A cold region ice-containing joint rock mass instability early warning method based on multi-source data fusion
By integrating multi-source data and clustering analysis, combined with machine learning models, the problem of insufficient accuracy and timeliness in early warning of instability of icy jointed rock masses in cold regions has been solved, and accurate risk assessment and early warning of rock mass stability have been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHAOXING UNIVERSITY
- Filing Date
- 2026-06-11
- Publication Date
- 2026-07-14
AI Technical Summary
Traditional monitoring methods are insufficient to fully capture the precursory features of instability in icy jointed rock masses in cold regions, resulting in inadequate accuracy and timeliness in early warning. Existing technologies have failed to effectively integrate multi-source heterogeneous data and lack collaborative perception of micro-fractures, stress redistribution, and temperature field evolution within the rock mass.
By dividing the target area into monitoring units, collecting multi-source heterogeneous time-series data, and constructing a standardized fusion dataset after preprocessing, statistical features and risk trend coefficients are extracted using cluster analysis and machine learning models to generate a visual early warning image of instability risk level.
It enables precise risk zoning identification and dynamic assessment of glacial jointed rock masses in cold regions, significantly improving the accuracy and timeliness of early warning.
Smart Images
Figure CN122392282A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of rock mass instability early warning technology, specifically to a method for early warning of instability of icy jointed rock masses in cold regions based on multi-source data fusion. Background Technology
[0002] Stability monitoring of glacial jointed rock masses in cold regions is a key technical challenge for engineering construction in high-altitude and frigid areas. Traditional monitoring methods mostly rely on single-type sensors for fixed-point measurements, lacking the ability to collaboratively perceive multi-physical field information such as micro-fractures, stress redistribution, and temperature field evolution within the rock mass. This makes it difficult to comprehensively capture the precursory characteristics of rock mass instability, resulting in insufficient accuracy and timeliness of early warnings.
[0003] In existing technologies, the artificial intelligence-based method and system for monitoring and early warning of cracks in ice-rock masses, published in CN120470545A, although using multimodal sensors and deep learning models for crack monitoring, focuses primarily on simulating the propagation path of existing cracks, and has the following shortcomings: First, it lacks sufficient perception of latent precursor information such as ice phase transitions and microseismic activity at joint surfaces within the rock mass; second, it fails to consider the spatial heterogeneity of rock masses in cold regions, making it difficult to achieve refined risk zoning assessment; and third, it lacks quantitative characterization of the dynamic trends of multi-source temporal characteristics, resulting in limited predictive power. Therefore, there is an urgent need for an early warning method for rock mass instability in cold regions that can integrate multi-source heterogeneous data, reveal the regional risk distribution patterns, and quantify evolution trends.
[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention
[0005] The purpose of this invention is to provide an early warning method for instability of glacial jointed rock masses in cold regions based on multi-source data fusion, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for early warning of instability of glacial jointed rock masses in cold regions based on multi-source data fusion, comprising the following steps: S1. The target cold-region glacial jointed rock mass area is divided into several monitoring units. Multi-source heterogeneous time-series data of each monitoring unit are collected, including microseismic monitoring data, acoustic emission data, surface displacement data, rock mass stress data and infrared thermal imaging data. The collected data are preprocessed to construct a standardized fusion dataset. S2, based on the standardized fusion dataset, extracts the energy and frequency of microseismic events, the ringing count and energy rate of acoustic emission events, the surface displacement rate and cumulative displacement, the rock mass stress change gradient, and the area ratio of infrared thermal image temperature anomalies in each monitoring unit within the current monitoring time window, thus forming the statistical feature set of each monitoring unit. S3, perform cluster analysis on the statistical feature set of each monitoring unit, and divide the monitoring units whose feature similarity index meets the preset threshold into the same unit cluster to obtain several unit clusters; S4. For each unit cluster, analyze the changes in its statistical feature set within a preset time window, and calculate the risk trend coefficient of the unit cluster within the preset time window. S5, input the statistical feature set and risk trend coefficient of each unit cluster into the pre-trained machine learning model to obtain the real-time instability risk probability of each unit cluster; S6 compares the instability risk probability of each cluster with the preset probability threshold range to determine the instability risk level of each cluster and generate a visual early warning image of the instability risk level covering the target cold-region icy jointed rock mass area.
[0007] Furthermore, the specific execution process of S1 is as follows: The target cold region containing ice-bearing jointed rock mass is evenly divided into multiple monitoring units. The geometric center of each monitoring unit is used as a monitoring node. Multi-source heterogeneous time series data are collected at each monitoring node, and the multi-source heterogeneous time series data at the monitoring node is labeled as the multi-source heterogeneous time series data of the corresponding monitoring unit. Microseismic monitoring data are collected from each monitoring node using microseismic sensors; Acoustic emission data from each monitoring node is collected using acoustic emission probes. Ground displacement data of each monitoring node is collected using GNSS positioning equipment; Rock mass stress data at each monitoring node are collected using stress sensors; Infrared thermal imaging data of each monitoring node are collected using an infrared thermal imager; The collected data underwent preprocessing, including outlier removal and missing value imputation. Outlier removal employed the Laida criterion, and the mean of each data category was calculated. with standard deviation The normal fluctuation range of the data is determined as [ Data that deviates from the normal fluctuation range will be identified as outliers and removed. Missing value completion uses linear interpolation: for each missing data point, all similar data points that were not removed are counted if their timestamp interval is less than a preset time, and the average value is taken as the completion value to complete the missing value. Then, the preprocessed multi-source heterogeneous time series data of each monitoring node are aligned and integrated according to timestamp to construct a standardized fusion dataset covering the target cold region.
[0008] Furthermore, the energy of the microseismic event is obtained by calculating the arithmetic mean of the energies of all microseismic events falling within the current monitoring time window; the frequency of the microseismic event is obtained by counting all microseismic events falling within the current monitoring time window. The acoustic emission event ringing count is obtained by counting the total number of times the waveform signals of all acoustic emission events falling within the current monitoring time window exceed a preset voltage threshold; the acoustic emission event energy rate is obtained by calculating the sum of the energies of all acoustic emission events falling within the current monitoring time window. The surface displacement rate is obtained by calculating the resultant displacement change of surface displacement data in three dimensions within the current monitoring time window; the cumulative displacement is obtained by summing all the resultant displacement changes obtained from the start of monitoring to the end of the current monitoring time window. The rock mass stress change gradient is obtained by calculating the absolute value of the maximum principal stress change in the rock mass stress data within the current monitoring time window. The percentage of the area of the abnormal temperature region in the infrared thermal image refers to the ratio of the area of the pixel region whose temperature exceeds the historical average temperature threshold range to the total imaging area, calculated based on the infrared thermal imaging data collected within the current monitoring time window, and the average value is taken as the percentage of the area of the abnormal temperature region in the infrared thermal image. The statistical features extracted above are used to construct the statistical feature set of each monitoring unit.
[0009] Furthermore, cluster analysis is performed on the statistical feature sets of each monitoring unit. Monitoring units whose feature similarity index meets a preset threshold are grouped into the same cluster, thus obtaining several unit clusters. The specific logic behind this is as follows: For each monitoring unit, the statistical feature set within its current monitoring time window is taken, and the statistical feature set is used as the current feature vector of that monitoring unit. Calculate the Euclidean distance between the current feature vectors of any two monitoring units, and use the calculation result as the feature similarity index of the two monitoring units; The formula used to calculate the feature similarity index is as follows: In the formula, Indicates monitoring unit With monitoring unit The feature similarity index between two units is such that the smaller the value, the more similar the features of the two units are. This represents the total dimension of the current feature vector, i.e., the total number of extracted statistical features; and Representing monitoring units With monitoring unit The current feature vector at the th The numerical values on the dimensional features, where p is the index of the statistical feature; Set similarity threshold and minimum cluster size The initial division of unit clusters is completed through the following steps: All monitoring units are aggregated to form a unit set, and all monitoring units in the unit set are defined as in an inaccessible state. Monitoring units are then randomly selected from the unit set. Based on the feature similarity index, the monitoring units are determined by traversing the unit set. neighborhood Specifically, if there exists a monitoring unit in the unit set whose status is unvisited, and this unit is related to unit... The feature similarity index between them is less than the similarity threshold. This unit will then be included in the monitoring unit. neighborhood middle; If monitoring unit neighborhood The number of internal cells is less than the minimum cluster size If the unvisited monitoring unit is selected from the unit set, then the unit will be reselected; otherwise, the unit will be removed. As the core point, and to create unit clusters Unit clustering By monitoring unit and neighboring areas Composition, and based on clustering Real-time updates of user status within a cell set, specifically: incorporating users from the cell set into clusters. The cell status has been changed to "accessed"; Another unvisited monitoring unit is randomly selected from the unit set to build a new cluster, until no new cluster can be built. This completes the initial division of the unit set. The initial clustering is then optimized by merging the clusters: if the minimum distance between two cluster centers is less than the merging threshold... If the two clusters are not found, then merge them into a new cluster. Finally obtained Clustering of individual units ,in, The number of unit clusters, ; And clustering for any unit Filter out clusters with units The monitoring unit with the smallest feature similarity index among the cluster centers is defined as the unit cluster. The representative unit is identified, and the statistical characteristics of this monitoring unit are labeled as unit clustering. Statistical characteristics.
[0010] Furthermore, for each unit cluster, the changes in its statistical feature set within a preset time window are analyzed. This preset time window consists of the most recent consecutive clusters. The cluster consists of several monitoring time windows. The risk trend coefficient of the cluster within the preset time window is calculated based on the following logic: For each unit cluster, select its nearest contiguous cluster. The statistical feature set of each monitoring time window, among which For each statistical feature, extract its value in this context. The representation values within each monitoring time window form a length of [value]. The statistical characteristics of this time series subsequence; The slope of the time-series subsequence of each statistical feature of this cluster is calculated over time using a linear fitting method. The specific formula used is as follows: In the formula, This indicates the first cluster in this unit cluster. The trend slope of a statistical feature; This indicates the total number of recent consecutive monitoring time windows; t Index for monitoring time windows ； in Represents the earliest monitoring time window, Represents the current monitoring time window; This indicates that the unit is clustered in the th order. Within the monitoring time window, the first The representational value of each statistical characteristic; An index for statistical features; The method for determining the risk trend coefficient is as follows: First, the trend slopes of each statistical feature are standardized to eliminate differences in their numerical magnitudes. Then, each standardized trend slope is multiplied by its corresponding weight coefficient. Finally, all weighted results are summed, and the resulting value is designated as the risk trend coefficient for that cluster. The specific formula used is as follows: In the formula, This represents the risk trend coefficient of the cluster. Indicates the first The preset weighting coefficients corresponding to the slope of each statistical characteristic trend satisfy the following conditions. ; and These represent the mean and standard deviation of the trend slopes of all statistical characteristics of the cluster in this unit, respectively.
[0011] Furthermore, for each unit cluster, its statistical feature set and risk trend coefficient are fused to construct a comprehensive feature vector for input to the machine learning model; the comprehensive feature vector consists of two parts: one part is based on the nearest continuous cluster of that unit cluster. The time-series statistics are calculated from the statistical feature set within a monitoring time window, specifically including the mean, standard deviation, maximum and minimum values of each statistical feature within that time period; another part is the risk trend coefficient of the clustering of that unit. The comprehensive feature vector is input into a pre-trained machine learning model, which is trained using a gradient boosting decision tree algorithm. Its output is a continuous probability value between 0 and 1, which is the real-time instability risk probability of unit clustering. The training process of the machine learning model specifically includes: collecting historical monitoring data of the target cold region or similar geological conditions area; using the statistical feature set of each unit cluster extracted from the historical data as the statistical feature set and risk trend coefficient as the model input features; using the rock mass instability state actually observed in the corresponding time period as the training label, where the stable state is marked as 0 and the unstable state is marked as 1; using the ten-fold cross-validation method to optimize the model parameters; determining the optimal tree depth, learning rate and number of trees through grid search; and finally obtaining the trained machine learning early warning model.
[0012] Furthermore, four consecutive probability threshold intervals are preset, namely the risk-free interval, low-risk interval, medium-risk interval and high-risk interval, which correspond to the risk-free level, low-risk level, medium-risk level and high-risk level respectively; the real-time instability risk probability of each unit cluster is compared with the above-mentioned preset probability threshold intervals to determine the instability risk level corresponding to each unit cluster. Based on the cluster location coordinates of each unit cluster and the determined instability risk level, a color-coded rendering method is used to mark clusters of different risk levels with corresponding characteristic colors on the geographic base map of the target cold-region glacial jointed rock mass area, generating a visual early warning image of instability risk level covering the entire target area. At the same time, corresponding early warning signals are triggered for clusters of medium risk and above. The early warning signals include risk level prompts, cluster location coordinates and risk development trend descriptions. The spatial coordinates of each cluster are represented by the spatial coordinates of the representative unit corresponding to that cluster. The risk development trend description is generated based on the recent changes in the risk trend coefficient and its statistical feature set of the clustering unit, including the direction of risk change, the rate of risk change, and the analysis of the main driving factors; specifically: when the risk trend coefficient is positive, it indicates that the risk is on an upward trend, and when it is negative, it indicates that the risk is on a downward trend; the rate of risk change is determined by calculating the change in the risk trend coefficient within the two most recent monitoring time windows; and the main driving factors are determined by identifying the statistical features with the largest changes in the statistical feature set.
[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention integrates multi-source monitoring data, including microseismic, acoustic emission, displacement, stress, and infrared thermal imaging, to construct a statistical feature set that comprehensively reflects the stability state of icy jointed rock masses. It employs cluster analysis to identify risk zones within the monitoring area and dynamically assesses the stability trends of each zone using risk trend coefficients. Finally, a machine learning model enables accurate probability prediction of zoned risks, significantly improving the accuracy and timeliness of early warning for instability of icy jointed rock masses in cold regions. Attached Figure Description
[0014] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 for , and 3D scatter image; Figure 3 for , and 3D scatter image. Detailed Implementation
[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.
[0016] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0017] Example: Please see Figures 1-3 The present invention provides a technical solution: A method for early warning of instability of glacial jointed rock masses in cold regions based on multi-source data fusion, comprising the following steps: S1. The target cold-region glacial jointed rock mass area is divided into several monitoring units. Multi-source heterogeneous time-series data of each monitoring unit are collected, including microseismic monitoring data, acoustic emission data, surface displacement data, rock mass stress data and infrared thermal imaging data. The collected data are preprocessed to construct a standardized fusion dataset. In this embodiment, the specific execution process of S1 is as follows: The target cold region containing ice-bearing jointed rock mass is evenly divided into multiple monitoring units. The geometric center of each monitoring unit is used as a monitoring node. Multi-source heterogeneous time series data are collected at each monitoring node, and the multi-source heterogeneous time series data at the monitoring node is labeled as the multi-source heterogeneous time series data of the corresponding monitoring unit. Microseismic monitoring data are collected from each monitoring node using microseismic sensors; Acoustic emission data from each monitoring node is collected using acoustic emission probes. Ground displacement data of each monitoring node is collected using GNSS positioning equipment; Rock mass stress data at each monitoring node are collected using stress sensors; Infrared thermal imaging data of each monitoring node are collected using an infrared thermal imager; The collected data underwent preprocessing, including outlier removal and missing value imputation. Outlier removal employed the Laida criterion, and the mean of each data category was calculated. with standard deviation The normal fluctuation range of the data is determined as [ Data that deviates from the normal fluctuation range will be identified as outliers and removed. Missing value completion uses linear interpolation: for each missing data point, all similar data points that were not removed are counted if their timestamp interval is less than a preset time, and the average value is taken as the completion value to complete the missing value. Then, the preprocessed multi-source heterogeneous time series data of each monitoring node are aligned and integrated according to timestamp to construct a standardized fusion dataset covering the target cold region.
[0018] Step S1 divides the target area into several monitoring units and deploys a multi-source sensor array at the center of each unit to achieve five-dimensional synchronous monitoring of microseismic activity, acoustic emission signals, surface displacement, rock mass stress, and infrared thermography of icy jointed rock masses in cold regions. This step uses the Laida criterion to remove abnormal data and fills in missing values through linear interpolation of time windows, constructing a standardized fusion dataset with spatiotemporal alignment. This data foundation not only solves the limitations of traditional single-sensor monitoring, but more importantly, provides high-quality multi-source heterogeneous time-series data support for subsequent extraction of statistical feature sets, cluster analysis, and calculation of risk trend coefficients.
[0019] S2, based on the standardized fusion dataset, extracts the energy and frequency of microseismic events, the ringing count and energy rate of acoustic emission events, the surface displacement rate and cumulative displacement, the rock mass stress change gradient, and the area ratio of infrared thermal image temperature anomalies in each monitoring unit within the current monitoring time window, thus forming the statistical feature set of each monitoring unit. In this embodiment, the energy of the microseismic event is obtained by calculating the arithmetic mean of the energies of all microseismic events falling within the current monitoring time window; the frequency of the microseismic event is obtained by counting all microseismic events falling within the current monitoring time window; wherein, the energy of the microseismic event is used to characterize the average level of energy released by microfractures inside the rock mass during this time period, and the frequency of the microseismic event is used to reflect the frequency of microfracture activity inside the rock mass. The acoustic emission event ring count is obtained by counting the total number of times the waveform signals of all acoustic emission events falling within the current monitoring time window exceed a preset voltage threshold; the acoustic emission event energy rate is obtained by calculating the sum of the energies of all acoustic emission events falling within the current monitoring time window; wherein, the acoustic emission event ring count is used to characterize the activity of crack propagation inside the rock mass; the acoustic emission event energy rate is used to reflect the energy scale of accumulated damage inside the rock mass during this time period; The surface displacement rate is obtained by calculating the resultant displacement change of surface displacement data in three dimensions within the current monitoring time window; the cumulative displacement is obtained by summing all the resultant displacement changes obtained from the start of monitoring to the end of the current monitoring time window; wherein, the cumulative displacement is used to reflect the historical cumulative effect of rock mass deformation.
[0020] The rock mass stress change gradient is obtained by calculating the absolute value of the maximum principal stress change in the rock mass stress data within the current monitoring time window, and is used to characterize the degree of rapid change in the internal stress state of the rock mass. The percentage of the area of the abnormal temperature region in the infrared thermal image refers to the ratio of the area of the pixel region whose temperature exceeds the historical average temperature threshold range to the total imaging area, calculated based on the infrared thermal imaging data collected within the current monitoring time window, and the average value is taken as the percentage of the area of the abnormal temperature region in the infrared thermal image. The statistical features extracted above are used to construct the statistical feature set of each monitoring unit.
[0021] This step extracts the above statistical features to comprehensively characterize the dynamic behavior of the rock mass from five dimensions: energy release, acoustic response, deformation characteristics, stress state, and thermal field distribution. These features together constitute the statistical feature set of each monitoring unit, providing multi-source feature support for subsequent cluster analysis and risk trend assessment.
[0022] S3, perform cluster analysis on the statistical feature set of each monitoring unit, and divide the monitoring units whose feature similarity index meets the preset threshold into the same unit cluster to obtain several unit clusters; In this embodiment, cluster analysis is performed on the statistical feature sets of each monitoring unit. Monitoring units whose feature similarity index meets a preset threshold are divided into the same unit cluster, thereby obtaining several unit clusters. The specific logic is as follows: For each monitoring unit, the statistical feature set within its current monitoring time window is taken, and the statistical feature set is used as the current feature vector of that monitoring unit. Calculate the Euclidean distance between the current feature vectors of any two monitoring units, and use the calculation result as the feature similarity index of the two monitoring units; The formula used to calculate the feature similarity index is as follows: In the formula, Indicates monitoring unit With monitoring unit The feature similarity index between two units is such that the smaller the value, the more similar the features of the two units are. This represents the total dimension of the current feature vector, i.e., the total number of extracted statistical features; and Representing monitoring units With monitoring unit The current feature vector at the th Numerical values on dimensional features An index for statistical features.
[0023] Dependent variable Used for characterization monitoring unit With monitoring unit The overall similarity in the multidimensional feature space, specifically, is reflected by calculating the Euclidean distance between two units across all statistical feature dimensions. It comprehensively reflects the overall differences between them in multi-source monitoring data such as microseismic, acoustic emission, displacement, stress, and infrared thermography. The smaller the value, the more similar the two units are in the dynamic response behavior of the rock mass, and they may belong to the same geological conditions or be in a similar stability state. Conversely, the larger the value, the more significant the differences in the characteristic performance of the two units, and they may correspond to different risk areas or mechanical behavior patterns, requiring independent assessment of their instability risk.
[0024] In this formula, the independent variable is... and Representing monitoring units and In the The numerical values of the statistical characteristics in the dimension affect the dependent variable through the sum of squared differences. Specifically, the differences in each feature dimension The similarity index is obtained by summing the squared values and then taking the square root. This mechanism significantly amplifies large differences on any feature dimension. The value of , i.e., feature differences have an amplifying effect in overall similarity assessment. Therefore, this method can sensitively capture local differences in key features such as energy release and displacement rate among different monitoring units, ensuring that the clustering results can accurately reflect the spatial differentiation characteristics of rock mass behavior.
[0025] Table 1: Feature Similarity Index Statistics Table It should be noted that in Table 1, F(a,1), F(a,2), F(a,3), F(b,1), F(b,2), F(b,3), and SI(a,b) correspond to... , and item; Analysis of the similarity index among the 15 monitoring units in Table 1 revealed that the monitoring units within the target area exhibited significant spatial differentiation in multiple characteristics such as microseismic activity, acoustic emission, displacement, stress, and infrared thermography. Most monitoring unit pairs, such as groups 1, 6, 10, and 14, showed high similarity, indicating that these areas share a consistent dynamic response behavior in the rock mass and can be grouped into the same risk cluster for unified assessment. However, some unit pairs, such as groups 3, 5, 8, and 13, showed relatively significant differences in characteristics, revealing the unique characteristics of local areas in terms of stress concentration, microfracture activity, or temperature field anomalies, requiring close attention to their independent evolution trends. This characteristic distribution pattern validates the necessity of using cluster analysis to achieve refined risk zoning, laying a reliable foundation for subsequent calculation of the risk trend coefficient and instability probability prediction of each unit cluster.
[0026] Set similarity threshold and minimum cluster size The initial division of unit clusters is completed through the following steps: All monitoring units are aggregated to form a unit set, and all monitoring units in the unit set are defined as in an inaccessible state. Monitoring units are then randomly selected from the unit set. Based on the feature similarity index, the monitoring units are determined by traversing the unit set. neighborhood Specifically, if there exists a monitoring unit in the unit set whose status is unvisited, and this unit is related to unit... The feature similarity index between them is less than the similarity threshold. This unit will then be included in the monitoring unit. neighborhood middle; If monitoring unit neighborhood The number of internal cells is less than the minimum cluster size If the unvisited cell is selected again from the cell set, then the cell will be monitored; otherwise, the cell will be monitored. As the core point, and to create unit clusters Unit clustering By monitoring unit and neighboring areas Composition, and based on clustering Real-time updates of user status within a cell set, specifically: incorporating users from the cell set into clusters. The cell status has been changed to "accessed"; Another unvisited monitoring unit is randomly selected from the unit set to build a new cluster, until no new cluster can be built. This completes the initial division of the unit set. The initial clustering is then optimized by merging the clusters: if the minimum distance between two cluster centers is less than the merging threshold... If so, these two clusters will be merged into a new cluster; the cluster center refers to the mean vector of all monitoring unit feature vectors within each unit cluster; Finally obtained Clustering of individual units ,in, The number of unit clusters, ; And clustering for any unit Filter out clusters with units The monitoring unit with the smallest feature similarity index among the cluster centers is defined as the unit cluster. The representative unit is identified, and the statistical characteristics of this monitoring unit are labeled as unit clustering. Statistical characteristics.
[0027] The similarity threshold This method is used to determine whether two monitoring units have sufficient similarity in the statistical feature space to be classified into the same neighborhood. Its value is determined based on the distribution characteristics of the pairwise feature similarity index (Euclidean distance) between all monitoring units. Specifically, first, the Euclidean distance set of all monitoring unit combinations is calculated, and a certain quantile of this set is taken as the benchmark value. The quantile is typically selected between 15% and 35%. Lower quantiles result in finer clustering and more partitions, while higher quantiles result in coarser clustering and fewer partitions. For cases where the data distribution is relatively concentrated, the "elbow rule" can also be used to help determine the optimal threshold. This involves plotting a curve showing the change in the number of clusters under different distance thresholds and selecting the distance value corresponding to the inflection point of the curve as the benchmark. Furthermore, when prior statistical data is lacking, an adaptive method can also be used to... The empirical coefficient is set to the mean of the standard deviations of each dimension of the feature vector of all monitoring units multiplied by an empirical coefficient, wherein the empirical coefficient ranges from 0.5 to 1.5.
[0028] Minimum cluster size The value of the clustering unit used to determine the core points and filter noise is related to the total number of monitoring units and the engineering geological complexity of the target area. It is generally set to 1% to 5% of the total number of monitoring units, and not less than 3%. For cold-region rock masses with complex geological conditions and strong spatial variability, a lower proportion (such as 1% to 2%) should be used to avoid misjudging small-scale but high-risk feature areas as noise. For areas with relatively uniform geological conditions, a higher proportion (such as 3% to 5%) can be used to enhance the statistical stability of clustering. In practical engineering applications, the contour coefficient of the clustering results under different parameter combinations can be compared through multiple experiments, and the parameter combination with the largest contour coefficient can be selected as the final setting value.
[0029] Step S3 uses a clustering algorithm based on Euclidean distance to divide monitoring units with similar multi-source statistical characteristics into several unit clusters. This step effectively identifies regions with similar response characteristics within the rock mass due to differences in geological conditions, mechanical state, or distribution of glacial joints. This divides the originally continuous and heterogeneous rock mass region into a finite number of risk assessment units with internal homogeneity and interval differences. This provides a reasonable spatial basis for subsequent differentiated trend analysis and risk probability calculation for different clusters.
[0030] S4. For each unit cluster, analyze the changes in its statistical feature set within a preset time window, and calculate the risk trend coefficient of the unit cluster within the preset time window. In this embodiment, for each unit cluster, the changes in its statistical feature set within a preset time window are analyzed. The preset time window consists of the most recent consecutive... The cluster consists of several monitoring time windows. The risk trend coefficient of the cluster within the preset time window is calculated based on the following logic: For each unit cluster, select its nearest contiguous cluster. The statistical feature set of each monitoring time window, among which For each statistical feature, extract its value in this context. The representation values within each monitoring time window form a length of [value]. The statistical characteristics of this time series subsequence; The slope of the temporal subsequence of each statistical feature in the cluster is calculated over time using a linear fitting method. The underlying logic is as follows: for any statistical feature in each unit cluster, its most recent continuous... Characteristic values within a monitoring time window As the dependent variable sequence, and index the corresponding time window. As the sequence of independent variables, construct a set of lengths... The time series data pairs were then used; subsequently, the least squares method was employed to perform linear regression on the time series, resulting in a fit of the form shown in the figure. The linear equation, where the obtained regression coefficients This refers to the trend slope of the statistical feature within the current time window. Its magnitude and sign represent the rate and direction of change of the feature over time, respectively.
[0031] The method for determining the risk trend coefficient is as follows: First, the trend slopes of each statistical feature are standardized to eliminate differences in their numerical magnitudes. Then, each standardized trend slope is multiplied by its corresponding weight coefficient. Finally, all weighted results are summed, and the resulting value is designated as the risk trend coefficient for that cluster. The specific formula used is as follows: In the formula, This represents the risk trend coefficient of the cluster. Indicates the first The preset weighting coefficients corresponding to the slope of each statistical characteristic trend satisfy the following conditions. This method characterizes the importance of different statistical features in contributing to the risk of rock mass instability. Its values are based on the correlation strength between each feature and instability events in historical monitoring data, and are determined through objective calculation using the entropy weight method or by combining domain knowledge with expert experience. First, the objective weight is calculated using the entropy weight method, assigning weights based on the degree of variation of each statistical feature in historical monitoring data; the greater the variation in feature values, the more information is contained and the higher the weight. Then, the subjective weight is calculated using the analytic hierarchy process (AHP). Based on the mechanical mechanism that microseismic and acoustic emission are the most direct precursors of instability, followed by displacement, then stress, and finally infrared thermography as an auxiliary criterion, experts compare and score each feature pairwise to determine its relative importance. Finally, the objective weight and the subjective weight are multiplied and normalized to obtain the final value. When historical data is insufficient, subjective weighting can be used temporarily. The value will be dynamically updated based on accumulated data. and These represent the mean and standard deviation of the trend slopes of all statistical characteristics of the cluster in this unit, respectively.
[0032] For this formula, the dependent variable This is used to characterize the evolution trend and direction of the overall stability state of unit clustering within a preset time window. Specifically, By combining the weighted standardized values of the trend slopes of various statistical characteristics, the concentrated change tendency of the dynamic behavior of the rock mass in the time dimension is reflected. The larger the value is, the faster the risk of the cluster area is rising and the greater the possibility of instability. The smaller the value is, the slower the risk is falling and the stability is improving. If it is close to zero, it reflects that the state of the area is relatively stable and no obvious systematic change trend has appeared.
[0033] Independent variable in the formula The dependent variable is determined jointly by standardization and weighted summation. Specifically, the slope of each feature is first standardized to eliminate dimensional differences, then multiplied by a corresponding preset weight and summed. This ensures that the slopes of features with large variations and high weights are more effective in controlling the overall slope. The contribution of any feature is more significant; that is, if any feature shows a sharp upward or downward trend and its weight is high, it will directly raise or lower the overall risk trend coefficient. Therefore, It can sensitively capture the dynamic changes of key driving characteristics, thereby achieving a comprehensive quantification of regional risk evolution trends.
[0034] Step S4 involves clustering each unit to analyze the changes in its statistical feature set within a preset time window and calculating the risk trend coefficient. This allows for a quantitative assessment of the evolution direction of rock mass stability; by extracting the temporal variation slope of each statistical feature and performing standardization and weighted fusion, This method comprehensively reflects the overall risk change trend of the rock mass across multiple monitoring dimensions: positive values indicate rising risk, while negative values indicate mitigating risk. This approach not only enhances the foresight of early warnings, enabling the system to identify potential areas of accelerated risk, but also provides key dynamic feature inputs for subsequent machine learning models, improving the timeliness and accuracy of early warnings for instability of icy jointed rock masses in cold regions.
[0035] S5, input the statistical feature set and risk trend coefficient of each unit cluster into the pre-trained machine learning model to obtain the real-time instability risk probability of each unit cluster; In this embodiment, for each unit cluster, its statistical feature set and risk trend coefficient are fused to construct a comprehensive feature vector for input to the machine learning model; the comprehensive feature vector consists of two parts: one part is based on the nearest continuous cluster of that unit cluster. The time-series statistics are calculated from the statistical feature set within a monitoring time window, specifically including the mean, standard deviation, maximum and minimum values of each statistical feature within that time period; another part is the risk trend coefficient of the clustering of that unit. The comprehensive feature vector is input into a pre-trained machine learning model, which is trained using a gradient boosting decision tree algorithm. Its output is a continuous probability value between 0 and 1, which is the real-time instability risk probability of unit clustering. The training process of the machine learning model specifically includes: collecting historical monitoring data of the target cold region or similar geological conditions region; using the statistical feature set of each unit cluster extracted from the historical data; using the statistical feature set and risk trend coefficient as the model input features; and using the rock mass instability state actually observed in the corresponding time period as the training label, wherein the stable state is marked as 0 and the unstable state is marked as 1. The unstable state includes observable instability phenomena such as rock mass sliding, collapse, and crack propagation. The gradient boosting decision tree algorithm is used for model training, and the model parameters are optimized by the ten-fold cross-validation method. The grid search strategy is used to find the optimal tree depth, learning rate and number of trees in the preset parameter space. The preset parameter space includes: tree depth ranging from 3 to 10, learning rate ranging from 0.01 to 0.2, and number of trees ranging from 50 to 200. Finally, a trained machine learning early warning model is obtained.
[0036] S6. The instability risk probability of each cluster is compared with the preset probability threshold range to determine the instability risk level of each cluster and generate a visual early warning image of the instability risk level covering the target cold-region icy jointed rock mass area. In this embodiment, four consecutive probability threshold intervals are preset, namely the risk-free interval, the low-risk interval, the medium-risk interval, and the high-risk interval, which correspond to the risk-free level, the low-risk level, the medium-risk level, and the high-risk level, respectively. The real-time instability risk probability of each unit cluster is compared with the above-mentioned preset probability threshold intervals to determine the instability risk level corresponding to each unit cluster. The probability threshold range is set based on historical instability event statistics and expert experience, and is specifically divided as follows: Risk-free zone: The probability value is located in This indicates that the current clustering of units is in a stable state with no obvious signs of instability. Low-risk range: probability value is within This indicates that the current unit clustering has a slight risk of instability and requires continued monitoring. Medium risk range: probability value is located in This indicates that the current unit clustering has a moderate risk of instability, and monitoring should be strengthened and preliminary preventive measures should be taken. High-risk range: probability value is located in This indicates that the current unit clustering is at high risk of instability, and an emergency response mechanism needs to be activated immediately.
[0037] Based on the cluster location coordinates of each unit cluster and the determined instability risk level, a color-coded rendering method is used to mark clusters of different risk levels with corresponding characteristic colors on the geographic base map of the target cold-region glacial jointed rock mass area, generating a visual early warning image of instability risk level covering the entire target area; the color-coded rendering method specifically is as follows: The risk-free level is marked in green; Low-risk levels are marked in yellow; Medium-risk levels are marked in orange. High-risk levels are marked in red; Simultaneously, corresponding early warning signals are triggered for clusters at medium-risk and above levels. These early warning signals include risk level indications, cluster location coordinates, and descriptions of risk development trends. The generation and release mechanism for these early warning signals includes: When a unit cluster is determined to be of medium risk level, the system automatically generates a level 2 warning, prompting on-site management personnel to strengthen patrols and record the location coordinates and risk trend description of the cluster; When a cluster is identified as high-risk, the system automatically generates a Level 1 warning and notifies relevant personnel through audible and visual alarms, SMS push notifications, or platform pop-ups. At the same time, the emergency plan is activated, providing detailed cluster location coordinates, risk level, risk trend descriptions, and suggested handling measures.
[0038] The spatial coordinates of each cluster are represented by the spatial coordinates of the representative unit corresponding to that cluster. The risk development trend description is generated based on the recent changes in the risk trend coefficient and its statistical feature set of the clustering unit, including the direction of risk change, the rate of risk change, and the analysis of the main driving factors; specifically: when the risk trend coefficient is positive, it indicates that the risk is on an upward trend, and when it is negative, it indicates that the risk is on a downward trend; the rate of risk change is determined by calculating the change in the risk trend coefficient within the two most recent monitoring time windows; and the main driving factors are determined by identifying the statistical features with the largest changes in the statistical feature set.
[0039] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.
[0040] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.
[0041] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.
[0042] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A method for early warning of instability of glacial jointed rock masses in cold regions based on multi-source data fusion, characterized in that, The specific steps include: S1. The target cold-region glacial jointed rock mass area is divided into several monitoring units. Multi-source heterogeneous time-series data of each monitoring unit are collected, including microseismic monitoring data, acoustic emission data, surface displacement data, rock mass stress data and infrared thermal imaging data. The collected data are preprocessed to construct a standardized fusion dataset. S2, based on the standardized fusion dataset, extracts the energy and frequency of microseismic events, the ringing count and energy rate of acoustic emission events, the surface displacement rate and cumulative displacement, the rock mass stress change gradient, and the area ratio of infrared thermal image temperature anomalies in each monitoring unit within the current monitoring time window, thus forming the statistical feature set of each monitoring unit. S3, perform cluster analysis on the statistical feature set of each monitoring unit, and divide the monitoring units whose feature similarity index meets the preset threshold into the same unit cluster to obtain several unit clusters; S4. For each unit cluster, analyze the changes in its statistical feature set within a preset time window, and calculate the risk trend coefficient of the unit cluster within the preset time window. S5, input the statistical feature set and risk trend coefficient of each unit cluster into the pre-trained machine learning model to obtain the real-time instability risk probability of each unit cluster; S6 compares the instability risk probability of each cluster with the preset probability threshold range to determine the instability risk level of each cluster and generate a visual early warning image of the instability risk level covering the target cold-region icy jointed rock mass area.
2. The method for early warning of instability of glacial jointed rock masses in cold regions based on multi-source data fusion according to claim 1, characterized in that: The specific execution process of S1 is as follows: The target cold region containing ice-bearing jointed rock mass is evenly divided into multiple monitoring units. The geometric center of each monitoring unit is used as a monitoring node. Multi-source heterogeneous time series data are collected at each monitoring node, and the multi-source heterogeneous time series data at the monitoring node is labeled as the multi-source heterogeneous time series data of the corresponding monitoring unit. Microseismic monitoring data are collected from each monitoring node using microseismic sensors; Acoustic emission data from each monitoring node is collected using acoustic emission probes. Ground displacement data of each monitoring node is collected using GNSS positioning equipment; Rock mass stress data at each monitoring node are collected using stress sensors; Infrared thermal imaging data of each monitoring node are collected using an infrared thermal imager; The collected data underwent preprocessing, including outlier removal and missing value imputation. Outlier removal employed the Laida criterion, and the mean of each data category was calculated. with standard deviation The normal fluctuation range of the data is determined as [ Data that deviates from the normal fluctuation range will be identified as outliers and removed. Missing value completion uses linear interpolation: for each missing data point, all similar data points that were not removed are counted if their timestamp interval is less than a preset time, and the average value is taken as the completion value to complete the missing value. Then, the preprocessed multi-source heterogeneous time series data of each monitoring node are aligned and integrated according to timestamp to construct a standardized fusion dataset covering the target cold region.
3. The method for early warning of instability of glacial jointed rock masses in cold regions based on multi-source data fusion as described in claim 1, characterized in that: The energy of the microseismic event is obtained by calculating the arithmetic mean of the energies of all microseismic events falling within the current monitoring time window; the frequency of the microseismic event is obtained by counting all microseismic events falling within the current monitoring time window. The acoustic emission event ringing count is obtained by counting the total number of times the waveform signals of all acoustic emission events falling within the current monitoring time window exceed a preset voltage threshold; the acoustic emission event energy rate is obtained by calculating the sum of the energies of all acoustic emission events falling within the current monitoring time window. The surface displacement rate is obtained by calculating the resultant displacement change of surface displacement data in three dimensions within the current monitoring time window; the cumulative displacement is obtained by summing all the resultant displacement changes obtained from the start of monitoring to the end of the current monitoring time window. The rock mass stress change gradient is obtained by calculating the absolute value of the maximum principal stress change in the rock mass stress data within the current monitoring time window. The percentage of the area of the abnormal temperature region in the infrared thermal image refers to the ratio of the area of the pixel region whose temperature exceeds the historical average temperature threshold range to the total imaging area, calculated based on the infrared thermal imaging data collected within the current monitoring time window, and the average value is taken as the percentage of the area of the abnormal temperature region in the infrared thermal image. The statistical features extracted above are used to construct the statistical feature set of each monitoring unit.
4. The method for early warning of instability of glacial jointed rock mass in cold regions based on multi-source data fusion according to claim 3, characterized in that: Cluster analysis is performed on the statistical feature sets of each monitoring unit. Monitoring units whose feature similarity index meets a preset threshold are grouped into the same cluster, thus obtaining several unit clusters. The specific logic behind this is as follows: For each monitoring unit, the statistical feature set within its current monitoring time window is taken, and the statistical feature set is used as the current feature vector of that monitoring unit. Calculate the Euclidean distance between the current feature vectors of any two monitoring units, and use the calculation result as the feature similarity index of the two monitoring units; Set similarity threshold and minimum cluster size The initial division of unit clusters is completed through the following steps: All monitoring units are aggregated to form a unit set, and all monitoring units in the unit set are defined as in an inaccessible state. Monitoring units are then randomly selected from the unit set. Based on the feature similarity index, the monitoring units are determined by traversing the unit set. neighborhood Specifically, if there exists a monitoring unit in the unit set whose status is unvisited, and this unit is related to unit... The feature similarity index between them is less than the similarity threshold. This unit will then be included in the monitoring unit. neighborhood middle; If monitoring unit neighborhood The number of internal cells is less than the minimum cluster size If the unvisited monitoring unit is selected from the unit set, then the monitoring unit will be reselected; otherwise, the monitoring unit will be... As the core point, and to create unit clusters Unit clustering From unit and neighboring areas Composition, and based on clustering Real-time updates of user status within a cell set, specifically: incorporating users from the cell set into clusters. The cell status has been changed to "accessed"; Another unvisited monitoring unit is randomly selected from the unit set to build a new cluster, until no new cluster can be built. This completes the initial division of the unit set.
5. The method for early warning of instability of glacial jointed rock masses in cold regions based on multi-source data fusion according to claim 4, characterized in that: Merge and optimize the initially divided clusters: if the minimum distance between the cluster centers of two units is less than the merging threshold... If the two clusters are not clustered, then the two clusters will be merged into a new cluster. Finally obtained Clustering of individual units ,in, The number of unit clusters, ; And clustering for any unit Filter out clusters with units The monitoring unit with the smallest feature similarity index among the cluster centers is defined as the unit cluster. The representative unit is identified, and the statistical characteristics of this monitoring unit are labeled as unit clustering. Statistical characteristics.
6. The method for early warning of instability of glacial jointed rock masses in cold regions based on multi-source data fusion according to claim 3, characterized in that: For each cluster, the changes in its statistical feature set within a preset time window are analyzed. The preset time window consists of the most recent consecutive clusters. The cluster consists of several monitoring time windows. The risk trend coefficient of the cluster within the preset time window is calculated based on the following logic: For each unit cluster, select its nearest contiguous cluster. The statistical feature set of each monitoring time window, among which For each statistical feature, extract its value in this context. The representation values within each monitoring time window form a length of [value]. The statistical characteristics of this time series subsequence; The slope of the trend of the time-series subsequence of each statistical feature of the cluster over time was calculated by linear fitting method; The method for determining the risk trend coefficient is as follows: First, the trend slopes of each statistical feature are standardized to eliminate differences in their numerical magnitudes. Then, each standardized trend slope is multiplied by its corresponding weight coefficient. Finally, all weighted results are summed, and the resulting value is designated as the risk trend coefficient of the cluster.
7. The method for early warning of instability of glacial jointed rock mass in cold regions based on multi-source data fusion according to claim 6, characterized in that: For each cluster, its statistical feature set and risk trend coefficient are fused to construct a comprehensive feature vector for input to the machine learning model; The comprehensive feature vector consists of two parts: one part is based on the nearest continuous cluster of the unit. The time-series statistics are calculated from the statistical feature set within a monitoring time window, specifically including the mean, standard deviation, maximum and minimum values of each statistical feature within that time period; another part is the risk trend coefficient of the clustering of that unit.
8. The method for early warning of instability of glacial jointed rock mass in cold regions based on multi-source data fusion according to claim 7, characterized in that: The comprehensive feature vector is input into a pre-trained machine learning model, which is trained using a gradient boosting decision tree algorithm. Its output is a continuous probability value between 0 and 1, which is the real-time instability risk probability of unit clustering. The training process of the machine learning model specifically includes: collecting historical monitoring data of the target cold region or similar geological conditions region, using the statistical feature set of each unit cluster extracted from the historical data as the model input feature set and the risk trend coefficient as the model input feature, and using the rock mass instability state actually observed in the corresponding time period as the training label, wherein the stable state is marked as 0 and the unstable state is marked as 1. The model parameters were optimized using a 10-fold cross-validation method. The optimal tree depth, learning rate, and number of trees were determined through grid search, resulting in a well-trained machine learning early warning model.
9. The method for early warning of instability of glacial jointed rock mass in cold regions based on multi-source data fusion according to claim 1, characterized in that: Four consecutive probability threshold intervals are preset, namely the risk-free interval, low-risk interval, medium-risk interval and high-risk interval, which correspond to the risk-free level, low-risk level, medium-risk level and high-risk level respectively; the real-time instability risk probability of each unit cluster is compared with the above preset probability threshold intervals to determine the instability risk level corresponding to each unit cluster. Based on the cluster location coordinates of each unit cluster and the determined instability risk level, a color-coded rendering method is used to mark clusters of different risk levels with corresponding characteristic colors on the geographic base map of the target cold-region glacial jointed rock mass area, generating a visual early warning image of instability risk level covering the entire target area. At the same time, corresponding early warning signals are triggered for clusters of medium risk and above. The early warning signals include risk level prompts, cluster location coordinates and risk development trend descriptions. The spatial coordinates of each cluster are represented by the spatial coordinates of the representative unit corresponding to that cluster.
10. A method for early warning of instability of glacial jointed rock masses in cold regions based on multi-source data fusion as described in claim 9, characterized in that: The risk development trend description is generated based on the recent changes in the risk trend coefficient and its statistical feature set of the clustering unit, including the direction of risk change, the rate of risk change, and the analysis of the main driving factors. Specifically: when the risk trend coefficient is positive, it indicates that the risk is on an upward trend; when it is negative, it indicates that the risk is on a downward trend. The rate of risk change is determined by calculating the change in the risk trend coefficient within the two most recent monitoring time windows; the main driving factors are determined by identifying the statistical features with the largest changes in the statistical feature set.