to below 0.2 mm, ultimately yielding the noise-suppressed time series. Building upon local noise suppression, trend correction is further applied to the smoothed time series to remove long-term drift and residual seasonal components. Trend correction employs a piecewise polynomial fitting method: the entire time series is divided into equal-length, non-overlapping segments, each containing 30 to 50 observation points; second-order polynomial fitting is then performed on each segment. The goodness-of-fit threshold was set to R² ≥ 0.7. For segments with insufficient goodness-of-fit, the local weighted regression method was automatically switched. The fitting window length was set to 15% of the total time series length. A cubic weighting kernel was used as the weighting function. The trend term obtained from the fitting was subtracted from the original sequence, and the residual sequence was zero-mean processed to eliminate systematic shifts. During the correction process, the standard deviation of the residuals within the fault main zone was monitored in real time to ensure that it did not exceed 0.3 mm, maintaining the fidelity of the structural signal. After adaptive smoothing and trend correction, a clean time series component sequence was output, which has high fidelity and high resolution: time dimension The sampling interval remains consistent with the original observation, and the sequence length remains unchanged. In terms of spatial dimension, it covers the target fault zone and its affected area, and the spatial resolution maintains the original pixel scale. In the output clean time series component sequence, non-tectonic noise (seasonal hydrological fluctuations, thermoelastic deformation, etc.) is significantly suppressed, and the time series signals related to fault creep are enhanced and preserved. The clean time series component sequence can be directly input into the subsequent modal decomposition analysis module to further extract multi-scale creep dynamic features and provide structured time series input for event state identification. The entire processing flow ensures repeatability and stability through parametric design, meeting the needs of engineering analysis.

[0062] The modal decomposition analysis module is used to decompose the temporal components in the clean temporal component sequence into multiple intrinsic mode functions using the variational mode decomposition (VMD) algorithm, extract creep dynamic features at different time scales, form a creep dynamic feature sequence, extract multi-scale creep dynamic features, and provide structured, multi-modal temporal input for event state recognition.

[0063] The event state recognition module establishes a creep event state transition model based on the Hidden Markov Model (HMM) algorithm. It takes the extracted creep dynamic features as input, automatically identifies the start, duration and termination states of creep events through probabilistic inference, and outputs the event time sequence. This enables intelligent segmentation and state labeling of the event time sequence, and improves the intelligence and accuracy of event boundary detection.

[0064] The slip quantification and classification module is used to accurately calculate the slip amount of each creep event based on event state identification, combined with temporal deformation variables and event duration. It also establishes a creep event classification model based on slip rate, duration, and deformation curve characteristics, automatically identifying and classifying steady-state creep and intermittent accelerated creep, improving the depth of understanding of fault slip behavior, realizing accurate calculation and intelligent classification of creep event slip amount, and enhancing the understanding of fault slip patterns and disaster assessment capabilities.

[0065] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the modal decomposition analysis module performs the following steps: receiving a clean time-series component sequence, decomposing it into multiple intrinsic mode functions using a variational mode decomposition algorithm, which can adaptively extract structural signal components at different time scales, enhance the analytical capability of creep dynamic features, and select modal components related to the creep dynamic process based on the center frequency and energy distribution of each intrinsic mode function, effectively eliminating noise and seasonal interference, retaining a subset of modes significantly related to fault activity, improving feature specificity, recombining the selected modal components to form a multi-time-scale creep dynamic feature sequence, constructing multi-scale time-series features with clear physical meaning, and providing structured, high-resolution input for subsequent event recognition;

[0066] The specific work involves: receiving a clean time-series component sequence, and then using a variational mode decomposition algorithm to decompose the time-series signal into multiple intrinsic mode functions (EMFs). First, the input signal is initialized with parameters set. The number of modes, K, is determined to be 6 to 10 based on the signal length and sampling frequency to ensure sufficient decomposition and avoid over-decomposition. The balancing parameter α is set to 2000 to control the bandwidth of each mode. The noise tolerance τ is set to 0 to ensure the decomposition process is insensitive to noise. An iterative solution using the alternating direction multiplier method is employed, with a maximum number of iterations set to 500 and a convergence tolerance of 1 × 10⁻⁶. -7During the decomposition process, the center frequency of each mode is estimated in real time using Hilbert transform, and its fluctuation around the initial center frequency is constrained to not exceed 0.5 Hz. This adaptively decomposes the clean time-series component sequence into a set of intrinsic mode functions with a defined frequency band structure. Based on the center frequency and energy distribution characteristics of each intrinsic mode function, mode components related to the creep dynamic process are screened. The center frequency analysis is based on Fourier spectral density calculation, with a screening range of 0.01 Hz to 0.5 Hz. This frequency band covers typical dynamic processes of fault creep, including long-term steady-state deformation and short-term acceleration events. The energy distribution characteristics are evaluated using the proportion of each mode's variance to the total signal variance, with a threshold set not less than 8% of the total energy. Simultaneously, the mean spatial coherence of each mode within the fault's main band is required to be not less than 0.5. For mode components that meet the conditions, their temporal autocorrelation is further examined, with a lag of 1 to 3 cycles. The autocorrelation coefficient of the period needs to be maintained above 0.4. Through multi-criteria joint screening, interference modes related to noise, seasonal changes or non-tectonic deformation are effectively eliminated, and a subset of modes significantly related to creep activity is retained. The screened modal components are recombined in chronological order to construct a creep dynamic feature sequence with multiple time scales. During the recombination process, the amplitude of each modal component is normalized to be unified to the same dimension. Each mode is superimposed in a linear weighting manner. The weight is determined based on its energy proportion and spatial coherence. The weight coefficients are normalized to a sum of 1. The creep dynamic feature sequence generated after recombination maintains the original temporal resolution and the spatial dimension covers the target fault zone and adjacent areas. To verify the effectiveness of the recombined sequence, the root mean square error of the residual between it and the clean time series component sequence in the main fault zone is calculated to ensure that it does not exceed 0.4 mm. Finally, a creep dynamic feature sequence with clear physical meaning and multi-scale temporal characteristics is output.

[0067] The event state recognition module performs the following steps: It constructs a creep event state transition model based on a Hidden Markov Model, defines three states for creep events: start, duration, and termination, and realizes physical modeling of event states, providing a probabilistic basis for automatic segmentation and improving the structure of detection. The creep dynamic feature sequence is input into the constructed creep event state transition model, and the event state at each time point is identified through probabilistic inference and Viterbi decoding algorithms. The most probable state at each time point is automatically output, achieving temporal intelligent recognition and boundary extraction of creep events. An event time sequence with state labels is output, marking the start time, duration, and termination time of each creep event, generating a structured event table that can be directly used for slip calculation, improving the efficiency and consistency of subsequent analysis.

[0068] The specific work involves constructing a state transition model for creep events based on a Hidden Markov Model (HMM). Creep events are defined as having three states: initiation, continuation, and termination. A state transition probability matrix is ​​set based on historical observation data and fault activity patterns. The transition probability from the initiation state to the continuation state is set to 0.85, the self-transition probability of the continuation state is set to 0.70, the transition probability from the continuation state to the termination state is set to 0.15, and the termination state, as an absorbing state, has a self-transition probability of 1.0. A Gaussian mixture model is used to model the observation probability density function, with each state corresponding to 2 to 3 Gaussian components. The mean and covariance matrices are estimated based on the historical distribution of the creep dynamic characteristic sequence. The model training uses the Baum-Welch algorithm, with a maximum of 100 iterations and a convergence tolerance of 1 × 10⁻⁶. -6 To ensure stable convergence of model parameters, after training, the model is capable of inferring the corresponding state at each time point based on the input feature sequence. The creeping dynamic feature sequence is input into the trained Hidden Markov Model (HMM), and the probability of each state at each time point is calculated using a forward-backward algorithm. The Viterbi decoding algorithm is then used to solve for the most probable state sequence. The input feature sequence needs to be standardized to have a mean of 0 and a variance of 1 to match the assumed distribution of the model's observation probabilities. During decoding, the state backtracking path length of the Viterbi algorithm is consistent with the time length of the input sequence, and each state transition is based on logarithmic probability to avoid numerical underflow. For each time point, its most probable state label is output: the initial state is labeled as 1, the continuous state as 2, and the final state as 3, achieving automatic segmentation of time-series data and separating continuous states. The signal is converted into an event sequence with well-defined state boundaries. Based on the decoded state label sequence, the start and end times and duration of each creep event are extracted. The event start time is defined as the moment when the state changes from non-start to start, and the end time is defined as the moment when the state changes from continuous to terminated. The event duration is calculated as the difference between the end time and the start time, in the number of observation periods. The output event time series is recorded in a structured table format, including fields for event number, start time index, end time index, and duration. The statistical measures of the mean and variance of the feature sequence within each event segment are calculated to further verify the rationality of the event detection and compare it with the background noise level to ensure that the detected events are statistically significant. The final output is an event time series that can be directly used for subsequent slip calculation and classification analysis.

[0069] The slip quantification and classification module performs the following steps: Based on the event time sequence, it extracts the time-series deformation and event duration corresponding to each creep event, accurately obtains the deformation data and duration within the event segment, provides basic data support for slip calculation and event classification, and improves the accuracy of subsequent analysis. Combining the fault slip mechanism and deformation curve characteristics, it establishes a slip rate calculation model, quantifies the slip amount and average slip rate of each creep event, realizes the physical quantification of slip amount and rate, enhances the objective description of the fault slip process and the ability to identify dynamic characteristics, and constructs a creep event classification model based on the average slip rate, duration and deformation curve morphology characteristics. It automatically identifies and classifies steady-state creep and intermittent accelerated creep, automates the event type discrimination, and improves the understanding of fault slip behavior patterns and the effectiveness of geological disaster risk classification management.

[0070] The specific work involves: after obtaining the event time series, extracting the time series deformation data corresponding to each creep event segment from the original InSAR deformation data based on the start and end time indices of the creep events. Deformation data is extracted in pixels, and spatial averaging is performed along the main fault zone (buffer width set to 1.5 km) to obtain the time-deformation curve for each event. The event duration is directly calculated from the start and end time difference in the state sequence, with the unit being the number of observation periods. To ensure data consistency, baseline correction is performed on the deformation data of all event segments, deducting the mean background deformation of the first 30 periods of the event, and uniformly applying Z-score standardization to eliminate dimensional differences. Finally, each event is represented as a deformation curve vector and the event duration. As input for slip calculation and classification; based on the fault slip mechanism, the slip amount is obtained by integrating the slip rate over time, and a slip rate model is established using a piecewise linear fitting method. Specifically, the event deformation curve is evenly divided into 5-8 segments according to time, and linear regression is performed on each segment. The slope is the average slip rate for that time period. slip The calculation is as follows:

[0071] ;

[0072] in, For each segment duration, For the first The average slip rate of the segment, The number of segments is a parameter in the model that dynamically adjusts based on the event duration, with shorter events (...). The period is taken as 5 segments, and the long event ( Eight segments are selected for the period. Cubic spline interpolation is used to smooth the rate abrupt change points (rate differences between adjacent segments exceeding 0.5 mm / cycle) to improve accuracy. Finally, the slip amount for each event is output. and average slip rate Extraction includes the coefficient of variation of slip rate. (Threshold set to 0.3), event duration kurtosis of deformation curve (Reflecting the severity of the event) and the cumulative deformation curvature integral Based on the characteristics of creep classification, and considering the average slip rate, duration, and deformation curve morphology, a creep event classification model was constructed, dividing events into two categories: steady-state creep and intermittent accelerated creep. A support vector machine classifier was used, with the radial basis function as the kernel function and regularization parameters as follows: kernel coefficient The training data comes from a historically validated event database, and five-fold cross-validation is used to ensure generalization performance. In classification decisions, if... and If it is determined to be steady-state creep; If there is a significant acceleration phase, i.e., the local rate exceeds twice the average rate, it is judged as intermittent accelerated creep. Finally, the event classification label and confidence probability are output and integrated into the event report.

[0073] The expression for the coefficient of variation of the slip rate is as follows:

[0074] ;

[0075] In the formula: The coefficient of variation of the slip rate; The standard deviation of the slip rate sequence; The average slip rate;

[0076] The expression for the kurtosis of the deformation curve is as follows:

[0077] ;

[0078] ;

[0079] ;

[0080] In the formula: Kubularity of the deformation curve; The fourth central moment of the deformation curve; The standard deviation of the deformation curve; For the first The deformation at each observation time; This represents the average value of the deformation curve; This represents the number of observation points on the event deformation curve.

[0081] The expression for the cumulative deformation curvature integral is as follows:

[0082] ;

[0083] ;

[0084] In the formula: This is the integral of the cumulative deformation curvature; For the deformation curve at time t The curvature; The first derivative (rate) of the deformation curve; The second derivative (acceleration) of the deformation curve; , This refers to the start and end times of the event (the observation time).

[0085] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0086] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for realizing InSAR deformation time series analysis and fault creep event detection, comprising an integrated monitoring platform, characterized in that, The integrated monitoring platform has the following communication modules: The temporal noise filtering module is used to extract creep deformation signals related to fault slip mechanism from the original InSAR deformation data of the target detection area by employing a spatiotemporal adaptive filtering algorithm, combined with wavelet transform and independent component analysis, while filtering out interference from non-tectonic noise such as seasonal hydrology and thermal expansion, and obtaining the filtered temporal component sequence. The creep signal enhancement module is used to further process the filtered non-structural noise into a clean time-series component sequence through time-series reconstruction and signal enhancement techniques, based on the filtering of non-structural noise. The mode decomposition analysis module is used to decompose the time-series components in the clean time-series component sequence into multiple intrinsic mode functions using the variational mode decomposition algorithm, extract the creep dynamic features at different time scales, and form a creep dynamic feature sequence. The event state recognition module establishes a creep event state transition model based on the Hidden Markov Model algorithm. It takes the extracted creep dynamic features as input, automatically identifies the start, duration and termination states of the creep event through probabilistic inference, and outputs the event time sequence. The slip quantization and classification module is used to accurately calculate the slip amount of each creep event based on the event state recognition, combined with the temporal deformation and event duration. Based on the slip rate, duration and deformation curve characteristics, it establishes a creep event classification model to automatically identify and classify steady-state creep and intermittent accelerated creep.

2. The system for realizing InSAR deformation time series analysis and fault creep event detection according to claim 1, characterized in that: The temporal noise filtering module includes a multi-scale temporal decomposition unit and an independent component separation unit; The multi-scale temporal decomposition unit is used to receive the original InSAR deformation data of the target detection area and use wavelet transform to perform multi-scale decomposition on the original InSAR deformation data to separate different frequency components. The independent component separation unit is used to further separate independent signal sources in InSAR deformation data based on the multi-scale decomposition of wavelet transform, combined with spatiotemporal adaptive filtering algorithm and independent component analysis, extract time-series components that are highly correlated with fault creep activity, and integrate the extracted time-series components to obtain a filtered time-series component sequence.

3. The system for realizing InSAR deformation time series analysis and fault creep event detection according to claim 2, characterized in that: The multi-scale temporal decomposition unit performs the following steps: The raw InSAR deformation data of the target detection area is received, and the raw InSAR deformation data is decomposed into multiple scales using wavelet transform to obtain a set of time-series components containing different frequency components. Frequency-spatial domain analysis was performed on each time series component in the time series component set to screen out high-frequency and low-frequency time series components related to the fault slip mechanism and to preliminarily separate non-tectonic noise interference. The selected high-frequency and low-frequency time-series components are normalized to output the preliminary decomposed time-series component sequence.

4. The system for realizing InSAR deformation time series analysis and fault creep event detection according to claim 2, characterized in that: The independent component separation unit performs the following steps: The temporal component sequences of the preliminary decomposition are smoothed in the spatial domain and filtered in the temporal domain based on the spatiotemporal adaptive filtering algorithm to suppress local noise interference. Independent component analysis was performed on the filtered time-series components to extract multiple statistically independent signal sources and identify independent time-series components that are highly correlated with fault creep activity. The extracted independent time-series components are integrated and reconstructed to form a filtered time-series component sequence after filtering out unconstructed noise.

5. The system for realizing InSAR deformation time series analysis and fault creep event detection according to claim 2, characterized in that: The creep signal enhancement module includes a timing reconstruction optimization unit and a signal-to-noise ratio enhancement unit; The time-series reconstruction optimization unit is used to selectively reconstruct the time-series components extracted from the filtered time-series component sequence, remove irrelevant components, highlight the time-series features related to the fault slip mechanism, and obtain the time-series component reconstruction sequence. The signal-to-noise ratio enhancement unit is used to further suppress residual noise in the time-series component reconstruction sequence by employing an adaptive smoothing and trend correction algorithm, and output a clean time-series component sequence.

6. The system for realizing InSAR deformation time series analysis and fault creep event detection according to claim 5, characterized in that: The timing reconstruction optimization unit performs the following steps: The filtered time series component sequence is received, and the time series component is segmented and analyzed based on a sliding time window to identify the characteristic components related to the fault slip mechanism in each segment. Temporal components unrelated to fault slip are removed, and the remaining characteristic components are weighted, superimposed, and time-aligned to reconstruct a temporal component reconstruction sequence that highlights the creep signal intensity. The temporal component reconstruction sequence is then subjected to trend fitting and baseline correction.

7. The system for realizing InSAR deformation time series analysis and fault creep event detection according to claim 5, characterized in that: The signal-to-noise ratio enhancement unit performs the following steps: An adaptive smoothing algorithm is used to suppress local noise in the reconstructed sequence of the time components, thus preserving the temporal abrupt change characteristics of the creeping signal; The smoothed sequence is fitted with an overall trend based on a trend correction algorithm to remove long-term drift and seasonal residual components, and output a clean time-series component sequence after noise suppression and trend correction.

8. The system for realizing InSAR deformation time series analysis and fault creep event detection according to claim 5, characterized in that: The modal decomposition analysis module performs the following steps: The clean time-series component sequence is received and decomposed into multiple intrinsic mode functions using a variational mode decomposition algorithm. Based on the center frequency and energy distribution of each intrinsic mode function, modal components related to the creep dynamic process are selected. The selected modal components are recombined to form a multi-timescale creep dynamic feature sequence.

9. A system for realizing InSAR deformation time series analysis and fault creep event detection according to claim 8, characterized in that: The event status identification module performs the following steps: A state transition model for creep events is constructed based on a hidden Markov model, defining three states for creep events: start, duration, and termination. The creep dynamic feature sequence is input into the constructed creep event state transition model, and the event state at each time point is identified by probability inference and Viterbi decoding algorithm; Output the event sequence with status labels, marking the start time, duration, and end time of each creep event.

10. A system for realizing InSAR deformation time series analysis and fault creep event detection according to claim 9, characterized in that: The slip quantization classification module performs the following steps: Based on the event time sequence, extract the time sequence deformation and event duration corresponding to each creep event; By combining the fault slip mechanism and deformation curve characteristics, a slip rate calculation model is established to quantify the slip amount and average slip rate of each creep event; Based on the average slip rate, duration, and deformation curve morphology, a creep event classification model is constructed to automatically identify and classify steady-state creep and intermittent accelerated creep.

Images