Method and system for monitoring vibration of underground facilities based on distributed sensor network

Abstract

The present invention discloses a vibration monitoring method and system for underground facilities based on a distributed sensor network, including: obtaining initial data of a distributed sensor network; constructing an initial network topology structure; collecting sensor node status data; performing dynamic topology adjustment; evaluating and optimizing network performance; obtaining an optimized network topology structure; receiving raw data collected by sensors and using the network topology structure to clean and standardize data; using a filtering algorithm to remove noise, and analyzing and extracting features; generating a preprocessed feature data set. A multi-source data fusion algorithm module is used to integrate information; an anomaly prediction model is constructed and parameters are optimized, and an anomaly detection is performed using the optimized anomaly prediction model; fused data and anomaly detection results are output; and for the anomaly detection results, whether to trigger an early warning is determined according to a preset threshold. Efficient and reliable real-time vibration monitoring is achieved, and the accuracy of anomaly detection and system adaptability are improved.

Description

Technical Field

[0001] The invention relates to a vibration monitoring method, in particular to an underground facility vibration monitoring method and system based on a distributed sensor network. Background Art

[0002] As an important part of modern urban infrastructure, the safety and stability of underground facilities are directly related to urban operations and the quality of life of residents. With the acceleration of urbanization and the high development and utilization of underground space, the risks and challenges faced by underground facilities are increasing. Vibration monitoring of underground facilities is of great significance for timely detection of potential risks, prevention of accidents, and extension of the service life of facilities. Traditional manual inspection methods can no longer meet the needs of large-scale, high-precision, and real-time monitoring.

[0003] At present, research on vibration monitoring of underground facilities based on distributed sensor networks has made certain progress. In terms of sensor technology, high-sensitivity, low-power MEMS accelerometers and optical fiber sensors have been developed to meet the needs of long-term and stable vibration monitoring. In terms of network communication, the application of low-power wide area network (LPWAN) technologies such as LoRa and NB-IoT solves the problem of long-distance and low-power communication in underground environments. In terms of data processing algorithms, time-frequency analysis, machine learning and other methods are widely used in feature extraction and anomaly detection of vibration signals. In addition, the introduction of edge computing technology effectively reduces the amount of data transmission and improves the real-time response capability of the system. Some studies have also explored blockchain-based data security storage solutions to ensure the integrity and traceability of monitoring data. The comprehensive application of these technologies has laid the foundation for the actual deployment and operation of underground facility vibration monitoring systems.

[0004] Although significant progress has been made in the research, there are still some technical problems that need to be solved urgently, including: underground facility vibration signals often show nonlinear and non-stationary characteristics, and existing feature extraction methods are difficult to effectively capture these complex patterns, resulting in insufficient accuracy and reliability of anomaly detection. In addition, the data synchronization and timestamp alignment problems in large-scale distributed sensor networks have not been well solved, which directly affects the accuracy and credibility of multi-point collaborative monitoring. At the same time, the existing network topology lacks adaptive capabilities and is difficult to cope with dynamic changes such as node failure or the addition of new nodes, which affects the robustness and scalability of the system.

[0005] Therefore, research and innovation are needed to improve the performance and practicality of underground facility vibration monitoring systems and provide more reliable technical support for urban safety management. Summary of the invention

[0006] The purpose of the invention is to provide a method and system for underground facility vibration monitoring based on a distributed sensor network, in order to solve the above-mentioned problems existing in the prior art.

[0007] Technical solution, according to one aspect of the present application, a method for monitoring vibration of underground facilities based on a distributed sensor network comprises the following steps:

[0008] Step S1, obtaining initial data of a distributed sensor network; constructing an initial network topology structure; collecting sensor node status data; performing dynamic topology adjustment; evaluating and optimizing network performance; and obtaining an optimized network topology structure;

[0009] Step S2, receiving the raw data collected by the sensor and using the network topology structure to clean and standardize the data; using a filtering algorithm to remove noise, and analyzing and extracting features; generating a preprocessed feature data set.

[0010] Step S3, input the preprocessed feature data set; use the multi-source data fusion algorithm module to integrate the information; build an anomaly prediction model and optimize the parameters, use the optimized anomaly prediction model to perform anomaly detection; output the fused data and anomaly detection results;

[0011] Step S4: for the abnormal detection results, determine whether to trigger an early warning according to the preset threshold; generate early warning information; execute early warning response measures; output monitoring report and early warning status.

[0012] According to another aspect of the present application, there is also provided an underground facility vibration monitoring system based on a distributed sensor network, comprising:

[0013] at least one processor; and,

[0014] a memory communicatively connected to at least one of the processors; wherein,

[0015] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the underground facility vibration monitoring method based on a distributed sensor network described in any of the above technical solutions.

[0016] Beneficial effect: efficient and reliable real-time vibration monitoring is achieved, and the accuracy of abnormality detection and system adaptability are improved. The relevant technical effects will be described in detail below in conjunction with the case. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the present invention.

[0018] Figure 2 It is a flow chart of step S1 of the present invention.

[0019] Figure 3 It is a flow chart of step S2 of the present invention.

[0020] Figure 4It is a flow chart of step S3 of the present invention.

[0021] Figure 5 It is a flow chart of step S4 of the present invention. DETAILED DESCRIPTION

[0022] like Figure 1 As shown, the underground facility vibration monitoring method based on the distributed sensor network includes the following steps:

[0023] Step S1, obtaining initial data of a distributed sensor network; constructing an initial network topology structure; collecting sensor node status data; performing dynamic topology adjustment; evaluating and optimizing network performance; and obtaining an optimized network topology structure;

[0024] Step S2, receiving the raw data collected by the sensor and using the network topology structure to clean and standardize the data; using a filtering algorithm to remove noise, and analyzing and extracting features; generating a preprocessed feature data set.

[0025] Step S3, input the preprocessed feature data set; use the multi-source data fusion algorithm module to integrate the information; build an anomaly prediction model and optimize the parameters, use the optimized anomaly prediction model to perform anomaly detection; output the fused data and anomaly detection results;

[0026] Step S4: for the abnormal detection results, determine whether to trigger an early warning according to the preset threshold; generate early warning information; execute early warning response measures; output monitoring report and early warning status.

[0027] First, the dynamic network topology optimization technology effectively copes with the impact of the complex underground environment on wireless communication, ensuring the reliable collection and transmission of monitoring data. Adaptive sampling and multi-scale feature extraction technology not only improves data quality, but also significantly reduces the transmission volume, which is particularly important for bandwidth-limited underground wireless networks. Multi-source data fusion and multi-scale anomaly detection methods significantly improve the system's ability to recognize complex vibration patterns, and can simultaneously capture microstructural anomalies and macroscopic systemic problems. Methods based on graph theory and space-time correlation analysis enhance the accuracy and interpretability of anomaly detection, providing reliable decision support for underground facility management. The introduction of dynamic thresholds and multi-level early warning strategies enables the system to flexibly respond to long-term changes and emergencies in the underground environment, improving the timeliness and reliability of early warnings. Resource demand prediction and adaptive network reconstruction technology ensure that the system maintains efficient operation in a complex and changing underground environment.

[0028] This comprehensive approach is particularly suitable for the special needs of vibration monitoring in underground facilities. It can effectively deal with challenges such as signal attenuation, electromagnetic interference and spatial limitations in underground environments, ensuring the continuity and reliability of monitoring data. The system's increased sensitivity to small vibration changes helps to detect potential structural problems early. The multi-scale analysis capability enables it to focus on both local anomalies (such as problems with a single support column) and systemic risks (such as instability of an entire tunnel section). The space-time correlation analysis effectively distinguishes between real vibration anomalies and environmental interference, reducing the false alarm rate. The adaptive capability enables the system to cope with the time-varying vibration characteristics of underground facilities, such as regular vibrations during subway operation hours.

[0029] In addition, the method has good scalability and flexibility. It can adapt to different types and sizes of underground facilities, from single tunnels to complex subway networks. Resource optimization and network reconstruction technology enable the system to maximize node service life and reduce maintenance costs while ensuring monitoring quality. This is especially important for long-term, large-scale underground facility monitoring projects.

[0030] According to one aspect of the present application, step S1 specifically comprises:

[0031] Step S11, receiving and constructing an initial network diagram based on the location coordinates and initial connection relationship data of the sensor network, wherein nodes represent sensors and edges represent connections between sensors; within a preset time window, collecting signal strength data of each sensor node; measuring the communication delay between adjacent nodes and recording the round-trip time; integrating the collected signal strength and communication delay data into a state matrix; wherein the matrix elements are used to represent the connection quality between node pairs; providing a basis for subsequent time synchronization.

[0032] Step 12: read the generated state matrix and initial network diagram, remove the connections in the state matrix that are below the set connection quality threshold, and update the network diagram; for isolated nodes formed by connection removal, apply the Delaunay triangulation algorithm to re-establish connections, while minimizing long-distance connections while maintaining network coverage; use the Kruskal minimum spanning tree algorithm to ensure the connectivity of the entire network; after completing the dynamic topology adjustment, obtain an updated network diagram; ensure that the network can be automatically reconstructed when a node fails.

[0033] Step S13, read the updated network diagram, calculate the network performance indicators and form a performance indicator set; call the multi-objective optimization function, evaluate the quality of the current network topology based on the performance indicator set, and output the optimization score and specific optimization suggestions; if the optimization score is lower than the preset threshold, make further adjustments according to the optimization suggestions; if it is not lower than the preset threshold, output the optimized network topology structure; where the network performance indicators include network connectivity, average node degree, network diameter, energy efficiency and load balancing. The continuous optimization mechanism enables the network to adapt to dynamic changes.

[0034] In this embodiment, by dynamically constructing and optimizing the network topology, the adaptability and reliability of the underground facility vibration monitoring system are significantly improved. Specifically, first, by constructing the initial network diagram and collecting real-time status data, the foundation is laid for subsequent optimization. Using signal strength and communication delay as connection quality indicators can accurately reflect the impact of the underground complex environment on wireless communication. Delaunay triangulation and Kruskal minimum spanning tree algorithm are used in the dynamic topology adjustment mechanism to minimize long-distance connections while ensuring network coverage, effectively reducing signal attenuation and energy consumption in the underground environment. The multi-objective optimization function comprehensively considers key indicators such as network connectivity, average node degree, network diameter, energy efficiency and load balancing, so that the optimized network structure can maintain stable and efficient data transmission in a complex underground environment. This adaptive network topology is particularly important for underground facility vibration monitoring, because obstacles, humidity changes and electromagnetic interference in the underground environment will continuously affect network performance. Through a continuous optimization process, the system can respond to these challenges in real time and ensure the reliable collection and transmission of monitoring data. In addition, the method also has good scalability, can easily adapt to the expansion or reduction of the monitoring range, and provides technical support for large-scale underground facility vibration monitoring.

[0035] It should be noted that in this embodiment, an optimized network structure is established from the beginning, rather than simply using a preset static configuration. This lays a good foundation for subsequent monitoring work. The system is able to adjust the network structure according to actual underground environmental conditions (such as signal propagation characteristics, obstacle distribution, etc.) instead of relying on a preset model that may be inaccurate. While ensuring network coverage, the communication overhead is minimized, and the energy efficiency and data transmission efficiency of the network are improved from the beginning. By "applying the Delaunay triangulation algorithm to reestablish connections for isolated nodes formed by connection removal", the system can identify and resolve potential network vulnerabilities in the initial stage, thereby improving overall reliability. The initial reconstruction process can help discover problems in deployment, such as improper sensor location and areas with severe signal interference. Early detection of these problems can make adjustments before the monitoring task officially begins, avoiding possible data loss or quality degradation in the later stage. The optimized network structure can ensure better data collection quality. For example, by optimizing the connection between nodes, packet loss and delay in data transmission can be reduced, thereby improving the integrity and real-time performance of monitoring data.

[0036] According to one aspect of the present application, step S2 further comprises:

[0037] Step S21, receiving the optimized network topology structure, extracting sub-networks, using the distributed Bellman-Ford algorithm to calculate the shortest path for each sub-network, selecting a central node as a data aggregation point, and constructing a data collection tree Ti with the central node as the root; passing the structural information of the data collection tree Ti to all nodes of each sub-network; this structure helps to achieve orderly collection and synchronization of data.

[0038] Step S22: On each sensor node, the structural information of the data collection tree is received, and the signal change rate is continuously calculated through an adaptive sampling algorithm. When the signal change rate is greater than a predetermined threshold, the sampling frequency is increased to a maximum value; otherwise, the sampling frequency is gradually reduced to a minimum value; and the original vibration signal is collected and stored according to the current sampling frequency;

[0039] Step S23, receiving the original vibration signal, using Daubechies wavelet as the basis function, applying wavelet packet transform to the original vibration signal, and obtaining a set of wavelet packet coefficients; wherein the decomposition level J of the wavelet packet transform is determined according to the highest frequency fmax and the lowest frequency fmin of the signal: J = log2(fmax / fmin);

[0040] Step S24, obtaining a set of wavelet packet coefficients, calculating a statistical feature vector for each frequency band, and using an improved local binary pattern to extract time-frequency domain texture features; combining the feature vector and the time-frequency domain texture features to form a multi-scale feature descriptor;

[0041] Step S25, receiving a multi-scale feature descriptor, reducing the dimension of the multi-scale feature descriptor, retaining the minimum principal components required to explain 95% of the variance, and obtaining a reduced-dimensional multi-scale feature descriptor; applying dynamic Huffman coding to compress the reduced-dimensional multi-scale feature descriptor, and dynamically updating the coding tree according to real-time data statistics to obtain reduced-dimensional compressed data;

[0042] Step S26. Receive the dimension reduction compressed data and the structural information of the data collection tree. Through the data collection tree, the dimension reduction compressed data is transmitted from the leaf node to the central node; using the jump data aggregation technology, the data from different child nodes are partially aggregated at each intermediate node to obtain aggregated data, and the dimension reduction feature matrix is ​​constructed based on the aggregated data, which is the preprocessed feature data set. Partial synchronization and alignment can be achieved during the data transmission process.

[0043] In this embodiment, the distributed Bellman-Ford algorithm is used to construct the optimal data collection tree, ensuring efficient data aggregation in a complex underground network environment. The adaptive sampling algorithm dynamically adjusts the sampling frequency according to the signal change rate, which can not only capture sudden vibration events, but also reduce energy consumption under normal conditions and extend the service life of sensor nodes. The application of wavelet packet transform enables the system to analyze vibration signals in different frequency bands and effectively identify various vibration sources in underground facilities, such as subways, heavy machinery or geological activities. The improved local binary pattern extracts time-frequency domain texture features, enhances the system's sensitivity to small vibration changes, and helps to detect structural anomalies of underground facilities as early as possible. The construction and dimensionality reduction compression of multi-scale feature descriptors greatly reduces the amount of data transmission while retaining key information, which is particularly important for underground wireless networks with limited bandwidth. The use of dynamic Huffman coding further optimizes the data compression effect and adapts to the dynamic changes of underground vibration modes. The jump data aggregation technology performs partial aggregation during data transmission, which not only reduces the network load but also improves the real-time performance of data. The comprehensive application of these technologies enables the system to achieve efficient and accurate vibration data collection and transmission in complex underground environments, providing a high-quality data foundation for subsequent abnormal detection and early warning. At the same time, this method also has good adaptability and can be flexibly adjusted according to different types of underground facilities and vibration characteristics to meet diverse monitoring needs.

[0044] In this embodiment, the adaptive sampling algorithm optimization process in step S22 is specifically as follows:

[0045] The signal change rate calculation method based on sliding window is used, combined with exponential backoff algorithm to adjust the sampling frequency. Specifically, the standard deviation of the signal is calculated in a fixed-size time window and used as a measure of the signal change rate. When the change rate exceeds the threshold, the sampling frequency is increased; when the change rate continues to be lower than the threshold, the sampling frequency is gradually reduced.

[0046] Let w be the sliding window size, x(t) be the signal value at time t, and f(t) be the sampling frequency at time t.

[0047] Calculate the signal standard deviation σ(t) within the sliding window:

[0048] σ(t) = sqrt(1 / w * Σ(x(i) - μ)²), i ∈ [t-w+1, t]; where μ is the signal average within the window.

[0049] Define the signal change rate threshold θ.

[0050] Sampling frequency adjustment rule: If σ(t) > θ, then f(t+1) = min(f_max, 2 * f(t)); if σ(t) ≤ θ, then f(t+1) = max(f_min, f(t) / (1 + α)); where f_max and f_min are the maximum and minimum sampling frequencies, respectively, and α is a small positive number (such as 0.1) used to control the speed of frequency reduction.

[0051] In this embodiment, the optimization process of the wavelet packet transform decomposition level J in S23 is as follows: in order to handle the situation where log2(fmax / fmin) may not be an integer, a method of rounding down and setting a minimum decomposition level is adopted. The maximum decomposition level needs to be limited by considering the bandwidth of the signal and the desired frequency resolution.

[0052] Let fmax be the highest frequency of the signal, fmin be the lowest frequency, and Δf be the minimum desired frequency resolution. J = max(Jmin, min(Jmax, floor(log2(fmax / fmin)))); where: Jmin = 3 (to ensure at least three levels of resolution); Jmax = floor(log2(fmax / Δf)).

[0053] In this embodiment, the optimization process of S25 dynamic Huffman coding is specifically as follows:

[0054] Adopt an adaptive update strategy to dynamically adjust the update frequency of the Huffman tree according to the speed of data change. Specifically, track the coding efficiency and trigger an update when the efficiency drops beyond a certain threshold.

[0055] Let L(t) be the average code length at time t, and N be the number of samples used to calculate the average length.

[0056] Calculate the average code length: L(t) = (1 / N) * Σ l(i), i ∈ [t-N+1, t]; where l(i) is the code length of the i-th sample. Define the efficiency drop threshold β (such as 10%); Update rule: If L(t) > (1 + β)* L(tN), update the Huffman tree; When updating the Huffman tree, use the frequency statistics of the past M samples, where M >N.

[0057] In this embodiment, the process of determining the preset threshold specifically includes:

[0058] Use a moving percentile approach and fine-tune based on the system's false positive and false negative rates.

[0059] Let X be the sequence of observations and p be the desired anomaly detection percentile (e.g. 99%).

[0060] Calculate the moving percentile: θ(t) = percentile(X[tw:t], p); where w is the moving window size. Introduce the adjustment factor λ: θ_adjusted(t) = λ * θ(t); the initial value of λ can be set to 1, and then adjusted according to the false alarm rate and false negative rate: if the false alarm rate > the target false alarm rate, then λ = λ * (1 + δ); if the false negative rate > the target false negative rate, then λ = λ * (1 - δ); where δ is a small positive number (such as 0.05).

[0061] According to one aspect of the present application, step S3 further comprises:

[0062] Step S31, obtain the dimension reduction feature matrix, apply the Dempster-Shafer evidence theory to perform multi-source data fusion, the fusion process is specifically as follows: according to the historical reliability and current data quality of each sensor, assign a basic probability distribution function to each data source; use the Dempster combination rule to fuse the evidence of multiple data sources, calculate the comprehensive belief function and likelihood function; based on the fused belief function, construct a comprehensive feature matrix;

[0063] Step S32, read the comprehensive feature matrix, use the fuzzy c-means clustering algorithm to perform cluster analysis on the fused data, and determine the optimal number of clusters and cluster centers through an iterative optimization process, and identify the main patterns and groups in the data; based on the clustering results, construct a Gaussian mixture model; for each cluster, estimate its mean vector and covariance matrix to form a weighted combination of multiple Gaussian distributions; optimize the GMM parameters through the expectation maximization algorithm; and obtain a GMM model that can describe the distribution characteristics of normal data;

[0064] Step S33, perform multi-scale anomaly detection through the GMM model or the pre-built graph analysis model. When the GMM model is used, the details are as follows: at the single node level, calculate the log-likelihood value of each data point, and mark the points that deviate significantly from the expected distribution as potential anomalies; at the local cluster level, apply the local anomaly factor algorithm, and consider the local density of the data points to identify samples that are abnormal relative to neighboring points. At the global network level, use the graph convolutional network anomaly detector to capture abnormal behaviors that violate the global pattern; combine abnormal evidence from different scales based on the Dempster-Shafer theory to resolve conflicts in the detection results; use the Moran's I index to evaluate the spatial autocorrelation of anomalies, perform spatial correlation analysis, and distinguish between local anomalies and systematic anomalies; output multi-dimensional anomaly detection results, including spatiotemporal feature descriptions of abnormal data points, abnormal regions, and abnormal patterns.

[0065] In this embodiment, the application of Dempster-Shafer evidence theory enables the system to effectively integrate data from different sensors, taking into account the historical reliability and current data quality of each sensor, thereby providing more reliable comprehensive information in complex underground environments. The combination of fuzzy c-means clustering algorithm and Gaussian mixture model enables the system to adaptively identify the main patterns and groups of underground vibration data, which is crucial for distinguishing normal underground activities (such as subway operation) from potential abnormal vibrations. The adoption of multi-scale anomaly detection methods, including analysis at the single node level, local cluster level, and global network level, enables the system to comprehensively capture abnormal behaviors at different scales. This multi-level detection strategy is particularly suitable for vibration monitoring of underground facilities because it can simultaneously identify local structural anomalies (such as vibration anomalies of a single support column) and systemic problems (such as abnormal vibration patterns of an entire tunnel section). The introduction of graph convolutional network anomaly detectors enables the system to effectively utilize the spatial correlation of vibration propagation and improves the ability to identify complex vibration patterns. The application of Moran's I index further enhances the system's ability to analyze spatial autocorrelation and helps to distinguish local anomalies from systemic anomalies, which is of great significance for accurately locating problem areas in underground facilities. In addition, when using the pre-built graph analysis model, the system is able to better capture the spatiotemporal characteristics of vibration data, improving the ability to identify long-term trends and emergencies. This comprehensive anomaly detection method not only improves detection accuracy but also reduces the false alarm rate, which is of great value for the safety monitoring and preventive maintenance of underground facilities.

[0066] According to one aspect of the present application, in step S33, a pre-built graph analysis model is used to perform multi-scale anomaly detection, further comprising:

[0067] Step S331, read the dimension reduction feature matrix, construct a multi-layer graph structure, use a multi-scale graph clustering algorithm to fuse information of different scales, and obtain a low-dimensional embedding space; in the low-dimensional embedding space, use an adaptive density peak clustering algorithm to fuse data and construct a hierarchical data representation; construct a graph neural network model and optimize parameters, update the node feature representation through a message passing mechanism, and obtain the final node feature representation set;

[0068] Step S332, read the node feature representation set and construct a graph structure, use graph theory to analyze the graph structure, calculate the Laplace matrix of the graph and solve its eigenvalue problem; call the graph-based anomaly score function, calculate the anomaly score of each node and normalize it; use the improved box plot method to calculate the local anomaly threshold, perform space-time correlation analysis, and output multi-dimensional anomaly detection results.

[0069] According to one aspect of the present application, step S331 is further:

[0070] Step S3311, read the dimension reduction feature matrix, and build a multi-layer graph structure based on the dimension reduction feature matrix, where each layer represents a different scale of the data; within each scale, calculate the relationship strength between nodes, and use cosine similarity to measure the similarity of node features; establish connections between different layers, and the connection strength depends on the similarity of node features at different scales; and obtain a multi-dimensional graph structure containing multi-scale topological information and node features;

[0071] Step S3312: read the multi-dimensional graph structure, construct a graph Laplacian matrix that comprehensively considers all scales, solve the eigenvalues ​​of the graph Laplacian matrix, select the eigenvectors corresponding to the smallest N non-zero eigenvalues, and each eigenvector forms a low-dimensional embedding space;

[0072] Step S3313: in the low-dimensional space, calculate the local density of each data point and its shortest distance to a point with higher density, select the cluster center through a dynamic threshold function, implement adaptive density peak clustering, and obtain a set of data clustering results;

[0073] Step S3314: construct a hierarchical data representation based on the clustering results, specifically: for each cluster, construct and calculate a feature vector that integrates the features of all nodes in the cluster and takes into account the importance of the nodes, merge similar clusters to form a multi-level structure, and obtain a hierarchical feature set;

[0074] Step S3315: call the pre-built graph neural network model containing multiple graph convolutional layers, with a multi-layer graph structure and hierarchical data representation as input; the calculation of each layer takes into account the characteristics of the node itself and the information of neighboring nodes; the node feature representation is updated through the message passing mechanism to obtain the final node feature representation set.

[0075] By constructing and analyzing a multi-layer graph structure, the system's ability to understand and detect anomalies in complex vibration patterns of underground facilities is greatly enhanced. The construction of the multi-layer graph structure takes into account the different scales of the data, allowing the system to capture vibration characteristics at both the micro and macro levels. This is particularly important for underground facilities, as different structural components may exhibit vibration anomalies at different scales. The application of cosine similarity ensures the accurate calculation of the strength of the relationship between nodes, effectively reflecting the similarity of vibration characteristics at different locations in the underground structure. The construction of the graph Laplacian matrix and eigenvalue analysis enable the system to extract the most representative low-dimensional features, effectively reducing the complexity of the data while retaining key information. This is of great significance for processing massive vibration data of large-scale underground facilities, and can improve computational efficiency while ensuring the quality of analysis. The application of the adaptive density peak clustering algorithm enables the system to automatically identify natural clusters in the data, which helps to discover vibration mode groups in underground facilities, such as distinguishing vibrations generated by normal operations from potential structural problems. The construction of the hierarchical data representation takes into account the importance of the nodes, allowing the system to better capture the vibration characteristics of key locations in underground facilities. The introduction of the graph neural network model effectively integrates local and global vibration information through a message passing mechanism, enhancing the system's ability to model complex spatial relationships. This is crucial for understanding the propagation pattern of underground vibration and identifying the source of abnormalities. Overall, this method based on a multi-layer graph structure and graph neural network greatly improves the system's ability to analyze complex vibration patterns of underground facilities, can more accurately identify potential structural problems and safety hazards, and provides strong technical support for preventive maintenance and safety management of underground facilities.

[0076] According to one aspect of the present application, step S332 is further:

[0077] Step S3321: Based on the node feature representation, a new graph structure is constructed, wherein the edge weights in the graph structure represent the similarity of the node features; based on the graph structure, an adjacency matrix and a degree matrix are calculated to obtain a Laplacian matrix and normalize it;

[0078] Step S3322, performing eigenvalue decomposition on the normalized Laplace matrix to obtain several columns of eigenvalues ​​and corresponding eigenvectors, and extracting the first M smallest non-zero eigenvalues ​​and their corresponding eigenvectors;

[0079] Step S3323: Based on the first M eigenvalues ​​and eigenvectors, construct an anomaly score function for representing the global structure and local anomaly performance; calculate the anomaly score of each node and normalize it using the moving Z-score method;

[0080] Step S3324: read the anomaly score of each node, and detect it through a distributed anomaly detection algorithm, mark the nodes exceeding the threshold as potential anomalies, and form a set of potential abnormal nodes;

[0081] Step S3325: For potential abnormal nodes, consider the node's own time series data and the data of its spatial neighbors, and calculate the cross-correlation function between them; if the correlation is lower than a certain threshold, the node is confirmed to be abnormal; obtain the space-time correlation analysis result, which is the multi-dimensional anomaly detection result.

[0082] Through graph theory and space-time correlation analysis, the accuracy and interpretability of vibration anomaly detection in underground facilities are significantly improved. First, the new graph structure constructed based on node feature representation effectively captures the similarity relationship of vibration features at different locations in underground facilities. The calculation of Laplace matrix and eigenvalue decomposition enable the system to extract the most representative global structural features, which is crucial for understanding the overall vibration mode of underground facilities. The anomaly score function constructed based on eigenvalues ​​and eigenvectors can simultaneously consider local and global anomaly performance, making the system more comprehensive and accurate in identifying vibration anomalies in underground facilities. The application of the moving Z-score method enables the normalization process of the anomaly score to adapt to the time-varying data distribution, which is of great significance for long-term monitoring of the dynamic changes of underground facilities. The use of distributed anomaly detection algorithms not only improves the computational efficiency of the system, but also enhances its adaptability to large-scale underground facility networks. The introduction of space-time correlation analysis enables the system to distinguish between real vibration anomalies and false positive results caused by sensor failure or environmental interference. This analysis method is particularly suitable for underground facilities because it considers the propagation characteristics of vibration in space and the continuity in time. By calculating the cross-correlation function between a node and its spatial neighbors, the system can effectively identify nodes that behave abnormally in both time and space, greatly improving the reliability of detection. In addition, this method can also help identify the propagation path of abnormal vibrations, providing important clues for locating the source of problems in underground facilities. In general, this anomaly detection method based on graph theory and space-time correlation not only improves the accuracy of detection, but also enhances the interpretability of the results, providing more reliable and specific decision-making support for underground facility managers, and helping to promptly discover and deal with potential safety hazards.

[0083] According to one aspect of the present application, in step 4, when the preset threshold is a dynamic threshold, it also includes building an adaptive threshold adjustment mechanism for different scenarios, using the dynamic threshold and anomaly detection results to build a multi-level warning strategy, and outputting multi-level warning information, as follows:

[0084] Step S41, using the exponentially weighted moving average method to calculate the long-term trend of each indicator as a dynamic baseline; using the seasonal adjustment factor to process the periodic changes in the data, using the cumulative sum CUSUM control chart technology to monitor the cumulative deviation of the data flow relative to the dynamic baseline in real time; using the bilateral cumulative sum CUSUM scheme to simultaneously detect anomalies of the upward and downward trends; based on the network topology, setting differentiated threshold strategies for nodes or regions of different importance; outputting a dynamically adjusted anomaly detection threshold set;

[0085] Step S42, construct a hierarchical warning architecture, including three levels: node level, cluster level and network level; at the node level, determine the initial warning level based on the degree and duration of abnormality of a single sensor; at the cluster level, comprehensively consider the number and distribution pattern of abnormal nodes in the cluster to evaluate the risk level of the local area; at the network level, analyze the spatial propagation trend and global impact of the abnormality, judge the systemic risk, and use a fuzzy logic controller to map quantitative abnormal indicators to qualitative warning levels; based on multi-dimensional fuzzy rules, consider the intensity, duration, scope of impact and potential consequences of the abnormality; call a historical case library that stores historical abnormal events and their processing experience; use a dynamic time warping algorithm to match the current situation with historical cases and extract relevant experience; according to the adaptive warning threshold mechanism, dynamically adjust the trigger conditions of different warning levels according to the real-time status of the system and the feedback of the operator.

[0086] The sensitivity and reliability of the vibration monitoring system for underground facilities are significantly improved by constructing dynamic thresholds and multi-level early warning strategies. The application of the exponentially weighted moving average method enables the system to accurately capture the long-term trends of various indicators, providing a reliable basis for the establishment of a dynamic baseline. This is particularly important for underground facilities, as their vibration characteristics may change slowly over time. The introduction of seasonal adjustment factors effectively handles periodic changes in the data, such as regular vibrations during subway operation hours. The use of cumulative and CUSUM control chart techniques enables the system to monitor the cumulative deviation of data streams relative to the dynamic baseline in real time, which is crucial for early detection of progressive problems in underground facilities. The adoption of bilateral cumulative and CUSUM schemes enables the system to detect anomalies in both rising and falling trends and fully grasp the changes in vibration patterns. The differentiated threshold strategy based on the network topology enables the system to perform refined management based on the importance of different nodes or areas, which is of great significance for optimizing resource allocation and improving monitoring efficiency. The establishment of a hierarchical early warning architecture, including three levels: node level, cluster level, and network level, enables the system to comprehensively evaluate the severity and impact of abnormal situations. This multi-level early warning mechanism is particularly suitable for underground facilities because it can focus on both local structural anomalies (such as problems with a single support column) and systemic risks (such as instability of an entire tunnel section). The application of fuzzy logic controllers enables the system to flexibly map quantitative anomaly indicators to qualitative early warning levels. This approach takes into account the uncertainty and complexity in vibration monitoring of underground facilities. The introduction of historical case libraries and the use of dynamic time warping algorithms enable the system to make decisions based on past experience, which is particularly valuable for dealing with abnormal conditions in complex underground environments. The implementation of an adaptive early warning threshold mechanism enables the system to dynamically adjust the early warning conditions based on real-time status and operator feedback. This flexibility ensures that the system can continuously optimize its performance in long-term operation. In general, this dynamic threshold and multi-level early warning strategy greatly improves the intelligence level of the vibration monitoring system for underground facilities, can identify potential risks more accurately and timely, and provides strong technical support for the safety management of underground facilities.

[0087] According to one aspect of the present application, step S5 is also included, i.e., a system-level optimization and reconstruction process, specifically, at predetermined intervals, based on early warning information and network status data, resource demand prediction and optimization allocation are performed to obtain a resource allocation plan; and adaptive network topology reconstruction is performed according to the resource allocation plan and the real-time status information of the system. It should be noted that this step is not necessary, but is a preferred step.

[0088] Step S51, obtaining warning information and network status data, applying time series analysis technology to generate short-term resource demand forecasts; inputting the forecast results into a multi-objective optimization model to generate a preliminary resource allocation plan; using the plan to train a reinforcement learning model, outputting a dynamic resource allocation strategy; applying the strategy to a resource scheduling simulator to generate an optimized resource allocation plan;

[0089] Step S52: Based on the resource allocation plan, perform resource reallocation operations, update and store the system resource distribution status; combine the real-time network topology data, call the graph theory analysis method to evaluate the current network performance, and output the evaluation results; based on the evaluation results, use the dynamic topology adjustment algorithm to generate a set of candidate topology structures; for each candidate topology structure, execute the cognitive radio channel allocation algorithm to determine the optimal communication channel configuration; input the optimized topology structure and channel configuration into the self-organizing network protocol module, and generate node role and connection relationship update instructions; according to these instructions, call the multi-path routing algorithm to build a new routing table; write the updated network topology structure, channel configuration and routing table into the network configuration database, and send update commands to each node to complete the adaptive reconstruction of the network topology.

[0090] The efficiency and adaptability of the underground facility vibration monitoring system are significantly improved through resource demand prediction, optimal allocation and adaptive network topology reconstruction. The application of time series analysis technology enables the system to accurately predict short-term resource demand, which is crucial for responding to emergencies that may occur in underground environments (such as geological changes or human activities). The use of multi-objective optimization models ensures that the resource allocation scheme can simultaneously meet multiple key performance indicators, such as monitoring coverage, energy efficiency and data quality. The introduction of reinforcement learning models enables the system to learn from historical experience and continuously optimize resource allocation strategies. This adaptive ability is particularly important for long-term underground facility monitoring systems. The application of resource scheduling simulators enables the system to evaluate and optimize resource allocation schemes before actual deployment, greatly reducing implementation risks. The use of graph analysis methods enables the system to comprehensively evaluate the current network performance and provide a reliable basis for subsequent topology adjustments. The candidate topology structure set generated by the dynamic topology adjustment algorithm provides the system with a variety of optimization options and enhances the network's ability to adapt to complex underground environments. The application of cognitive radio channel allocation algorithms enables the system to achieve optimal communication channel configuration in underground environments with strong electromagnetic interference, improving the reliability and efficiency of data transmission. The introduction of the self-organizing network protocol module enables the system to automatically update the node roles and connection relationships according to the optimized topology and channel configuration. This adaptive capability is essential for maintaining the long-term stable operation of the underground monitoring network. The use of the multipath routing algorithm enhances the robustness of the network and can maintain the continuity of data transmission when some nodes or links fail, which is of great significance for ensuring the integrity and timeliness of underground facility monitoring data. Overall, this resource optimization and network reconstruction method greatly improves the flexibility and reliability of the underground facility vibration monitoring system. It enables the system to dynamically adjust resource allocation and network structure according to real-time status and predicted needs, and effectively respond to various challenges in the underground environment, such as signal attenuation, energy limitations, and topology changes. This adaptive capability not only improves the overall efficiency of the system, but also extends the service life of the network and reduces maintenance costs. At the same time, it also enhances the system's ability to respond to emergencies, and can quickly reconfigure the network when abnormal vibration occurs in underground facilities to ensure the monitoring quality of key areas. This method provides strong technical support for large-scale and long-term vibration monitoring of underground facilities, and has important practical significance for improving the safety and reliability of underground infrastructure.

[0091] In another embodiment of the present application, in S13, the adjustment process is specifically as follows: using the gradient descent method, with the connection weight as the optimization variable and the network performance score as the objective function, fine-tuning the network; through multiple iterations, gradually adjusting the connection weight until a local optimal solution is reached or a predetermined termination condition is met.

[0092] In another embodiment of the present application, step S5 is specifically as follows:

[0093] Step S51: Resource demand prediction and optimal allocation

[0094] Based on the warning information and network status data output in step S4, resource demand forecasting and optimal allocation are performed. First, time series analysis techniques, such as the ARIMA (autoregressive integrated moving average) model, are applied to predict the resource demand trends of each node and region in the short term. Long short-term memory networks (LSTM) are combined to capture complex nonlinear patterns and long-term dependencies. Then, a multi-objective optimization model is constructed, taking into account energy efficiency, data quality, and system reliability. Genetic algorithms are used to solve the optimization problem and generate a Pareto optimal solution set. To deal with the uncertainty in the optimization process, robust optimization techniques are introduced to ensure the stability of the solution under different scenarios. Next, an adaptive resource allocation strategy is designed. Reinforcement learning algorithms, such as deep Q networks (DQN), are used to learn the optimal resource allocation strategy in a dynamic environment. The reward function design takes into account both short-term performance and long-term sustainability. At the same time, a distributed learning architecture is implemented so that each node can make resource allocation decisions based on local information, reducing the burden on central control. A resource scheduling simulator is developed to test and evaluate the effects of different strategies in a virtual environment. Digital twin technology is used to build a high-fidelity virtual model of the system for large-scale scenario analysis and risk assessment. This step outputs a dynamically optimized resource allocation plan to provide decision support for efficient operation and risk management of the system.

[0095] Step S52: Adaptive network topology reconstruction

[0096] Using the resource allocation scheme of step S51 and the real-time status information of the system, adaptive network topology reconstruction is performed. First, the performance indicators of the current network topology, including connectivity, robustness, and energy distribution balance, are evaluated based on graph theory analysis technology. Identify key nodes and fragile links in the network as the focus of topology optimization. Then, a dynamic topology adjustment algorithm is designed. Using the simulated annealing method, possible topologies are explored under the constraint of maintaining network connectivity. The optimization objective function comprehensively considers network life, data transmission efficiency, and load balancing. In order to adapt to the rapidly changing environment, an incremental topology update mechanism is introduced to allow local adjustment rather than global reconstruction. Next, cognitive radio technology is implemented to enable nodes to dynamically select the best communication channel, reduce interference, and improve spectrum utilization. A distributed collaborative perception algorithm is developed to enable node groups to effectively identify and utilize available spectrum resources. At the same time, a self-organizing network protocol is designed to support the autonomous joining, exiting, and role conversion of nodes, thereby improving the scalability and flexibility of the network. A fault-tolerant mechanism for topology control is implemented. Multipath routing technology is used to enhance network resilience and dynamically balance network load. A predictive maintenance model is developed to actively adjust the topology structure based on the health status of nodes to prevent potential failures. This step outputs an adaptively optimized network topology, providing a dynamic, robust and efficient communication foundation for the system. Step S5 is specifically aimed at dynamic adjustment of the network, improving the robustness and scalability of the system.

[0097] It should be noted that the network reconstruction in step S1 is performed during system initialization and short-term operation, with the main purpose of establishing and maintaining a basic, functional network structure. It is a periodic and relatively frequent process. The network reconstruction in step S5 is based on longer-term observations and more comprehensive information, taking into account longer-term data trends. The reconstruction of S1 focuses on basic network connections and topology, while the reconstruction of S5 is more comprehensive and in-depth, including not only topology but also optimization of resource allocation and communication strategies. The reconstruction of S1 mainly responds to short-term, local changes in the network, such as the failure of a single node or a temporary decrease in communication quality. The reconstruction of S5 targets larger-scale, more persistent changes, which may be caused by factors such as long-term changes in the environment, adjustments to monitoring requirements, or system upgrades. The reconstruction of S1 provides S5 with basic data and a preliminarily optimized network structure. S5 uses this information, combined with long-term observations and more complex algorithms, to perform deeper optimization. The combination of these two steps forms a closed-loop system that enables the network to be continuously optimized at different time scales. By reconfiguring the network at different levels and time scales, the system can better cope with various possible changes and challenges, thereby enhancing overall robustness and adaptability.

[0098] In another embodiment of the present application, a weighted sum method is used to comprehensively consider multiple network performance indicators, and a gradient descent method is used to optimize the network topology. Let {f1, f2, ..., fn} be n network performance indicators, and {w1, w2, ..., wn} be the corresponding weights. Define a comprehensive performance function: F(x) = Σ(wi* fi(x)), i = 1 to n; where x represents a parameterized representation of the network topology.

[0099] Optimize using gradient descent: x(t+1) = x(t) - η * LA F(x(t)); where η is the learning rate. LA is the Laplace operator.

[0100] Generate optimization suggestions: Compare x(t+1) and x(t), extract parameters with significant changes, and convert them into specific network adjustment suggestions.

[0101] In another embodiment of the present application, a basic probability is assigned based on the historical accuracy of the sensor and the quality index of the current data. Then, the Dempster combination rule is used to fuse the evidence of multiple sensors.

[0102] Let Ω be the frame set, and m1, m2, ..., mn be the basic probability distribution of n evidence sources.

[0103] Basic probability distribution: m(A) = α * historical_accuracy + (1 - α) * data_quality; where A Ω, α is a weight factor (0 ≤ α ≤ 1).

[0104] Dempster's combination rule: (m1 O m2)(A) = (Σ(m1(B) * m2(C))) / (1 - K); where B∩ C = A, K = Σ(m1(B) * m2(C)), B ∩ C = Φ; for n sources of evidence, repeatedly apply the combination rule: m = ((m1 O m2) O m3) O ... O mn. O is the combination symbol.

[0105] In another embodiment of the present application, a multi-scale and rotation-invariant LBP variant is used to improve the robustness of the feature. Specifically, LBP is calculated at multiple radii, and an equivalent pattern is used to reduce the feature dimension. Let gc be the gray value of the center pixel, {g0, ..., g7} be the gray values ​​of the surrounding 8 pixels, R be the radius of the LBP operator, and P be the number of sampling points.

[0106] Basic LBP operator: LBP_R, P = Σ s(gp - gc) * 2p, p = 0 to P-1;

[0107] where s(x) = 1 if x ≥ 0, otherwise 0;

[0108] Rotation-invariant LBP: LBP_R, Pri = min(ROR(LBP_R, P, i)), i = 0 to P-1; where ROR is a circular right shift operation.

[0109] Equivalent pattern LBP: An equivalent pattern is defined as a pattern in which the number of 0-1 or 1-0 transitions in a cyclic binary string does not exceed 2 times.

[0110] Multi-scale LBP: LBP is calculated on multiple radii R1, R2, ..., Rk to obtain the feature vector: [LBP_R1, P^riu2, LBP_R2, P^riu2, ..., LBP_Rk, P^riu2]; ^ is a superscript symbol; where riu2 represents the rotationally invariant equivalent mode.

[0111] In another embodiment of the present application, an improved k-means++ method is used to initialize the cluster center, and the maximum number of iterations and the minimum improvement threshold are combined as termination conditions.

[0112] Let X = {x1, x2, ..., xn} be the data set, c be the number of clusters, and m > 1 be the fuzzy factor.

[0113] Randomly select the first cluster center

[0114] For the remaining cluster centers, select the points that are farther away from the selected centers, and the probability is proportional to the square of the distance

[0115] Update the membership matrix U = (uij):uij = 1 / Σ((||xi - vj|| / ||xi - vk||)^(2 / (m-1))), k = 1 to c.

[0116] Update cluster center V = (vj):vj = (Σ(uij^m * xᵢ)) / (Σ(uij^m)), i = 1 ton;

[0117] Termination condition: reaching the maximum number of iterations T_max; or the maximum moving distance of the cluster center is less than the threshold ε: max(||vj_new - vj_old||) < ε.

[0118] In another embodiment of the present application, the graph convolution network anomaly detector includes a network structure of multiple layers of graph convolution and fully connected layers, and uses the reconstruction error as the anomaly score. The autoencoder method is used during training, and normal samples are used for unsupervised learning.

[0119] Let A be the adjacency matrix and X be the node feature matrix.

[0120] Graph convolution layer: H^(l+1) = σ(D^(-1 / 2)AD^(-1 / 2)H^(l)W^(l)); ^ is a superscript symbol;

[0121] Where D is the degree matrix, σ is the activation function, and W is the weight matrix.

[0122] Network structure: Encoder: GCN1 -> GCN2 -> FC1; Decoder: FC2 -> GCN3 -> GCN4.

[0123] Loss function: L = ||X - X'||² + λ * ||A - A'||²; where X' and A' are the reconstructed node features and adjacency matrix respectively, and λ is the trade-off parameter.

[0124] Anomaly score: score(x) = ||x - x'||² + β * ||N(x) - N(x')||²; where N(x) represents the set of neighbor nodes of x and β is a trade-off parameter.

[0125] In another embodiment of the present application, the graph neural network model is specifically: a graph neural network including a multi-layer graph attention mechanism to capture the complex relationship between nodes. At the same time, residual connections and layer normalization are introduced to improve the expression ability and training stability of the model.

[0126] Let h = {h1, h2, ..., hN} be the node feature and eij be the edge feature.

[0127] Graph attention layer: α_ij = softmax_j(a(Wh_i, Wh_j, e_ij)); h_i' = σ(Σ(α_ij *Wh_j))

[0128] Multi-head attention: h_i' = ||k σ(Σ(α_ij^k * W^k h_j)); where || represents the concatenation operation and k is the number of attention heads.

[0129] Residual connection and layer normalization: h_i^(l+1) = LayerNorm(h_i^l + MLP(GAT(h_i^l)));

[0130] The overall network structure is: Input -> GAT1 -> GAT2 -> GAT3 -> FC -> Output; each GAT layer is followed by residual connection and layer normalization.

[0131] In another embodiment of the present application, the adaptive threshold method may also be:

[0132] Let X_t be the observation at time t and α be the smoothing factor (0 < α < 1).

[0133] Calculate EWMA: S_t = α * X_t + (1 - α) * S_(t-1); Calculate dynamic threshold: θ_t = μ_t + k* σ_t;

[0134] Where μ_t and σ_t are the mean and standard deviation of S_t respectively, and k is the adjustment factor.

[0135] Threshold fine-tuning: If FPR > FPR_target, then k = k * (1 + δ); if FNR > FNR_target, then k = k * (1 - δ); where FPR is the false alarm rate, FNR is the false negative rate, and δ is a small positive number (such as 0.05).

[0136] In another embodiment of the present application, the process of the multi-objective optimization function is specifically as follows:

[0137] A multi-objective optimization method based on Pareto optimality is adopted to optimize the network topology using NSGA-II (Non-dominated Sorting Genetic Algorithm II). This method can simultaneously consider multiple potentially conflicting objectives, such as network connectivity, energy efficiency, and load balancing. Let f = (f1, f2, ..., fk) be k optimization objectives.

[0138] Define the dominance relation: A solution x dominates a solution y if for all i, f_i(x) ≤ f_i(y), and there exists at least one j such that f_j(x) < f_j(y).

[0139] NSGA-II main steps: initialize population P_t; generate offspring Q_t; merge R_t = P_t ∪ Q_t; perform non-dominated sorting on R_t to obtain the frontier F = (F1, F2, ...); select a new population P_(t+1) until |P_(t+1)| + |F_i| > N; use crowding distance sorting to select individuals in F_i; repeat steps bf until the termination condition is met;

[0140] Optimization objective definition: f1: network connectivity = minimum cut set size; f2: energy efficiency = -Σ(transmission energy consumption) / number of nodes; f3: load balance = -max(node ​​load) / avg(node ​​load).

[0141] In another embodiment of the present application, the time synchronization problem in a distributed system is solved as follows:

[0142] Divide the data into K partitions: D = {D1, D2, ..., Dk} Compute local statistics in parallel on each partition: S_i = f(D_i) Merge local statistics: S = g(S1, S2, ..., Sk);

[0143] For attribute a, calculate the information gain ΔG_a. If ΔG_a - ΔG_b > ε, select attribute a for splitting, where ε = sqrt((R² * ln(1 / δ)) / (2n)) R is the attribute range, δ is the confidence, and n is the number of samples;

[0144] Maintain a sliding window W of dynamic size. If there is a cut point c such that |μ_W1 - μ_W2| > ε_cut, discard the older subwindow where ε_cut = sqrt((1 / m) * ln(2 / δ') * (1 / 2m)) m is the current window size and δ' is the confidence parameter;

[0145] NTP synchronization: offset θ = ((T2 – T1) + (T3 – T4)) / 2; round-trip delay δ = (T4 – T1)- (T3 – T2);

[0146] Where: T1 is the time when the client sends the request; T2 is the time when the server receives the request; T3 is the time when the server sends the response; T4 is the time when the client receives the response. By calculating the offset θ and the round-trip delay δ, the client can adjust its local clock to synchronize it with the server time. The adjusted time can be expressed as:

[0147] Adjusted time = current local time + θ;

[0148] Logical clock: Each event e is assigned a timestamp LC(e); if a -> b (a occurs before b), then LC(a) < LC(b); when process P_i sends message m: LC(m) = LC(P_i) + 1; when process P_j receives message m: LC(P_j) = max(LC(P_j), LC(m)) + 1;

[0149] Vector clock: Each process P_i maintains a vector VC_i = [c_i1, c_i2, ..., c_in]; when a local event occurs in P_i: c_ii = c_ii + 1; when P_i sends a message: first perform the local event update, then send VC_i; when P_j receives a message from P_i: for all k, c_jk = max(c_jk, c_ik); then c_jj = c_jj + 1;

[0150] Causal consistency judgment: Event a causally precedes event b if and only if: VC(a) < VC(b), that is, for all k, VC(a)[k] ≤ VC(b)[k], and there exists at least one l such that VC(a)[l] < VC(b)[l].

[0151] In summary, several innovative techniques are used in step S2 to effectively capture complex vibration patterns, including:

[0152] The original vibration signal is analyzed using wavelet packet transform, which can analyze vibration signals in different frequency bands and effectively identify nonlinear and non-stationary characteristics. A multi-scale feature descriptor is constructed to capture the characteristics of the signal at different scales and adapt to the changing characteristics of non-stationary signals. Specifically, the statistical feature vector is calculated for each frequency band, and the time-frequency domain texture features are extracted using the improved local binary pattern; the feature vector and the time-frequency domain texture features are combined to form a multi-scale feature descriptor. The dimensionality reduction technology is applied to retain key information, which reduces the complexity of the data while retaining the key nonlinear characteristics of the signal.

[0153] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for monitoring vibration of underground facilities based on a distributed sensor network, characterized in that: The steps include: Step S1, obtaining initial data of a distributed sensor network; constructing an initial network topology structure; collecting sensor node status data; performing dynamic topology adjustment; evaluating and optimizing network performance; and obtaining an optimized network topology structure; Step S2, receiving the raw data collected by the sensor and using the network topology structure to clean and standardize the data; Use filtering algorithms to remove noise and analyze and extract features; Generate preprocessed feature dataset; Step S3, input the preprocessed feature data set; use the multi-source data fusion algorithm module to integrate the information; Build an anomaly prediction model and optimize parameters, use the optimized anomaly prediction model to perform anomaly detection; output fused data and anomaly detection results; Step S4: for the abnormal detection result, determine whether to trigger an early warning according to a preset threshold; Generate early warning information; implement early warning response measures; output monitoring reports and early warning status; The step S1 is specifically as follows: Step S11, receiving and constructing an initial network graph based on the location coordinates and initial connection relationship data of the sensor network, wherein nodes represent sensors and edges represent connections between sensors; collecting signal strength data of each sensor node within a preset time window; measuring the communication delay between adjacent nodes and recording the round-trip time; integrating the collected signal strength and communication delay data into a state matrix; wherein the matrix elements are used to represent the connection quality between node pairs; Step S12, reading the generated state matrix and the initial network diagram, removing the connections in the state matrix that are lower than the threshold according to the set connection quality threshold, and updating the network diagram; for isolated nodes formed by the connection removal, applying the Delaunay triangulation algorithm to re-establish the connection, while minimizing the long-distance connection while maintaining the network coverage; using the Kruskal minimum spanning tree algorithm to ensure the connectivity of the entire network; after completing the dynamic topology adjustment, obtaining an updated network diagram; Step S13, read the updated network diagram, calculate the network performance indicators and form a performance indicator set; call the multi-objective optimization function, evaluate the quality of the current network topology based on the performance indicator set, and output the optimization score and specific optimization suggestions; if the optimization score is lower than the preset threshold, make further adjustments according to the optimization suggestions; if it is not lower than the preset threshold, output the optimized network topology structure; wherein the network performance indicators include network connectivity, average node degree, network diameter, energy efficiency and load balancing.

2. The underground facility vibration monitoring method based on a distributed sensor network as claimed in claim 1, characterized in that: The step S2 is further as follows: Step S21, receiving the optimized network topology structure, extracting sub-networks, using the distributed Bellman-Ford algorithm to calculate the shortest path for each sub-network, selecting a central node as a data aggregation point, and constructing a data collection tree Ti with the central node as the root; passing the structural information of the data collection tree Ti to all nodes of each sub-network; Step S22: On each sensor node, the structural information of the data collection tree is received, and the signal change rate is continuously calculated through an adaptive sampling algorithm. When the signal change rate is greater than a predetermined threshold, the sampling frequency is increased to a maximum value; otherwise, the sampling frequency is gradually reduced to a minimum value; and the original vibration signal is collected and stored according to the current sampling frequency; Step S23, receiving the original vibration signal, using Daubechies wavelet as the basis function, applying wavelet packet transform to the original vibration signal, and obtaining a set of wavelet packet coefficients; wherein the decomposition level J of the wavelet packet transform is determined according to the highest frequency fmax and the lowest frequency fmin of the signal: J = log2(fmax / fmin); Step S24, obtaining a wavelet packet coefficient set, calculating a statistical feature vector for each frequency band, and using an improved local binary pattern to extract time-frequency domain texture features; The feature vector and the time-frequency domain texture features are combined to form a multi-scale feature descriptor; Step S25, receiving the multi-scale feature descriptor, reducing the dimension of the multi-scale feature descriptor, retaining the minimum principal components required to explain 95% of the variance, and obtaining the reduced-dimensional multi-scale feature descriptor; Dynamic Huffman coding is applied to compress the multi-scale feature descriptors with reduced dimensions. The coding tree is dynamically updated according to real-time data statistics to obtain reduced-dimensional compressed data. Step S26, receiving the dimension reduction compressed data and the structural information of the data collection tree, and transmitting the dimension reduction compressed data from the leaf node to the central node through the data collection tree; Using the jump data aggregation technology, the data from different sub-nodes are partially aggregated at each intermediate node to obtain aggregated data, and a dimensionality reduction feature matrix is ​​constructed based on the aggregated data, which is the preprocessed feature data set.

3. The underground facility vibration monitoring method based on a distributed sensor network as claimed in claim 2, characterized in that: The step S3 is further as follows: Step S31, obtain the dimension reduction feature matrix, apply the Dempster-Shafer evidence theory to perform multi-source data fusion, the fusion process is specifically as follows: according to the historical reliability and current data quality of each sensor, assign a basic probability distribution function to each data source; use the Dempster combination rule to fuse the evidence of multiple data sources, calculate the comprehensive belief function and likelihood function; based on the fused belief function, construct a comprehensive feature matrix; Step S32, read the comprehensive feature matrix, use the fuzzy c-means clustering algorithm to perform cluster analysis on the fused data, and determine the optimal number of clusters and cluster centers through an iterative optimization process, and identify the main patterns and groups in the data; based on the clustering results, construct a Gaussian mixture model; for each cluster, estimate its mean vector and covariance matrix to form a weighted combination of multiple Gaussian distributions; optimize the GMM parameters through the expectation maximization algorithm; and obtain a GMM model that can describe the distribution characteristics of normal data; Step S33, perform multi-scale anomaly detection through the GMM model or the pre-built graph analysis model. When the GMM model is used, the specific steps are as follows: at the single node level, calculate the log-likelihood value of each data point, and mark the points that deviate significantly from the expected distribution as potential anomalies; at the local cluster level, apply the local anomaly factor algorithm, and consider the local density of the data points to identify samples that are abnormal relative to neighboring points; at the global network level, use the graph convolutional network anomaly detector to capture abnormal behaviors that violate the global pattern; combine anomaly evidence from different scales based on the Dempster-Shafer theory to resolve conflicts in the detection results; use the Moran's I index to evaluate the spatial autocorrelation of anomalies, perform spatial correlation analysis, and distinguish between local anomalies and systematic anomalies; Output multi-dimensional anomaly detection results, including spatiotemporal feature descriptions of abnormal data points, abnormal regions, and abnormal patterns.

4. The underground facility vibration monitoring method based on a distributed sensor network as claimed in claim 3, characterized in that: In step S33, a pre-built graph analysis model is used to perform multi-scale anomaly detection, further comprising: Step S331, read the dimension reduction feature matrix, construct a multi-layer graph structure, use a multi-scale graph clustering algorithm to fuse information of different scales, and obtain a low-dimensional embedding space; in the low-dimensional embedding space, use an adaptive density peak clustering algorithm to fuse data and construct a hierarchical data representation; construct a graph neural network model and optimize parameters, update the node feature representation through a message passing mechanism, and obtain the final node feature representation set; Step S332: read the node feature representation set, construct a graph structure, use graph theory to analyze the graph structure, calculate the Laplacian matrix of the graph and solve its eigenvalue problem; Call the graph-based anomaly score function to calculate the anomaly score of each node and normalize it; The improved box plot method is used to calculate the local anomaly threshold, and spatial-temporal correlation analysis is performed to output multi-dimensional anomaly detection results.

5. The underground facility vibration monitoring method based on a distributed sensor network as claimed in claim 4, characterized in that: The step S331 is further as follows: Step S3311, read the dimension reduction feature matrix, and build a multi-layer graph structure based on the dimension reduction feature matrix, where each layer represents a different scale of the data; within each scale, calculate the relationship strength between nodes, and use cosine similarity to measure the similarity of node features; Connections are established between different layers, and the strength of the connections depends on the similarity of node features at different scales. A multi-dimensional graph structure containing multi-scale topological information and node features is obtained. Step S3312: read the multi-dimensional graph structure, construct a graph Laplacian matrix that comprehensively considers all scales, solve the eigenvalues ​​of the graph Laplacian matrix, select the eigenvectors corresponding to the smallest N non-zero eigenvalues, and each eigenvector forms a low-dimensional embedding space; Step S3313: in the low-dimensional space, calculate the local density of each data point and its shortest distance to a point with higher density, select the cluster center through a dynamic threshold function, implement adaptive density peak clustering, and obtain a set of data clustering results; Step S3314: construct a hierarchical data representation based on the clustering results, specifically: for each cluster, construct and calculate a feature vector that integrates the features of all nodes in the cluster and takes into account the importance of the nodes, merge similar clusters to form a multi-level structure, and obtain a hierarchical feature set; Step S3315: call the pre-built graph neural network model containing multiple graph convolutional layers, with a multi-layer graph structure and hierarchical data representation as input; the calculation of each layer takes into account the characteristics of the node itself and the information of neighboring nodes; the node feature representation is updated through the message passing mechanism to obtain the final node feature representation set.

6. The underground facility vibration monitoring method based on a distributed sensor network as claimed in claim 4, characterized in that: The step S332 is further as follows: Step S3321: Based on the node feature representation, a new graph structure is constructed, wherein the edge weights in the graph structure represent the similarity of the node features; based on the graph structure, an adjacency matrix and a degree matrix are calculated to obtain a Laplacian matrix and normalize it; Step S3322, performing eigenvalue decomposition on the normalized Laplace matrix to obtain several columns of eigenvalues ​​and corresponding eigenvectors, and extracting the first M smallest non-zero eigenvalues ​​and their corresponding eigenvectors; Step S3323: Based on the first M eigenvalues ​​and eigenvectors, construct an anomaly score function for representing the global structure and local anomaly performance; calculate the anomaly score of each node and normalize it using the moving Z-score method; Step S3324: read the anomaly score of each node, and detect it through a distributed anomaly detection algorithm, mark the nodes exceeding the threshold as potential anomalies, and form a set of potential abnormal nodes; Step S3325: For a potential abnormal node, consider the node's own time series data and the data of its spatial neighbors, and calculate the cross-correlation function between them; if the correlation is lower than a certain threshold, the node is confirmed to be abnormal; The spatial-temporal correlation analysis results are obtained, which are multi-dimensional anomaly detection results.

7. The underground facility vibration monitoring method based on a distributed sensor network as claimed in claim 1, characterized in that: In step 4, when the preset threshold is a dynamic threshold, it also includes building an adaptive threshold adjustment mechanism for different scenarios, using the dynamic threshold and anomaly detection results to build a multi-level warning strategy and output multi-level warning information, as follows: Step S41, using the exponentially weighted moving average method to calculate the long-term trend of each indicator as a dynamic baseline; using the seasonal adjustment factor to process the cyclical changes in the data, using the cumulative sum CUSUM control chart technology to monitor the cumulative deviation of the data flow relative to the dynamic baseline in real time; using the bilateral cumulative sum CUSUM scheme to simultaneously detect anomalies in both upward and downward trends; Based on the network topology, differentiated threshold strategies are set for nodes or areas of different importance; Output dynamically adjusted anomaly detection threshold set; Step S42: construct a hierarchical warning architecture, including three levels: node level, cluster level and network level; at the node level, determine the initial warning level based on the abnormality degree and duration of a single sensor; At the cluster level, the number and distribution pattern of abnormal nodes in the cluster are comprehensively considered to assess the risk level of the local area; at the network level, the spatial propagation trend and global impact of the anomaly are analyzed to judge the systemic risk, and the fuzzy logic controller is used to map the quantitative anomaly indicators to the qualitative warning level; based on multi-dimensional fuzzy rules, the intensity, duration, impact range and potential consequences of the anomaly are considered; Call the historical case library that stores historical abnormal events and their handling experience; use the dynamic time warping algorithm to match the current situation with historical cases and extract relevant experience; based on the adaptive warning threshold mechanism, dynamically adjust the trigger conditions of different warning levels according to the real-time status of the system and the feedback from the operator.

8. The underground facility vibration monitoring method based on a distributed sensor network as claimed in claim 1, characterized in that: The method further comprises step S5, performing resource demand prediction and optimal allocation based on the early warning information and the network status data at predetermined intervals to obtain a resource allocation plan; performing adaptive network topology reconstruction according to the resource allocation plan and the real-time status information of the system; Step S51, obtaining warning information and network status data, applying time series analysis technology to generate short-term resource demand forecasts; inputting the forecast results into a multi-objective optimization model to generate a preliminary resource allocation plan; Use this solution to train a reinforcement learning model and output a dynamic resource allocation strategy; Apply the strategy to the resource scheduling simulator to generate an optimized resource allocation plan; Step S52: Based on the resource allocation plan, perform resource reallocation operations, update and store the system resource distribution status; combine the real-time network topology data, call the graph theory analysis method to evaluate the current network performance, and output the evaluation results; based on the evaluation results, use the dynamic topology adjustment algorithm to generate a set of candidate topology structures; for each candidate topology structure, execute the cognitive radio channel allocation algorithm to determine the optimal communication channel configuration; input the optimized topology structure and channel configuration into the self-organizing network protocol module, and generate node role and connection relationship update instructions; according to these instructions, call the multi-path routing algorithm to build a new routing table; write the updated network topology structure, channel configuration and routing table into the network configuration database, and send update commands to each node to complete the adaptive reconstruction of the network topology.

9. The underground facility vibration monitoring system based on distributed sensor network is characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the underground facility vibration monitoring method based on a distributed sensor network as described in any one of claims 1 to 8.

Images