Geological disaster early-stage monitoring and early-warning method based on micro-deformation feature extraction and analysis

Abstract

The invention relates to the technical field of geological disaster monitoring and early warning, in particular to a geological disaster early monitoring and early warning method based on micro-deformation feature extraction and analysis. According to the method, micro-deformation data of a geological disaster area is collected through a multi-source cooperative monitoring network, and after preprocessing, a time series data set is generated by using a multi-source data fusion model based on an attention mechanism. Basic features and depth features based on improved LSTM are extracted, and a core feature set is formed through screening; and periodically training an early warning threshold model based on a historical core feature set of a sliding window, and outputting a dynamic early warning interval by adjusting a learning rate. And matching the current core feature set with the interval, generating a risk early warning signal, and sending the risk early warning signal to the emergency terminal. And then, according to the early warning verification rate, the risk missing report coefficient and the average effective response time, calculating an early warning characterization value so as to evaluate an early warning state, and according to the early warning characterization value, adaptively adjusting a model training period and a learning rate. According to the invention, the accuracy of geological disaster monitoring and early warning is improved.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and early warning technology, and in particular to an early monitoring and early warning method for geological disasters based on micro-deformation feature extraction and analysis. Background Technology

[0002] Geological disasters are often accompanied by micro-deformation processes of the earth's surface and rock mass. Accurate capture and analysis of early micro-deformation characteristics are key to achieving early warning of disasters. Existing geological disaster monitoring and early warning technologies are mainly divided into two categories: ground-based contact monitoring and space-based remote sensing monitoring.

[0003] However, existing technologies have several shortcomings: First, traditional ground-based contact monitoring methods have the drawbacks of limited monitoring range and point-to-area coverage, making it difficult to comprehensively cover a wide range of disaster hazard points. Furthermore, they are difficult to deploy and costly to maintain in harsh terrain conditions, and are susceptible to environmental interference, leading to data loss or excessive errors. Second, existing micro-deformation feature analysis often uses single-dimensional data and does not fully integrate multi-source monitoring information. Moreover, feature recognition algorithms are poorly adaptable to nonlinear and uncertain deformation time-series features, resulting in rigid warning threshold settings and a tendency for false alarms and missed alarms. Third, existing early warning systems are mostly based on fixed threshold triggering modes, and cannot dynamically adjust warning standards according to differences in the geological environment and seasonal changes of disaster hazard points, resulting in insufficient targeting and reliability of warnings.

[0004] Chinese Patent Publication No. CN119580474A discloses a landslide geological disaster monitoring and early warning system based on multi-method fusion, including a data acquisition module, a data fusion module, a landslide prediction module, a landslide disaster intelligent zoning and response module, and an early warning signaling module. The data acquisition module collects landslide-related data in real time; the data fusion module fuses landslide-related data from different sources; the landslide prediction module predicts the probability of landslide occurrence; the landslide disaster intelligent zoning and response module combines real-time monitored landslide-related data and geological structure, and dynamically adjusts post-disaster response measures to generate early warning information; the early warning signaling module automatically issues early warning signals based on the generated early warning information.

[0005] It is evident that existing technologies suffer from the following problems: inaccurate micro-deformation feature analysis and poor feature recognition adaptability lead to inaccurate geological disaster monitoring and early warning. Summary of the Invention

[0006] Therefore, this invention provides a geological disaster early monitoring and warning method based on micro-deformation feature extraction and analysis, in order to overcome the problem that the existing technology has inaccurate geological disaster monitoring and warning due to inaccurate micro-deformation feature analysis and poor feature recognition adaptability.

[0007] To achieve the above objectives, this invention provides a method for early monitoring and warning of geological disasters based on micro-deformation feature extraction and analysis, comprising: A multi-source collaborative micro-deformation monitoring network is used to collect multi-source micro-deformation data of the monitoring area; The multi-source micro-deformation data is preprocessed, and the preprocessed multi-source data is fused using a multi-source data fusion model based on an attention mechanism to generate a micro-deformation time series dataset. The basic and deep features are extracted from the micro-deformation time series dataset, and the basic and deep features are filtered using a filtering algorithm to generate a core feature set. The deep features are extracted based on at least an improved LSTM neural network model, which is an improved LSTM neural network model that introduces a multi-head self-attention mechanism. The early warning threshold model is periodically trained based on the multi-source micro-deformation data extracted from historical moments using a preset sliding window, and the core feature set is input into the trained early warning threshold model to output dynamic early warning intervals for different risk levels. The learning rate of the early warning threshold model is adjusted at least during training. Based on the current core feature set and the dynamic early warning interval, a warning signal corresponding to the risk level is generated and output to each emergency response terminal. The early warning characterization value is determined based on the obtained early warning verification rate, risk underreporting coefficient, and average effective response time; The warning status is determined based on the warning characterization value, and the training period of the warning threshold model is adjusted based on the warning status. The learning rate of the warning threshold model is then adjusted based on the adjusted training period.

[0008] Furthermore, the multi-source collaborative micro-deformation monitoring network includes an airborne monitoring platform, ground-based monitoring sensors, and underground monitoring sensors, wherein data transmission in the monitoring network adopts dual-mode redundant communication technology.

[0009] Furthermore, the early warning verification rate V is ,in, The actual number of warnings confirmed through on-site verification. The total number of warnings issued by the system; the risk underreporting coefficient O is... ,in, The potential severity level of the i-th disaster event. Let be the time difference between the occurrence of the i-th disaster event and the successful detection of the anomaly. Let R be the total time from the inception to the occurrence of the i-th disaster event; the average effective response time R is... ,in, This is the average time difference from when the system issues a warning to when the deformation enters the acceleration phase. The preset response time is; the warning characterization value M is... ,in, These are the weight coefficients for the corresponding parameters, and the sum of the weight coefficients is ensured to be 1.

[0010] Furthermore, the process of determining the warning status based on the warning characterization value includes: comparing the warning characterization value with a preset warning characterization value; if the warning characterization value is greater than or equal to the preset warning characterization value, the warning status is determined to be qualified; if the warning characterization value is less than the preset warning characterization value, the warning status is determined to be unqualified.

[0011] Furthermore, the process of adjusting the training period of the warning threshold model based on the warning status includes: obtaining the warning verification rate at multiple historical moments when the warning status is unqualified; calculating the average value of the multiple warning verification rates; if the average value is less than the preset average value, adjusting the training period of the warning threshold model based on the ratio of the average value to the preset average value.

[0012] Furthermore, the process of adjusting the training period of the warning threshold model based on the ratio of the average value to the preset average value includes: reducing the training period of the warning threshold model based on the ratio of the average value to the preset average value, and the reduction in the training period is inversely proportional to the ratio.

[0013] Furthermore, the process of adjusting the learning rate of the warning threshold model based on the adjusted training period includes: increasing the learning rate of the warning threshold model by the ratio of the adjusted training period to the preset training period, and the increase in the learning rate is proportional to the ratio.

[0014] Furthermore, the method also includes: if the control state is unqualified after adjusting the learning rate of the warning threshold model, obtaining the data time-lapse coefficient recorded by the warning threshold model at the beginning of the training cycle; if the data time-lapse coefficient within the window is greater than a preset value, adjusting the preset sliding window based on the ratio of the data time-lapse coefficient within the window to the preset value.

[0015] Furthermore, the process of adjusting the preset sliding window based on the ratio of the data time-related decay coefficient within the window to the preset value includes: reducing the preset sliding window based on the ratio of the data time-related decay coefficient within the window to the preset value, and the reduction of the preset sliding window is proportional to the ratio.

[0016] Furthermore, the method further includes: if the warning status is unqualified after adjusting the preset sliding window, repeatedly adjusting the preset sliding window at least once until the number of adjustments is less than the preset number and the warning status is qualified, or the number of adjustments is equal to the preset number and the adjustment is stopped; if the warning status is unqualified after stopping the adjustment, calculate the difference between the preset warning characterization value and the warning characterization value corresponding to multiple historical times; calculate the average and variance of the multiple differences, if the average is greater than the preset difference and the variance is less than the preset variance, then reduce the calibration cycle of the sensor in the monitoring network based on the ratio of the variance to the preset variance, and the reduction in the calibration cycle is inversely proportional to the ratio.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention collects micro-deformation data of geological disaster areas through a multi-source collaborative monitoring network. After preprocessing, a time-series dataset is generated using a multi-source data fusion model based on an attention mechanism. Basic features and deep features based on an improved LSTM are extracted and filtered to form a core feature set. An early warning threshold model is periodically trained based on the historical core feature set using a sliding window, and a dynamic early warning interval is output by adjusting the learning rate. The current core feature set is matched with the interval to generate a risk early warning signal and send it to the emergency terminal. Subsequently, an early warning characterization value is calculated based on the early warning verification rate, risk miss coefficient, and average effective response time to assess the early warning status, and the model training cycle and learning rate are adaptively adjusted accordingly. This invention improves the accuracy of geological disaster monitoring and early warning.

[0018] Furthermore, this invention monitors geological conditions using multi-source sensors and transmits data using dual-redundant communication technology, enabling more accurate acquisition of multi-source micro-deformation data and more comprehensive coverage of wide-area disaster hazard points, thereby further improving the accuracy of geological disaster monitoring and early warning.

[0019] Furthermore, this invention determines the early warning characterization value based on the weighted sum of the early warning verification rate, the risk underreporting coefficient, and the average effective response time. This allows for a more accurate determination of the early warning characterization value, thereby more accurately determining the early warning status and further improving the accuracy of geological disaster monitoring and early warning.

[0020] Furthermore, the present invention determines the warning status based on the comparison between the warning characterization value and the preset warning characterization value, which can more accurately determine the warning status and thus further improve the accuracy of geological disaster monitoring and early warning.

[0021] Furthermore, this invention determines the reasons for unqualified early warning status based on the average of the early warning verification rate at multiple historical moments. This allows for a more accurate determination of whether false alarms are caused by the inability of the dynamic early warning interval constructed based on historical data to adapt to new deformation patterns. Consequently, the reasons for unqualified early warning status can be used to more effectively adjust relevant parameters and further improve the accuracy of geological disaster monitoring and early warning.

[0022] Furthermore, this invention adjusts the training period of the early warning threshold model based on the ratio of the average value to the preset average value, which can more accurately adjust the training period and make the early warning threshold model more adaptable to the current geological disaster situation, thereby further improving the accuracy of geological disaster monitoring and early warning.

[0023] Furthermore, this invention increases the learning rate of the early warning threshold model by adjusting the ratio of the training period after the model is adjusted to the preset training period. This effectively increases the learning rate, enabling the model to adapt to minor changes in new data more quickly, thereby further improving the accuracy of the dynamic early warning interval output by the early warning threshold model, and further improving the accuracy of geological disaster monitoring and early warning.

[0024] Furthermore, this invention determines the reasons for unqualified early warning status based on the data time-effect decay coefficient recorded within the window at the beginning of the training cycle of the acquired early warning threshold model. It can determine whether the model is insensitive to recent changes due to excessive historical information, resulting in delayed or invalid early warnings. This further improves the accuracy of the dynamic early warning interval output by the early warning threshold model, and thus further improves the accuracy of geological disaster monitoring and early warning.

[0025] Furthermore, this invention reduces the preset sliding window based on the ratio of the data time-effect decay coefficient within the window to a preset value, which can more accurately adjust the preset sliding window, thereby making the extracted core feature set of historical moments more effective. This, in turn, enables the early warning threshold model to more effectively output dynamic early warning intervals for different risk levels, further improving the accuracy of geological disaster monitoring and early warning.

[0026] Furthermore, the average value and variance of the difference between the preset warning characterization value and the warning characterization value corresponding to multiple historical moments in this invention determine the reason for the failure of the warning state. This can more accurately determine whether the failure of the warning state is due to the offset or failure of the sensors in the monitoring network, thereby further improving the accuracy of geological disaster monitoring and early warning. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the structure of the geological disaster early monitoring and warning system based on micro-deformation feature extraction and analysis according to an embodiment of the present invention; Figure 2This is a flowchart illustrating the steps of the geological disaster early monitoring and warning method based on micro-deformation feature extraction and analysis, as described in an embodiment of the present invention. Figure 3 This is a flowchart illustrating the steps of determining the warning characterization value based on the comparison result between the warning characterization value and the preset warning characterization value in an embodiment of the present invention. Figure 4 This is a flowchart illustrating the steps of determining the warning status based on adjusting a preset sliding window, as described in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0029] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0030] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0031] Please see Figure 1 As shown, it is a schematic diagram of the structure of the geological disaster early monitoring and warning system based on micro-deformation feature extraction and analysis according to an embodiment of the present invention.

[0032] The system includes an acquisition unit, a preprocessing unit, a filtering unit, an output unit, a matching unit, a calculation unit, and an analysis unit.

[0033] The acquisition unit collects multi-source micro-deformation data of the monitoring area based on a multi-source collaborative micro-deformation monitoring network; The preprocessing unit is connected to the acquisition unit and is used to preprocess the multi-source micro-deformation data and fuse the preprocessed multi-source data based on the multi-source data fusion model with attention mechanism to generate a micro-deformation time series dataset. The filtering unit is connected to the preprocessing unit and is used to extract basic features and deep features from the micro-deformation time series dataset. It also uses a filtering algorithm to filter the basic features and deep features to generate a core feature set. The deep features are extracted based on at least an improved LSTM neural network model, which is an improved LSTM neural network model that introduces a multi-head self-attention mechanism. The output unit is connected to the acquisition unit and the filtering unit respectively. It is used to periodically train the early warning threshold model based at least on the multi-source micro-deformation data extracted from historical moments using a preset sliding window, and input the core feature set into the trained early warning threshold model to output dynamic early warning intervals of different risk levels. The learning rate of the early warning threshold model is adjusted at least during training. The matching unit is connected to the filtering unit and the output unit respectively. It is used to match the current core feature set with the dynamic early warning interval, generate an early warning signal with the corresponding risk level, and output it to each emergency response terminal. The calculation unit is connected to the matching unit and is used to determine the early warning characterization value based on the acquired early warning verification rate, risk underreporting coefficient and average effective response time. The analysis unit is connected to the calculation unit and is used to determine the warning status based on the warning characterization value, adjust the training period of the warning threshold model based on the warning status, and adjust the learning rate of the warning threshold model based on the adjusted training period.

[0034] Specifically, constructing a multi-source collaborative micro-deformation monitoring network includes: selecting geological hazard hazard points and surrounding areas as monitoring target areas, and deploying a multi-source collaborative monitoring network composed of airborne, ground-based, and underground three-dimensional monitoring methods. The airborne monitoring platform uses the Sentinel-1 satellite equipped with SBAS-InSAR technology, setting a 12-day revisit cycle to acquire large-scale surface deformation data of the monitoring target area. Short baseline set interferometry is used to eliminate atmospheric delay and spatiotemporal incoherence effects, and preliminary extraction of macroscopic micro-deformation trends of the surface is achieved. Ground-based monitoring sensors include high-precision micro-deformation radar, a binocular vision monitoring system, and fiber optic sensor arrays deployed in key areas of the hazard points. The micro-deformation radar is set to a sampling frequency of 200 times per second, achieving real-time micro-deformation monitoring with an accuracy of 0.01mm. The binocular vision system achieves non-contact quantitative monitoring of surface deformation through a sliding rail mounting, and outputs planar deformation data in combination with deep learning algorithms. The fiber optic sensor array is deployed along the potential sliding surface to collect micro-strain data inside the rock mass. The underground monitoring sensor consists of miniature displacement sensors drilled at potential hazard points to monitor micro-displacement data of rock masses at different underground depths. The sampling cycle is dynamically adjusted to 1-5 minutes / time according to the geological environment. Each monitoring platform achieves real-time data transmission through 5G and Beidou dual-mode communication modules, and the transmission process uses encryption protocols to ensure data security.

[0035] Specifically, the preprocessing unit involves cleaning the collected multi-source deformation data, filling in missing data using random forest imputation, and eliminating noise interference through Kalman filtering. It also standardizes the data across different dimensions, unifying the data units and time reference. Then, by constructing a multi-source data fusion model based on an attention mechanism, it weights and fuses macroscopic deformation data from airborne monitoring, high-precision surface deformation data from ground-based monitoring, and internal micro-displacement data from underground monitoring. The weights are dynamically adjusted based on the monitoring accuracy and environmental adaptability of each monitoring unit, outputting a fused micro-deformation time-series dataset. This process is existing technology and will not be elaborated further.

[0036] Specifically, the basic feature extraction in the screening unit involves extracting basic feature parameters from the fused micro-deformation time series data, including cumulative deformation, deformation rate, deformation acceleration, and deformation stability coefficient. Deep feature extraction involves introducing multifractal detrending fluctuation analysis to process the micro-deformation time series data, calculating the multifractal spectral width Δa and spectral skewness Δf(a), and capturing nonlinear feature mutations during the micro-deformation process. Simultaneously, an improved LSTM neural network model is used for feature mining of the time series data. This improved LSTM neural network model enhances the identification ability of key deformation nodes by introducing a multi-head self-attention mechanism, outputting deep feature vectors representing early disaster precursors. Finally, the Relief-F algorithm is used to screen the basic feature parameters and deep feature vectors, eliminating redundant features to generate a core feature set.

[0037] Specifically, the output unit collects historical micro-deformation data and disaster occurrence records under different geological types such as landslides, collapses, and debris flows, as well as different environmental conditions such as rainfall, earthquakes, and seasonal changes, to construct a feature-disaster association sample library. Based on the sample library, a dynamic early warning threshold model is constructed using a support vector machine combined with a Bayesian optimization algorithm. The core feature set after screening is input into the early warning threshold model, and the output is a dynamic threshold range for different risk levels such as safe, attention, warning, and emergency. The model can adaptively adjust the threshold boundaries of each risk level according to the feature changes of real-time monitoring data.

[0038] Please see Figure 2 The diagram shown is a flowchart illustrating the steps of an early monitoring and warning method for geological disasters based on micro-deformation feature extraction and analysis, according to an embodiment of the present invention.

[0039] The specific steps of the early monitoring and warning of geological disasters based on micro-deformation feature extraction and analysis in this embodiment of the invention are as follows: S1, multi-source micro-deformation data of the monitoring area are collected by the acquisition unit based on a multi-source collaborative micro-deformation monitoring network; S2, the multi-source micro-deformation data is preprocessed by the preprocessing unit connected to the acquisition unit, and the preprocessed multi-source data is fused by the multi-source data fusion model based on the attention mechanism to generate a micro-deformation time series dataset. S3, the basic features and deep features in the micro-deformation time series dataset are extracted by the filtering unit connected to the preprocessing unit, and the basic features and deep features are filtered by the filtering algorithm to generate a core feature set. The deep features are extracted based on at least an improved LSTM neural network model, and the improved LSTM neural network model is to introduce a multi-head self-attention mechanism. S4, the early warning threshold model is periodically trained by the output unit connected to the acquisition unit and the filtering unit respectively, based at least on the multi-source micro-deformation data extracted from historical moments by a preset sliding window, and the core feature set is input into the trained early warning threshold model to output dynamic early warning intervals of different risk levels, wherein the learning rate of the early warning threshold model is adjusted at least during training; S5, the matching unit, which is connected to the filtering unit and the output unit respectively, performs matching based on the current core feature set and the dynamic early warning interval, generates an early warning signal with the corresponding risk level, and outputs it to each emergency response terminal. S6, the calculation unit connected to the matching unit determines the early warning characterization value based on the acquired early warning verification rate, risk false alarm coefficient, and average effective response time, wherein the early warning verification rate V is: In the formula, The actual number of warnings confirmed through on-site verification. The total number of warnings issued by the system; the risk underreporting coefficient O is... In the formula, The potential severity level of the i-th disaster event. Let be the time difference between the occurrence of the i-th disaster event and the successful detection of the anomaly. Let R be the total time from the inception to the occurrence of the i-th disaster event; the average effective response time R is... In the formula, This is the average time difference from when the system issues a warning to when the deformation enters the acceleration phase. The preset response time is; the warning characterization value M is... ,in, These are the weight coefficients for the corresponding parameters, and it is ensured that the sum of the three weight coefficients is 1; S7, the analysis unit connected to the computing unit determines the warning status based on the warning characterization value, adjusts the training period of the warning threshold model based on the warning status, and adjusts the learning rate of the warning threshold model based on the adjusted training period.

[0040] Please see Figure 3 The diagram shown is a flowchart illustrating the steps of determining the warning characterization value based on the comparison result between the warning characterization value and the preset warning characterization value in an embodiment of the present invention.

[0041] Specifically, in early monitoring and warning methods for geological disasters, which focus on the extraction and analysis of micro-deformation features, the system sets corresponding preset or critical parameters based on the limits of our understanding of the physical mechanisms of soil and rock deformation and the requirements for engineering safety tolerance. This is done in conjunction with several precursor data sequences obtained through statistical analysis and inversion from a historical disaster case database. These parameters are not fixed values ​​but are continuously optimized through historical data accumulation and model learning. This ensures that the warning system can capture the earliest abnormal signals within physical limits while meeting the tolerance requirements of actual engineering projects for the reliability and timeliness of warnings.

[0042] Specifically, determining the preset early warning characterization value typically requires a comprehensive approach, considering historical disaster case data, regional geological risk background, and industry early warning effectiveness standards. Specifically, firstly, a retrospective analysis is conducted based on historical monitoring data and known disaster events to calculate the baseline ranges for various indicators such as verification rate, false negative rate, and response time. Then, using mathematical models such as the weighted comprehensive evaluation method, weights are assigned to each indicator, and initial thresholds are set. This preset value is not fixed but serves as a benchmark reference point for the initial operation of the system and subsequent adaptive optimization. In this embodiment of the invention, the preset early warning characterization value L0 = 0.86 is set. The comparison process between the early warning characterization value L0 and the preset early warning characterization value L0 is as follows: If the warning indicator value L is greater than or equal to the preset warning indicator value L0, then the warning status is determined to be qualified. If the warning indicator value L is less than the preset warning indicator value L0, then the warning status is determined to be unqualified.

[0043] Specifically, when the warning status is unqualified, the warning verification rate at multiple historical moments is obtained; the average of the multiple warning verification rates is calculated. If the average is less than the preset average, it indicates that the dynamic warning interval constructed based on historical data cannot adapt to the new deformation mode, resulting in false alarms. In this case, the training period of the warning threshold model is adjusted based on the ratio of the average to the preset average. Here, the preset ratio P0 of the average to the preset average is set to 0.81. The comparison process between the ratio P of the average to the preset average and the preset ratio P0 is as follows: If the ratio P of the average value to the preset average value is less than or equal to the preset ratio P0, the training period of the warning threshold model will be adjusted to 0.59 times the original training period, where the adjusted training period is rounded up. If the ratio P of the average value to the preset average value is greater than the preset ratio P0, the training period of the warning threshold model will be adjusted to 0.73 times the original training period, where the adjusted training period is rounded up.

[0044] Specifically, more frequent training may mean smaller data increments each time, allowing for a more appropriate increase in the learning rate. This enables the model to adapt more quickly to minor changes in new data. Therefore, the learning rate of the early warning threshold model is adjusted based on the ratio of the adjusted training period to the preset training period. In this embodiment, the preset training period needs to comprehensively consider the update frequency of monitoring data and the time scale of geological disaster development and evolution. Typically, it is based on the natural collection cycle of the core monitoring data source and accumulates sufficient data to form an effective time window that reflects trends. Therefore, in this embodiment, the preset training period is set to 4 months, and this preset value will serve as a benchmark for dynamic adjustment, subsequently shortening or extending according to the actual early warning status. Simultaneously, the preset ratio Q0 of the adjusted training period to the preset training period is set to 0.75. The comparison process between the ratio Q0 of the adjusted training period and the preset training period is as follows: If the ratio Q of the adjusted training period of the warning threshold model to the preset training period is less than or equal to the preset ratio Q0, then the learning rate of the warning threshold model will be adjusted to 5 times the original learning rate. If the ratio Q of the adjusted training period of the warning threshold model to the preset training period is greater than the preset ratio Q0, then the learning rate of the warning threshold model will be adjusted to 8 times the original learning rate.

[0045] Specifically, if the control state is unqualified after adjusting the learning rate of the warning threshold model, the data timeliness decay coefficient within the window recorded at the beginning of the training cycle of the warning threshold model is obtained. If the data timeliness decay coefficient within the window is greater than a preset value, it indicates that the model contains too much historical information, including outdated data that may no longer be applicable to the current deformation mode, causing the model to be insensitive to recent changes, resulting in delayed or ineffective warnings. In this case, the preset sliding window is adjusted based on the ratio of the data timeliness decay coefficient within the window to the preset value. The data timeliness decay coefficient within the window can be directly obtained from the system metadata without complex calculations. By systematically correlating the data timeliness status within historical training periods—that is, the data timeliness decay coefficient within the window—with the model's actual performance in subsequent applications, such as early warning accuracy and false negative rate, the critical inflection point of the data timeliness decay coefficient within the window that leads to a significant performance decline is objectively identified. This inflection point value is then determined as the preset value corresponding to the data timeliness decay coefficient within the window used to diagnose whether the window is too large. Therefore, the preset value is set to 0.68, and the preset ratio R0 between the data timeliness decay coefficient within the window and the preset value is set to 1.27. The specific process of comparing the ratio R0 between the data timeliness decay coefficient within the window and the preset value is as follows: If the ratio R of the data time decay coefficient in the window to the preset value is less than or equal to the preset ratio R0, the preset sliding window will be adjusted to 0.83 times the original preset sliding window, where the adjusted preset sliding window will be rounded up. If the ratio R of the data time decay coefficient in the window to the preset value is greater than the preset ratio R0, the preset sliding window will be adjusted to 0.61 times the original preset sliding window, where the adjusted preset sliding window is rounded up.

[0046] Please see Figure 4 The diagram shown is a flowchart illustrating the steps of determining the warning status based on adjusting the preset sliding window in an embodiment of the present invention.

[0047] Specifically, if the warning status is unqualified after adjusting the preset sliding window, the preset sliding window is adjusted at least once until the number of adjustments is less than the preset number and the warning status is qualified, or the number of adjustments is equal to the preset number, at which point the adjustment stops. If the warning status is still unqualified after stopping the adjustment, the difference between the preset warning characterization value and the warning characterization value corresponding to multiple historical times is calculated. The average and variance of the multiple differences are calculated. If the average is greater than the preset difference and the variance is less than the preset variance, it indicates that the sensor in the monitoring network has drifted or malfunctioned. The ratio of variance to preset variance is then used to adjust the calibration cycle of the sensor in the monitoring network. Here, the preset ratio of variance to preset variance is set to T0 = 0.72. The comparison process based on the ratio of variance to preset variance T and the preset ratio T0 is as follows: If the ratio T of the variance to the preset variance is less than or equal to the preset ratio T0, the calibration period of the sensors in the monitoring network will be adjusted to 0.59 of the original calibration period, where the adjusted calibration period is rounded up. If the ratio T of the variance to the preset variance is greater than the preset ratio T0, the calibration period of the sensor in the monitoring network will be adjusted to 0.78 of the original calibration period, where the adjusted calibration period is rounded up.

[0048] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A geological disaster early monitoring and warning method based on micro-deformation feature extraction and analysis, characterized in that, include: A multi-source collaborative micro-deformation monitoring network is used to collect multi-source micro-deformation data of the monitoring area; The multi-source micro-deformation data is preprocessed, and the preprocessed multi-source data is fused using a multi-source data fusion model based on an attention mechanism to generate a micro-deformation time series dataset. The basic and deep features are extracted from the micro-deformation time series dataset, and the basic and deep features are filtered using a filtering algorithm to generate a core feature set. The deep features are extracted based on at least an improved LSTM neural network model, which is an improved LSTM neural network model that introduces a multi-head self-attention mechanism. The early warning threshold model is periodically trained based on the multi-source micro-deformation data extracted from historical moments using a preset sliding window, and the core feature set is input into the trained early warning threshold model to output dynamic early warning intervals for different risk levels. The learning rate of the early warning threshold model is adjusted at least during training. Based on the current core feature set and the dynamic early warning interval, a warning signal corresponding to the risk level is generated and output to each emergency response terminal. The early warning characterization value is determined based on the obtained early warning verification rate, risk underreporting coefficient, and average effective response time; The warning status is determined based on the warning characterization value, and the training period of the warning threshold model is adjusted based on the warning status. The learning rate of the warning threshold model is then adjusted based on the adjusted training period.

2. The method according to claim 1, wherein, The multi-source collaborative micro-deformation monitoring network includes an airborne monitoring platform, ground-based monitoring sensors, and underground monitoring sensors. Data transmission in the monitoring network employs dual-mode redundant communication technology.

3. The method according to claim 2, wherein, The early warning verification rate V is wherein, is the number of true early warnings confirmed by on-site verification, is the total number of early warnings issued by the system; The risk underreporting coefficient O is ,in, The potential severity level of the i-th disaster event. Let be the time difference between the occurrence of the i-th disaster event and the successful detection of the anomaly. Let i be the total time from the gestation to the occurrence of the i-th disaster event; The average effective response time R is ,in, This is the average time difference from when the system issues a warning to when the deformation enters the acceleration phase. Preset response time; The early warning characteristic value M is wherein, are weight coefficients of corresponding parameters, and ensure that the weight coefficients and are 1.

4. The method according to claim 3, wherein, The process of determining the warning status based on the warning characterization value includes: The warning characterization value is compared with the preset warning characterization value; If the warning indicator value is greater than or equal to the preset warning indicator value, then the warning status is determined to be qualified. If the warning indicator value is less than the preset warning indicator value, then the warning status is determined to be unqualified.

5. The method according to claim 4, wherein, The process of adjusting the training cycle of the early warning threshold model based on the early warning status includes: If the warning status is not up to standard, obtain the warning verification rate at multiple historical moments; Calculate the average of the multiple warning verification rates. If the average is less than the preset average, adjust the training period of the warning threshold model based on the ratio of the average to the preset average.

6. The method according to claim 5, wherein, The process of adjusting the training cycle of the early warning threshold model based on the ratio of the average value to the preset average value includes: The training period of the warning threshold model is reduced based on the ratio of the average value to the preset average value, and the reduction in the training period is inversely proportional to the ratio.

7. The method according to claim 6, wherein, The process of adjusting the learning rate of the early warning threshold model based on the adjusted training cycle includes: The learning rate of the early warning threshold model is increased by the ratio of the adjusted training period to the preset training period, and the increase in the learning rate is proportional to the ratio.

8. The method according to claim 7, wherein, The method further includes: If the control state is unqualified after adjusting the learning rate of the warning threshold model, obtain the data time decay coefficient recorded by the warning threshold model at the beginning of the training cycle. If the data attenuation coefficient within the window is greater than a preset value, the preset sliding window is adjusted based on the ratio of the data attenuation coefficient within the window to the preset value.

9. The method according to claim 8, wherein, The process of adjusting the preset sliding window based on the ratio of the data attenuation coefficient within the window to a preset value includes: The preset sliding window is reduced based on the ratio of the data time-related decay coefficient within the window to a preset value, and the reduction of the preset sliding window is proportional to the ratio.

10. The method according to claim 9, wherein, The method further includes: If the warning status is not satisfactory after adjusting the preset sliding window, repeat the adjustment of the preset sliding window at least once until the number of adjustments is less than the preset number and the warning status is satisfactory, or the number of adjustments is equal to the preset number and then stop adjusting. If the warning status after the adjustment is not qualified, calculate the difference between the preset warning characterization value and the warning characterization value corresponding to multiple historical moments; Calculate the average and variance of multiple differences. If the average is greater than the preset difference and the variance is less than the preset variance, then reduce the calibration cycle of the sensors in the monitoring network based on the ratio of the variance to the preset variance. The reduction in calibration cycle is inversely proportional to the ratio.

Images