A deep dimension reduction simulation system and method for time-frequency feature fusion of main aftershock sequence

Abstract

The application relates to the field of data processing, and discloses a main aftershock sequence time-frequency feature fusion deep dimension reduction simulation system and method, wherein the main aftershock sequence time-frequency feature fusion deep dimension reduction simulation method comprises the following steps: adaptively segmenting original waveform data from multiple seismic monitoring nodes according to a non-fixed window; performing cross-window aggregation on local amplitude anomalies and burst energy features existing in preliminary sequence representation; dynamically mapping potential aftershock signals and main shock features by constructing a bidirectional association network; iteratively fusing feature differences in a response evolution diagram and node distribution information; introducing a hierarchical dimension reduction processing mechanism based on a digital constraint model, and performing selective compression on signal modes of different granularities; and dynamically comparing and weighting integrating sequence features of different time segments by using an index system. The application has the advantage of improving the real-time performance of large-scale seismic data processing.

Description

Technical Field

[0001] This invention relates to the field of data processing, specifically to a deep dimensionality reduction simulation system and method for fusing time-frequency features of mainshock and aftershock sequences. Background Technology

[0002] With the construction of urban earthquake monitoring networks and the continuous increase in observation stations, the amount of data generated from mainshock and aftershock sequences is growing exponentially. Existing earthquake analysis methods mainly rely on single feature extraction in the time or frequency domains, performing statistical processing or manual screening on high-dimensional time-frequency data. However, in practical applications, observation data suffers from uneven sampling, signal-noise interference, and difficulty in identifying low-amplitude aftershock events, leading to low computational efficiency and insufficient information utilization in traditional methods for feature extraction and sequence simulation. Especially in the digital processing of large-scale, multi-source earthquake observation data, the data redundancy and storage overhead brought by high-dimensional time-frequency features increase significantly, while traditional dimensionality reduction methods struggle to achieve real-time computation while ensuring information integrity. Therefore, it is essential to design a deep dimensionality reduction simulation method that integrates the time-frequency features of mainshock and aftershock sequences to improve the real-time performance of large-scale earthquake data processing. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a deep dimensionality reduction simulation system and method for fusing time-frequency features of mainshock and aftershock sequences. This system has the advantage of improving the real-time performance of large-scale earthquake data processing and solves the problems mentioned in the background technology.

[0004] To achieve the aforementioned goal of improving the real-time performance of large-scale earthquake data processing, this invention provides the following technical solution: a deep dimensionality reduction simulation method for fusing time-frequency features of mainshock and aftershock sequences, comprising the following steps:

[0005] The raw waveform data from multiple seismic monitoring nodes are adaptively segmented using non-fixed windows. The transient changes and local trends of each signal segment are recorded in a multi-channel mapping form to generate a preliminary sequence representation that can be dynamically adjusted.

[0006] Local amplitude anomalies and sudden energy characteristics present in the preliminary sequence representation are aggregated across windows. By constructing a bidirectional correlation network, potential aftershock signals are dynamically mapped with mainshock characteristics to form a traceable response evolution map.

[0007] The feature differences in the response evolution graph are iteratively fused with the node distribution information. The non-obvious dependencies between signal changes are identified by using time-delay correlation analysis and asynchronous matching strategy, and a digital constraint model of sequence features is generated.

[0008] Based on the digital constraint model, a hierarchical dimensionality reduction processing mechanism is introduced to selectively compress signal patterns of different granularities, and key changes are enhanced through the residual feedback mechanism to form an index system that reflects the stability of the sequence and the potential aftershock trend.

[0009] The indicator system is used to dynamically compare and weight the sequence features of different time segments, and to calculate the information retention and feature recognition rate.

[0010] Preferably, the process of generating a preliminary sequence representation that can be dynamically adjusted is as follows:

[0011] The raw waveform data of each monitoring node is dynamically segmented and collected according to a non-fixed time window, and transient changes and local trend features are extracted in each signal segment.

[0012] Multi-channel mapping processing is performed on each signal segment, and the feature weights of each channel can be dynamically adjusted according to changes in node status and data quality.

[0013] By iteratively updating the mapping results of each channel, the multi-channel signal features of all nodes are integrated to form a preliminary sequence representation.

[0014] Preferably, the process of cross-window aggregation of local amplitude anomalies and burst energy features present in the preliminary sequence representation is as follows:

[0015] The preliminary sequence representation is scanned within a sliding time window to identify the temporal location and amplitude changes of local amplitude anomalies and sudden energy characteristics within each time segment;

[0016] Align and fuse the local amplitude anomalies and energy burst features detected in adjacent windows to establish a sequence chain that reflects the continuity of amplitude fluctuations and energy release;

[0017] Based on the fused amplitude and energy characteristic distribution, a traceable local signal pattern chain is constructed, and through cross-window aggregation, the abnormal responses of node signals in the spatial and temporal dimensions are uniformly integrated to generate cross-window aggregation results.

[0018] Preferably, the process of forming a traceable response evolution map is as follows:

[0019] A bidirectional correlation network between nodes is generated based on the cross-window aggregation results, and the similarity of signal features and the degree of temporal evolution coupling of each node are quantified.

[0020] Based on the interaction intensity and edge weight changes of nodes in the bidirectional correlation network, statistical correlation information between nodes is extracted and a node correlation matrix is ​​generated to describe the dynamic relationship between the main shock and potential aftershocks.

[0021] Potential aftershock signals and mainshock characteristics are projected onto the response evolution diagram through a nonlinear mapping function, and the time evolution path and weight distribution are determined based on the node correlation matrix.

[0022] Potential abnormal fluctuation regions are identified in the bidirectional correlation network, and the disturbance information is fed back to the node correlation matrix to perform weight adjustment. The corrected node correlation relationship and time series evolution data are integrated to form a traceable response evolution diagram.

[0023] Preferably, the process of iteratively fusing the feature differences and node distribution information in the response evolution graph is as follows:

[0024] The signal characteristic differences of each node in the response evolution diagram are weighted and combined with the node's geographical location and monitoring coverage information;

[0025] The fusion weights of each node are updated through multiple iterations, and the feature associations and spatial distribution relationships between nodes are calculated synchronously in each iteration.

[0026] The features of each node after iterative fusion are integrated to form a fused feature representation.

[0027] Preferably, the process of generating a digital constraint model for sequence features is as follows:

[0028] The signals of each node in the fused feature representation are paired in pairs according to the time series, the delay correlation coefficient of the signals is calculated, and potential nonlinear dependencies are identified by delay scanning;

[0029] For node signals with time misalignment or asynchronous sampling, an asynchronous matching strategy is applied to quantify the evolution synchronization and feature differences between nodes through phase adjustment and local feature comparison.

[0030] By integrating the results of delay correlation analysis with asynchronous matching metrics, a digital constraint model for sequence features is formed for hierarchical dimensionality reduction processing.

[0031] Preferably, the process of selectively compressing signal modes of different granularities is as follows:

[0032] The node signals and dependencies in the digital constraint model are divided into different granularities according to the hierarchy, and a hierarchical dimensionality reduction processing mechanism is applied to each layer to perform structured mapping and feature extraction of the signal patterns.

[0033] Under the hierarchical dimensionality reduction processing mechanism, selective compression is performed on the signal patterns of each granularity layer to retain key evolutionary features and reduce redundant information, while generating low-dimensional representations that can be connected between layers.

[0034] By integrating the dimensionality reduction results from each layer, a dimensionality-reduced signal representation with a hierarchical structure and containing multi-granularity signal features is formed.

[0035] Preferably, the process of forming an indicator system that reflects sequence stability and potential aftershock trends is as follows:

[0036] The difference between the dimensionality-reduced signal representation and the original sequence is calculated, and residual information is extracted to identify key change points.

[0037] By using a residual feedback mechanism, residual information is injected into the dimension-reduced signal representation to amplify key changes;

[0038] The integrated and enhanced signal representation forms an indicator system that reflects sequence stability and potential aftershock trends.

[0039] Preferably, the process of calculating information preservation and feature recognition rate is as follows:

[0040] The indicator system dataset is dynamically compared and weighted by time segment, and the information retention rate and feature recognition rate of each segment are calculated.

[0041] By comparing the trend of feature changes across different time segments, an evaluation result reflecting the preservation of sequence information and the ability to identify features is generated.

[0042] A deep dimensionality reduction simulation system for fusing time-frequency features of mainshock and aftershock sequences includes:

[0043] Adaptive segmentation module: Performs non-fixed window adaptive segmentation on seismic monitoring data based on waveform change characteristics to generate preliminary sequence representation;

[0044] Association and aggregation module: Aggregates local amplitude anomalies and energy mutation characteristics to construct a dynamic management mapping between the mainshock and aftershocks;

[0045] Feature fusion module: fuses response evolution features with node distribution information to identify implicit dependencies between signal changes;

[0046] Hierarchical dimensionality reduction module: Performs selective dimensionality reduction and residual enhancement on multi-level signal patterns to extract stability and trend indicators;

[0047] Dynamic evaluation module: Performs dynamic comparison and weighted analysis on sequence features of different time segments, and calculates information retention and feature recognition rate.

[0048] Compared with existing technologies, this invention provides a deep dimensionality reduction simulation system and method for fusing time-frequency features of mainshock and aftershock sequences, which has the following beneficial effects:

[0049] This invention achieves precise capture of transient changes and local trends by adaptively segmenting and mapping multi-node raw waveform data of the mainshock and aftershock sequences, fully preserving the fine-grained features of the signal. Through cross-window aggregation and bidirectional correlation network construction, it effectively integrates the spatial and temporal coupling information of potential aftershocks and mainshock characteristics, enabling dynamic tracking and evolution visualization of aftershock signals. Iterative fusion of response evolution diagram feature differences, combined with time-delay correlation analysis and asynchronous matching strategies, identifies non-explicit dependencies, improving the accuracy of sequence feature modeling. A hierarchical dimensionality reduction mechanism based on a digital constraint model and residual feedback enhancement effectively compresses signal patterns of different granularities and enhances key changes, ensuring that the dimensionality-reduced signal maintains information integrity while highlighting potential aftershock trends. Dynamic comparison and weighted integration of sequence features through an index system quantifies information retention and feature recognition rate, providing a high-precision, traceable quantitative assessment tool for earthquake sequence analysis. This invention significantly improves the monitoring, identification, and prediction capabilities of mainshocks, aftershocks, and potential aftershock activity, and has significant application value for earthquake risk assessment, disaster early warning, and seismological research. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the method of the present invention;

[0051] Figure 2 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0052] 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.

[0053] Example 1: Please refer to Figure 1 As shown in the figure, a deep dimensionality reduction simulation method for fusing time-frequency features of mainshock and aftershock sequences in an embodiment of the present invention includes the following steps:

[0054] S1: The raw waveform data from multiple earthquake monitoring nodes are adaptively segmented using a non-fixed window, and the transient changes and local trends of each signal segment are recorded in a multi-channel mapping form to generate a preliminary sequence representation that can be dynamically adjusted.

[0055] The process of generating a dynamically adjustable preliminary sequence representation in S1 is as follows:

[0056] The raw waveform data of each monitoring node is dynamically segmented and acquired according to non-fixed time windows, and transient changes and local trend features are extracted within each signal segment. In the mainshock and aftershock observation network, each monitoring node continuously records seismic waveform data. To avoid the truncation of key signals caused by fixed windows, a non-fixed time window division strategy based on energy adaptation is adopted. The local energy density and frequency drift of the continuous waveform sequence are calculated. When the local energy exceeds the threshold or the frequency change rate increases significantly, the time window boundary is automatically adjusted to make the segmentation more consistent with the non-stationary characteristics of seismic activity. Within each segment, transient changes (including amplitude change rate and frequency drift rate) and local trend features (such as average energy, amplitude gradient, and spectral shape changes) of the signal are extracted.

[0057] Multi-channel mapping processing is performed on each signal segment, and the feature weights of each channel can be dynamically adjusted according to the node status and data quality changes. To address the differences in data reliability and signal quality among different monitoring nodes, a multi-channel feature mapping model is established. Each channel corresponds to different physical quantities or feature dimensions of the waveform (e.g., time-domain amplitude characteristics, frequency-domain energy distribution characteristics, wave velocity variation characteristics, and source mechanism-related parameters). Initial weights are assigned to each channel based on the node's noise level, signal-to-noise ratio, and geological conditions. During operation, the model dynamically corrects the channel weights based on real-time quality indicators (e.g., packet loss rate, signal drift, and equipment calibration frequency) fed back by the monitoring nodes. If a node experiences signal distortion or noise enhancement during a specific time period, the weight ratio of that node in the multi-channel mapping is automatically reduced to ensure the stability and authenticity of the overall feature representation. The resulting multi-channel mapping results can simultaneously characterize the multi-dimensional attributes of the seismic signal.

[0058] By iteratively updating the mapping results of each channel and integrating the multi-channel signal features of all nodes to form a preliminary sequence representation, the mapping results of each channel are iteratively optimized to correct time drift and inter-channel deviation. In each iteration, the feature similarity and time alignment error between adjacent time segments and adjacent node channels are calculated. The channel mapping parameters are updated by the weighted minimum mean square error criterion until the overall error converges. The optimized multi-channel signal features of all monitoring nodes are integrated in chronological order and spatial location to form a preliminary sequence representation that reflects the joint temporal characteristics of multiple nodes and multiple channels, which can dynamically describe the response differences of different nodes during vibration propagation.

[0059] S2: Cross-window aggregation of local amplitude anomalies and sudden energy characteristics present in the preliminary sequence representation, and dynamic mapping of potential aftershock signals with mainshock characteristics by constructing a bidirectional correlation network to form a traceable response evolution map.

[0060] The process of cross-window aggregation of local amplitude anomalies and burst energy features present in the preliminary sequence representation in S2 is as follows:

[0061] The algorithm scans the preliminary sequence representation within a sliding time window to identify the temporal location and amplitude changes of local amplitude anomalies and sudden energy features within each time segment. A sliding time window is established based on the time axis of the preliminary sequence representation, with the window length adaptively adjusted according to the signal change rate. When the local signal energy exceeds a set threshold or the amplitude change rate is higher than the average level, the location is marked as a local amplitude anomaly point. The instantaneous energy distribution and peak energy density of each window are calculated in the frequency domain. Regions with sudden energy increases are identified as sudden energy features. To avoid misjudgments caused by noise interference, the algorithm combines signal smoothing and threshold self-adjustment mechanisms to jointly judge the amplitude change trend of continuous windows. Only when the duration of the anomaly exceeds the minimum effective duration interval is the anomaly determined to be a valid detection result.

[0062] Alignment and fusion of local amplitude anomalies and energy burst features detected in adjacent windows are performed to establish a sequence chain reflecting the continuity of amplitude fluctuations and energy release. The time interval and amplitude difference of anomalies in adjacent windows are calculated. If the time interval is less than a set overlap threshold and the amplitude difference is less than the similarity limit, the two points are considered to belong to the same anomaly event. Based on the Euclidean distance temporal matching method, anomalies in adjacent windows are aligned to eliminate the time misalignment caused by different window boundaries. After alignment, the abnormal energy peaks and corresponding amplitude changes of multiple adjacent windows are weighted and fused. The weights are automatically adjusted according to the node data quality and local signal-to-noise ratio, thereby obtaining a continuous amplitude fluctuation trajectory and energy release sequence. This can unify and integrate the anomaly information that was originally scattered in multiple windows into a continuous response chain, eliminate the time window boundary effect, and preserve the natural evolution characteristics of seismic signals in the time dimension.

[0063] Based on the fused amplitude and energy feature distributions, a traceable local signal pattern chain is constructed. Through cross-window aggregation, the anomalous responses of node signals in both spatial and temporal dimensions are unified and integrated to generate a cross-window aggregation result. The amplitude and energy feature distributions of each node within the same time period are extracted. Through time alignment and node index matching, a spatial distribution matrix is ​​constructed to characterize the response differences of different nodes to the same seismic event. In the temporal dimension, adjacent time segments are connected according to the continuity of the signal to generate a traceable local signal pattern chain, which includes amplitude evolution trends, energy release rates, and node spatial correlation. Through the cross-window aggregation mechanism, the signal pattern chains of all nodes are simultaneously integrated in both spatial and temporal dimensions to output a unified cross-window aggregation result.

[0064] The process of forming a traceable response evolution diagram in S2 is as follows:

[0065] A bidirectional correlation network between nodes is generated based on the cross-window aggregation results, quantifying the similarity of signal features and the degree of temporal evolution coupling of each node. The cross-window aggregation results output from the previous steps are obtained, which contain characteristic information such as amplitude changes, energy release, and relative phase differences of multiple earthquake monitoring nodes in continuous time segments. Each monitoring node is regarded as a node unit in the network, and its signal feature vector is used as the node attribute input. The similarity of signal features between nodes is quantified by calculating the correlation coefficient, mutual information, or cosine similarity of feature vectors between any two nodes. A temporal evolution coupling factor is introduced to measure the synchronicity and delay relationship of signal changes between nodes. The coupling factor can be calculated by the delay cross-correlation function, thereby reflecting the temporal dependence between the mainshock propagation and the aftershock response. The feature similarity and temporal coupling information between nodes are simultaneously encoded into the edge weight parameters to form a bidirectional correlation network with bidirectional connection characteristics (i.e., node A can be weighted separately for node B, and node B can be weighted separately for node A).

[0066] Based on the node interaction intensity and edge weight changes in the bidirectional correlation network, statistical correlation information between nodes is extracted and a node correlation matrix is ​​generated to describe the dynamic relationship between the mainshock and potential aftershocks. Statistical analysis is performed on the edge weight distribution of the bidirectional correlation network to extract node interaction intensity indicators, including the average edge weight, rate of change, and temporal continuity. The interaction intensity between each pair of nodes is standardized to form node correlation weight vectors. These weight vectors are then arranged according to the node index to generate a node correlation matrix. The rows and columns of the matrix correspond to the monitoring node numbers, and the elements in the matrix represent the combined value of the signal similarity and evolutionary coupling degree between two nodes within a specific time window. The establishment of the node correlation matrix enables the system to describe the coupling dynamic relationship between the mainshock and potential aftershocks with a unified data structure.

[0067] Potential aftershock signals and mainshock characteristics are projected onto the response evolution diagram using a nonlinear mapping function, and the temporal evolution path and weight distribution are determined based on the node correlation matrix. A nonlinear mapping function (such as Gaussian radial basis mapping based on kernel function or multilayer perceptron mapping model) is selected to project the node signal characteristics into a low-dimensional evolution space. During the mapping process, the connection path between different nodes is determined based on the edge weight strength in the node correlation matrix, and the temporal evolution connectivity between nodes is adjusted according to the weight size. A set of temporal evolution paths is established in the mapping space, and each path corresponds to the influence chain of the mainshock signal on the potential aftershock response. The path length and node weight together reflect the evolution trend and persistence of seismic activity. Through the nonlinear mapping and path reconstruction process, a response evolution prototype containing node attributes, connection edge weights and temporal trajectories is formed.

[0068] In a bidirectional correlation network, potential abnormal fluctuation regions are identified, and perturbation information is fed back to the node correlation matrix to perform weight adjustments. The corrected node correlation relationships and temporal evolution data are integrated to form a traceable response evolution graph. The generated response evolution prototype is dynamically monitored, and the energy fluctuation amplitude and similarity mutation rate of each node and its adjacent edges are calculated. If a node experiences a sharp increase in energy release or its edge weight change rate exceeds a set threshold within a short period of time, the region is identified as a potential abnormal fluctuation region. The perturbation information (including location, time, and amplitude change rate) of the nodes in the abnormal region is fed back to the node correlation matrix. The edge weights between related nodes are adjusted using a weight correction function, thereby dynamically correcting the temporal evolution path. Through multiple rounds of iterative adjustments, the path gradually converges to the true temporal propagation structure. The corrected node correlation relationships and temporal evolution data are integrated to generate a traceable response evolution graph.

[0069] S3: Iteratively fuse the feature differences in the response evolution graph with the node distribution information, and use time-delay correlation analysis and asynchronous matching strategy to identify the non-obvious dependencies between signal changes, generating a digital constraint model of sequence features.

[0070] The process of iteratively fusing the feature differences and node distribution information in the response evolution graph in S3 is as follows:

[0071] The signal characteristic differences of each node in the response evolution diagram are weighted and combined with the node's geographical location and monitoring coverage information. The signal characteristics of each monitoring node are analyzed, including amplitude changes, energy release rate, spectral energy density, and phase shift. By comparing the signal response differences between nodes in the same time period, areas with significant characteristic changes can be identified. These characteristic differences are correlated with information such as the geographical coordinates, altitude, geological conditions, and monitoring equipment coverage of each node. Based on the location density of the nodes in spatial distribution and their representativeness to the surrounding area, different weights are assigned to different nodes to highlight monitoring points with large signal characteristic differences and representative geographical distribution. The weighted combination results form a comprehensive node weight that includes signal difference attributes and geospatial characteristics, which is used to reflect the degree of influence of each node in the overall response evolution diagram.

[0072] The fusion weights of each node are updated through multiple iterations, and the feature correlations and spatial distribution relationships between nodes are calculated synchronously in each iteration. During each iteration, the correlation between nodes in time series is calculated based on the similarity of signal features between nodes, and node pairs with synchronous response trends or energy coupling characteristics are identified. Combined with the spatial distribution relationship of nodes, the distance influence between adjacent nodes is calculated. Nodes that are closer have higher spatial correlation in the calculation. Taking into account both temporal feature correlation and spatial proximity, the node weights are dynamically updated. As the number of iterations increases, the fusion weights of nodes tend to stabilize, and the combined influence of signal features and geographical distribution is fully reflected.

[0073] The features of each node after iterative fusion are integrated to form a fused feature representation; the temporal response features, energy change trends and geographical distribution of the nodes are uniformly mapped into a comprehensive feature space to reflect the relative influence and temporal evolution of each node in the entire seismic response process.

[0074] The process of generating a digital constraint model for sequence features in S3 is as follows:

[0075] The signals of each node in the fused feature representation are paired in pairs according to time series, and the time delay correlation coefficient of the signals is calculated. Potential nonlinear dependencies are identified through delay scanning. The time series signals of each pair of nodes are combined and analyzed. By comparing the amplitude changes, energy distribution and spectral characteristics of the node signals, possible inter-temporal dependencies are identified. In order to capture nonlinear dependencies and potential hysteresis effects, the signals are delayed and the response correlation at different time offsets is examined in turn. By analyzing the similarity changes at different time offsets, potential coupling modes and non-obvious dependencies between nodes are identified. Even if there is hysteresis or fluctuation in the signal, the dynamic correlation between nodes can be reflected.

[0076] For node signals with time misalignment or asynchronous sampling, an asynchronous matching strategy is applied. Phase adjustment and local feature comparison are used to quantify the evolutionary synchronicity and feature differences between nodes. The sampling times of different nodes in the earthquake monitoring system may not be completely synchronized, or there may be short-term data gaps. An asynchronous matching strategy is introduced to correct and compare time series signals. Each pair of node signals is first aligned through phase adjustment to ensure that local features can be matched on the time axis. For each local time segment, the differences in amplitude changes, energy peaks, and short-term spectral characteristics are calculated to quantify the evolutionary synchronicity between nodes and the degree of signal feature differences. Through this strategy, even if there is asynchronous sampling or slight delay between nodes, the dynamic relationship between signals can be effectively captured.

[0077] By integrating the results of delay correlation analysis with asynchronous matching metrics, a digital constraint model for sequence features is formed for hierarchical dimensionality reduction processing. The dependencies and synchronization metrics between nodes are comprehensively integrated to form a unified constraint information structure, which not only reflects the nonlinear coupling relationship between nodes, but also quantifies the temporal consistency and differences of signal evolution, providing a digital representation for sequence features. The integrated constraint model serves as input to guide the hierarchical dimensionality reduction process, enabling selective compression of signal patterns at different granularities and preservation of key features, while ensuring the reasonable expression of sequence features in space and time.

[0078] S4: Based on the digital constraint model, a hierarchical dimensionality reduction processing mechanism is introduced to selectively compress signal patterns of different granularities, and key changes are enhanced through the residual feedback mechanism to form an indicator system that reflects the stability of the sequence and the potential aftershock trend.

[0079] The process of selectively compressing signal modes of different granularities in S4 is as follows:

[0080] In the digital constraint model, node signals and dependencies are divided into different granularities according to hierarchy, and a hierarchical dimensionality reduction mechanism is applied to each layer to perform structured mapping and feature extraction of signal patterns. Based on the feature complexity, dependency strength, and evolution time scale of the node signals, the node signals are divided into several levels, each representing a signal pattern of different granularities. This hierarchical division allows for the separate processing of high-frequency detail signals, low-frequency trend signals, and intermediate mixed patterns to meet the feature extraction needs of different granularities. A hierarchical dimensionality reduction mechanism is introduced in each layer to perform structured mapping of the node signals, mapping the original high-dimensional signals to representative patterns in the feature space, while extracting amplitude changes, local trends, and energy peaks.

[0081] Under the hierarchical dimensionality reduction mechanism, selective compression is performed on the signal patterns of each granularity layer to retain key evolutionary features and reduce redundant information, while generating low-dimensional representations that can be connected between layers. After completing the signal hierarchy division and feature extraction, the signal patterns of each layer are selectively compressed through the hierarchical dimensionality reduction mechanism. The core features reflecting the dynamic evolution of nodes, local anomaly responses, and key trends in each granularity layer are retained, while duplicate information and noise components in the signal are removed, thereby significantly reducing data dimensionality and storage overhead. During the compression process, the continuity and connectivity between different layers are also considered. Through a unified feature mapping and time alignment strategy, it is ensured that the low-dimensional representation can be transmitted and combined across layers.

[0082] By integrating the dimensionality reduction results of each layer, a dimensionality-reduced signal representation with a hierarchical structure and containing multi-granularity signal features is formed. After selective compression of each layer, the low-dimensional signal representations of each layer are integrated. By maintaining the hierarchical relationship and time series order, key signal features of different granularities are uniformly arranged into a multi-level dimensionality-reduced signal representation. This integration process not only preserves the core evolutionary features and anomalous response information of each layer, but also ensures the continuity and comparability between high-level trend signals and low-level local detail signals. The resulting dimensionality-reduced signal representation has a clear hierarchical structure and can simultaneously reflect the multi-granularity features of node signals in spatial, temporal, and feature dimensions.

[0083] The process of forming an indicator system in S4 that reflects sequence stability and potential aftershock trends is as follows:

[0084] The difference between the dimensionality-reduced signal representation and the original sequence is calculated to extract residual information and identify key change points. After obtaining the multi-granularity dimensionality-reduced signal representation, the multi-granularity dimensionality-reduced signal is compared and analyzed with the original waveform sequence. By calculating the differences of signals at each time point and node, residual information is extracted. The residual information mainly includes key features such as local amplitude anomalies, sudden changes in energy, and small trend shifts that are lost or compressed during the dimensionality reduction process. These residuals can reflect the dynamic changes in the original signal that were not fully captured, and help identify key change points that have an important impact on the stability of the sequence and the potential aftershock trend during the time-series evolution.

[0085] By using a residual feedback mechanism, residual information is injected into the dimensionality-reduced signal representation to amplify key changes. After extracting the residual information, it is re-injected into the dimensionality-reduced signal representation through the residual feedback mechanism. The residual feedback mechanism will weight and adjust the key time segments and node signals in the dimensionality-reduced signal, so that the amplitude changes, local energy peaks and trend characteristics that were compressed or weakened in the dimensionality reduction process are amplified. It can dynamically compensate for the information loss caused by dimensionality reduction, so that the signal representation can more accurately reflect the evolution pattern of potential aftershock events, while highlighting important evolutionary change points in the sequence.

[0086] The integrated and enhanced signal representations form an indicator system reflecting sequence stability and potential aftershock trends. After completing residual feedback and key change enhancement, the signal representations of each node and time segment are integrated to form a unified indicator system, including sequence stability assessment indicators and potential aftershock trend characteristic indicators. The sequence stability indicators reflect the evolution continuity, abnormal fluctuation amplitude, and overall stability of the signals at each node, while the potential aftershock trend indicators reflect the possible sudden aftershock signal patterns and their evolution trends. The integrated indicator system can not only intuitively assess the dynamic stability of the earthquake sequence, but also provide systematic and quantitative reference data for potential aftershock prediction and risk warning.

[0087] S5: Utilize an indicator system to dynamically compare and weight the sequence features of different time segments, and calculate the information retention and feature recognition rate.

[0088] The process of calculating information preservation and feature recognition rate in S5 is as follows:

[0089] The indicator system dataset is dynamically compared and weighted by time segments to calculate the information retention rate and feature recognition rate of each segment. The indicator system dataset formed by residual feedback and dimensionality reduction is divided into preset time segments. Each time segment contains multiple node signals and their evolution characteristics. For each time segment, the signal information before and after dimensionality reduction and residual enhancement is compared to analyze the amount of key information retained, such as amplitude changes, energy peaks, and trend characteristics. The dataset is then weighted and integrated according to the importance of each feature. Through this dynamic comparison and weighting process, the degree of information retention of each time segment during dimensionality reduction, integration, and enhancement can be quantified, thus forming an information retention rate index. The feature recognition rate index is generated by comparing the recognition ability of the original signal and the features of each time segment in the indicator system to evaluate the feature fidelity and recognizability after dimensionality reduction and integration.

[0090] By comparing the trend of feature changes between different time segments, an evaluation result reflecting the preservation of sequence information and the ability to identify features is generated. After calculating the information preservation rate and feature recognition rate for each time segment, the trend of feature changes between different time segments is analyzed. By comparing the changes in amplitude, energy, waveform morphology and temporal evolution of key features in each segment, potential abnormal fluctuations and trend shifts in the sequence are identified. These trends are integrated to generate a comprehensive evaluation result reflecting the preservation of information and the ability to identify features of the entire sequence. This result can intuitively reveal the information integrity, key feature retention and identification effect of the sequence in temporal evolution.

[0091] Example 2: As Figure 2 As shown, a deep dimensionality reduction simulation system for fusing time-frequency features of mainshock and aftershock sequences includes:

[0092] Adaptive segmentation module: Performs non-fixed window adaptive segmentation on seismic monitoring data based on waveform change characteristics to generate preliminary sequence representation;

[0093] Association and aggregation module: Aggregates local amplitude anomalies and energy mutation characteristics to construct a dynamic management mapping between the mainshock and aftershocks;

[0094] Feature fusion module: fuses response evolution features with node distribution information to identify implicit dependencies between signal changes;

[0095] Hierarchical dimensionality reduction module: Performs selective dimensionality reduction and residual enhancement on multi-level signal patterns to extract stability and trend indicators;

[0096] Dynamic evaluation module: Performs dynamic comparison and weighted analysis on sequence features of different time segments, and calculates information retention and feature recognition rate.

[0097] 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 process, method, article, or apparatus.

[0098] 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 deep dimensionality reduction simulation method for fusing time-frequency features of mainshock and aftershock sequences, characterized in that, Includes the following steps: The raw waveform data from multiple seismic monitoring nodes are adaptively segmented using non-fixed windows. The transient changes and local trends of each signal segment are recorded in a multi-channel mapping form to generate a preliminary sequence representation that can be dynamically adjusted. Local amplitude anomalies and sudden energy characteristics present in the preliminary sequence representation are aggregated across windows. By constructing a bidirectional correlation network, potential aftershock signals are dynamically mapped with mainshock characteristics to form a traceable response evolution map. The cross-window aggregation results are obtained, which include the amplitude changes, energy release and relative phase difference of multiple earthquake monitoring nodes in a continuous time segment. Each monitoring node is regarded as a node unit in the network. The signal feature vector of the node unit is used as the node attribute input. The similarity of signal features between nodes is quantified by calculating the correlation coefficient, mutual information or cosine similarity between feature vectors of any two nodes. A time evolution coupling factor is introduced to measure the synchronicity and delay relationship of signal changes between nodes. The coupling factor can be calculated by the delay cross-correlation function, thereby reflecting the temporal dependence between the mainshock propagation and the aftershock response. The feature similarity and time coupling information between nodes are simultaneously encoded into the edge weight parameters to form a bidirectional correlation network with bidirectional connection characteristics. The feature differences in the response evolution graph are iteratively fused with the node distribution information. The non-obvious dependencies between signal changes are identified by using time-delay correlation analysis and asynchronous matching strategy, and a digital constraint model of sequence features is generated. Based on the digital constraint model, a hierarchical dimensionality reduction processing mechanism is introduced to selectively compress signal patterns of different granularities, and key changes are enhanced through the residual feedback mechanism to form an index system that reflects the stability of the sequence and the potential aftershock trend. The indicator system is used to dynamically compare and weight the sequence features of different time segments, and to calculate the information retention and feature recognition rate.

2. The deep dimensionality reduction simulation method for fusing time-frequency features of mainshock and aftershock sequences according to claim 1, characterized in that, The process of generating a preliminary sequence representation that can be dynamically adjusted is as follows: The raw waveform data of each monitoring node is dynamically segmented and collected according to a non-fixed time window, and transient changes and local trend features are extracted in each signal segment. Multi-channel mapping processing is performed on each signal segment, and the feature weights of each channel can be dynamically adjusted according to changes in node status and data quality. By iteratively updating the mapping results of each channel, the multi-channel signal features of all nodes are integrated to form a preliminary sequence representation.

3. The deep dimensionality reduction simulation method for fusing time-frequency features of mainshock and aftershock sequences according to claim 2, characterized in that, The process of cross-window aggregation of local amplitude anomalies and burst energy features present in the preliminary sequence representation is as follows: The preliminary sequence representation is scanned within a sliding time window to identify the temporal location and amplitude changes of local amplitude anomalies and sudden energy characteristics within each time segment; Align and fuse the local amplitude anomalies and energy burst features detected in adjacent windows to establish a sequence chain that reflects the continuity of amplitude fluctuations and energy release; Based on the fused amplitude and energy characteristic distribution, a traceable local signal pattern chain is constructed, and through cross-window aggregation, the abnormal responses of node signals in the spatial and temporal dimensions are uniformly integrated to generate cross-window aggregation results.

4. The deep dimensionality reduction simulation method for fusing time-frequency features of mainshock and aftershock sequences according to claim 3, characterized in that, The process of forming a traceable response evolution map is as follows: A bidirectional correlation network between nodes is generated based on the cross-window aggregation results, and the similarity of signal features and the degree of temporal evolution coupling of each node are quantified. Based on the interaction intensity and edge weight changes of nodes in the bidirectional correlation network, statistical correlation information between nodes is extracted and a node correlation matrix is ​​generated to describe the dynamic relationship between the main shock and potential aftershocks. Potential aftershock signals and mainshock characteristics are projected onto the response evolution diagram through a nonlinear mapping function, and the time evolution path and weight distribution are determined based on the node correlation matrix. Potential abnormal fluctuation regions are identified in the bidirectional correlation network, and the disturbance information is fed back to the node correlation matrix to perform weight adjustment. The corrected node correlation relationship and time series evolution data are integrated to form a traceable response evolution diagram.

5. The deep dimensionality reduction simulation method for fusing time-frequency features of mainshock and aftershock sequences according to claim 4, characterized in that, The process of iteratively fusing the feature differences and node distribution information in the response evolution graph is as follows: The signal characteristic differences of each node in the response evolution diagram are weighted and combined with the node's geographical location and monitoring coverage information; The fusion weights of each node are updated through multiple iterations, and the feature associations and spatial distribution relationships between nodes are calculated synchronously in each iteration. The features of each node after iterative fusion are integrated to form a fused feature representation.

6. The deep dimensionality reduction simulation method for fusing time-frequency features of mainshock and aftershock sequences according to claim 5, characterized in that, The process of generating a digital constraint model for sequence features is as follows: The signals of each node in the fused feature representation are paired in pairs according to the time series, the delay correlation coefficient of the signals is calculated, and potential nonlinear dependencies are identified by delay scanning; For node signals with time misalignment or asynchronous sampling, an asynchronous matching strategy is applied to quantify the evolution synchronization and feature differences between nodes through phase adjustment and local feature comparison. By integrating the results of delay correlation analysis with asynchronous matching metrics, a digital constraint model for sequence features is formed for hierarchical dimensionality reduction processing.

7. The deep dimensionality reduction simulation method for fusing time-frequency features of mainshock and aftershock sequences according to claim 6, characterized in that, The process of selectively compressing signal modes of different granularities is as follows: The node signals and dependencies in the digital constraint model are divided into different granularities according to the hierarchy, and a hierarchical dimensionality reduction processing mechanism is applied to each layer to perform structured mapping and feature extraction of the signal patterns. Under the hierarchical dimensionality reduction processing mechanism, selective compression is performed on the signal patterns of each granularity layer to retain key evolutionary features and reduce redundant information, while generating low-dimensional representations that can be connected between layers. By integrating the dimensionality reduction results from each layer, a dimensionality-reduced signal representation with a hierarchical structure and containing multi-granularity signal features is formed.

8. The deep dimensionality reduction simulation method for fusing time-frequency features of mainshock and aftershock sequences according to claim 7, characterized in that, The process of developing an indicator system that reflects sequence stability and potential aftershock trends is as follows: The difference between the dimensionality-reduced signal representation and the original sequence is calculated, and residual information is extracted to identify key change points. By using a residual feedback mechanism, residual information is injected into the dimension-reduced signal representation to amplify key changes; The integrated and enhanced signal representation forms an indicator system that reflects sequence stability and potential aftershock trends.

9. The deep dimensionality reduction simulation method for fusing time-frequency features of mainshock and aftershock sequences according to claim 8, characterized in that, The process of calculating information preservation and feature recognition rate is as follows: The indicator system dataset is dynamically compared and weighted by time segment, and the information retention rate and feature recognition rate of each segment are calculated. By comparing the trend of feature changes across different time segments, an evaluation result reflecting the preservation of sequence information and the ability to identify features is generated.

10. A deep dimensionality reduction simulation system for fusing time-frequency features of mainshock and aftershock sequences, applied to the method described in any one of claims 1-9, characterized in that, include: Adaptive segmentation module: Performs non-fixed window adaptive segmentation on seismic monitoring data based on waveform change characteristics to generate preliminary sequence representation; Association and aggregation module: Aggregates local amplitude anomalies and energy mutation characteristics to construct a dynamic management mapping between the mainshock and aftershocks; Feature fusion module: fuses response evolution features with node distribution information to identify implicit dependencies between signal changes; Hierarchical dimensionality reduction module: Performs selective dimensionality reduction and residual enhancement on multi-level signal patterns to extract stability and trend indicators; Dynamic evaluation module: Performs dynamic comparison and weighted analysis on sequence features of different time segments, and calculates information retention and feature recognition rate.

Images