Intelligent slope health degree evaluation system and method based on multi-source data fusion

Abstract

The present application relates to the technical field of slope health degree evaluation, and specifically discloses a slope health degree intelligent evaluation system and method based on multi-source data fusion, which constructs a multi-dimensional feature matrix fused with change rate by receiving displacement time sequence and rainfall environment data of multiple monitoring terminals of the slope. The core technology is that each characteristic value is mapped through a Gaussian membership function and converted into an initial reliability distribution under different risk levels to form a risk reliability vector. The reliability vector is differentially weighted to generate evidence items representing local risk contribution. Finally, all weighted evidence items are input into an improved evidence reasoning model to effectively resolve multi-source information contradictions, calculate the global membership probability of the overall slope at each risk level, and fuse to generate a continuous overall health degree score that can dynamically reflect the stability state, realizing the leap from local heterogeneous perception to overall intelligent judgment and providing a core basis for slope safety early warning and precise control.

Description

Technical Field

[0001] This invention belongs to the field of slope health assessment technology, and relates to an intelligent slope health assessment system and method based on multi-source data fusion. Background Technology

[0002] Slope geological hazards, as a major hidden danger threatening infrastructure safety and the lives and property of residents, have complex formation mechanisms, are highly concealed and sudden, and are difficult to prevent and control. Multi-source monitoring data, as the "nerve endings" for sensing slope conditions, includes key dimensions such as displacement, rainfall, and deep displacement. The accuracy of its collection and the depth of its analysis directly determine the timeliness and reliability of geological hazard early warning, and are an indispensable basis for decision-making in landslide prevention and control.

[0003] It is worth noting that multi-source monitoring data do not exist in isolation; they collectively carry information reflecting the overall state of the system. However, they often differ in dimensions, accuracy, and reliability, and contain significant uncertainties. For example, abnormal changes in local displacement may be coupled with environmental factors, and an alarm from a single sensor may not be sufficient to determine the overall risk.

[0004] However, most existing health assessment methods still have many limitations in this regard. Traditional methods often rely on threshold judgments for single or a few key indicators, making it difficult to comprehensively utilize the complementary and corroborative value of multi-dimensional information. They are also prone to false alarms or missed alarms due to individual sensor false alarms or data fluctuations. More importantly, existing methods are relatively weak in handling the inherent ambiguity and uncertainty of monitoring data. They usually use simple weighted averages or deterministic models, failing to explicitly characterize and reason about the differences in credibility from different information sources and the contradictions between them. This leads to the final assessment result potentially being a fragile "consensus," failing to effectively distinguish between the "certain" and "uncertain" parts of the information. This makes the health index insensitive to early, subtle abnormalities, resulting in delayed warnings and difficulty in interpretation. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention provides an intelligent assessment system and method for slope health based on multi-source data fusion to solve the above-mentioned technical problems.

[0006] To achieve the above and other objectives, the technical solution adopted by the present invention is as follows:

[0007] The first aspect of this invention provides an intelligent assessment method for slope health based on multi-source data fusion, the method comprising:

[0008] It receives multi-source slope monitoring data collected by multiple monitoring terminals, which includes displacement time series and rainfall environmental data;

[0009] The multi-source monitoring data of the slope is segmented by a sliding window with a predetermined step size. The rate of change of displacement time series is extracted by least squares regression calculation. Combined with rainfall environmental data, a multi-dimensional feature matrix is ​​formed for a single monitoring terminal.

[0010] Each eigenvalue in the multidimensional feature matrix is ​​mapped to a preset Gaussian membership function to calculate the initial confidence distribution of the monitoring terminal under different risk levels, and a risk confidence vector is constructed.

[0011] Retrieve the pre-stored weight configuration information, which includes the relative weight values ​​assigned to sensors of different monitoring dimensions and locations using the analytic hierarchy process (AHP).

[0012] The risk confidence vector is weighted by applying relative weight values ​​to generate a weighted evidence term that represents the risk contribution of a local area;

[0013] The weighted evidence items corresponding to multiple monitoring terminals are input into the weighted evidence reasoning model. Evidence conflict identification and joint probability synthesis are performed inside the weighted evidence reasoning model. Data conflicts between different monitoring dimensions are eliminated through iterative synthesis algorithm, and the global membership probability of the slope as a whole at each risk assessment level is calculated.

[0014] The global membership probability is quantitatively weighted and summed to generate a health score that represents the overall stability of the slope. The health score is then sent to the early warning terminal to trigger the corresponding safety control logic.

[0015] The second aspect of this invention provides an intelligent slope health assessment system based on multi-source data fusion, the system comprising:

[0016] Multi-source data acquisition module: Receives multi-source slope monitoring data collected by multiple monitoring terminals. The multi-source slope monitoring data includes displacement time series and rainfall environmental data.

[0017] Risk vector construction module: The multi-source monitoring data of the slope is segmented using a sliding window with a predetermined step size. The rate of change of displacement time series is extracted by least squares regression calculation. Combined with rainfall environmental data, a multi-dimensional feature matrix for a single monitoring terminal is formed. Each feature value in the multi-dimensional feature matrix is ​​mapped to a preset Gaussian membership function to calculate the initial confidence distribution of the monitoring terminal under different risk levels, and a risk confidence vector is constructed.

[0018] Risk weighting calculation module: retrieves pre-stored weight configuration information, which includes relative weight values ​​assigned to sensors of different monitoring dimensions and locations using the analytic hierarchy process; applies the relative weight values ​​to perform weighted mapping on the risk confidence vector to generate weighted evidence items characterizing the risk contribution of a local area;

[0019] The slope health assessment module inputs weighted evidence items from multiple monitoring terminals into a weighted evidence inference model. Within the model, evidence conflict identification and joint probability synthesis are performed. An iterative synthesis algorithm eliminates data conflicts between different monitoring dimensions, and the global membership probability of the slope at each risk assessment level is calculated. The global membership probability is then quantitatively weighted and summed to generate a health score that characterizes the overall stability of the slope. This health score is then sent to the early warning terminal to trigger the corresponding safety control logic.

[0020] As described above, the intelligent slope health assessment system and method based on multi-source data fusion provided by the present invention has at least the following beneficial effects:

[0021] This invention segments multi-source slope monitoring data using a sliding window with a predetermined step size and extracts the rate of change features of displacement time series using a least squares regression algorithm. Combined with rainfall and environmental data, a multi-dimensional feature matrix is ​​constructed, achieving accurate capture and quantification of the dynamic evolution characteristics of slopes. In particular, by introducing a Gaussian membership function to transform heterogeneous physical monitoring data into a risk confidence vector, it effectively solves the problem that a single monitoring indicator cannot objectively describe the complex nonlinear state of slopes. It overcomes the feature expression distortion caused by data noise or random fluctuations in traditional methods, laying a solid data foundation for subsequent risk quantification analysis and significantly improving the ability of monitoring data to express the actual stability of slopes.

[0022] This invention introduces the analytic hierarchy process (AHP) to construct pre-stored weight configuration information and applies relative weight values ​​to weighted map the risk confidence vector, generating weighted evidence terms that can characterize the risk contribution of local areas. This process fully considers the differential contributions of sensors from different monitoring dimensions and locations in slope stability evaluation, avoiding the problem of key information being diluted due to neglecting differences in sensor importance in traditional fusion models. It effectively enhances the weight proportion of monitoring data in high-risk areas, making the characterization of local risk features more consistent with geomechanical mechanisms, thereby significantly improving the rationality and accuracy of multi-source data fusion.

[0023] This invention utilizes a weighted evidence reasoning model to fuse weighted evidence items from multiple monitoring terminals. Through internal evidence conflict identification and joint probability synthesis mechanisms, it effectively eliminates conflicts and uncertainties among multi-source heterogeneous data, calculates the overall global membership probability of the slope, and generates a health score. This method overcomes the bottleneck of effectively fusing multi-source information in traditional assessments, achieving a leap from "local perception" to "holistic judgment." It can output continuous and intuitive quantitative indicators of health, providing reliable triggering basis for early warning terminals and avoiding false alarms or missed alarms caused by single sensor failures, significantly improving the scientific rigor and foresight of slope safety management decisions. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram showing the connections between the steps of the method of the present invention.

[0026] Figure 2 A schematic diagram of the global membership probability calculation logic provided in this application.

[0027] Figure 3 This is a schematic diagram showing the connections of the various modules in the system of the present invention. Detailed Implementation

[0028] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

[0029] In traditional slope monitoring and early warning systems, fixed threshold values ​​and isolated indicator evaluation rules cannot adapt to the nonlinear characteristics and multi-source coupling effects of dynamic geological evolution. When slope stability undergoes complex changes under the coupled effects of rainfall infiltration and stress redistribution, the system cannot establish a dynamic correlation between multi-source heterogeneous monitoring data and geological safety status, leading to a mismatch between risk assessment results and the actual disaster incubation process. This static threshold mechanism reduces the depth of multi-source data fusion, causing key precursor features to be submerged in data noise, ultimately affecting the timeliness and accuracy of early warning decisions.

[0030] For example, in an emergency monitoring scenario of a landslide induced by heavy rainfall, after three hours of continuous rainfall, the deep horizontal displacement rate of the slope increased from an initial 0.5 mm / h to 2.0 mm / h, with the surface displacement exhibiting a non-linear accelerating trend, and the rainfall intensity exceeding historical extremes. At this point, traditional systems still use preset fixed displacement thresholds for judgment, failing to capture the strong correlation between rainfall environmental data and the accelerating displacement trend. This leads to key precursor features characterizing soil saturation and softening being misjudged as sensor errors. Single-perspective monitoring and analysis loses crucial data dimensions characterizing the overall slope instability, and the risk assessment model fails to identify the critical state from gradual to abrupt change in a timely manner, still displaying a "low risk" safety status.

[0031] If the above problems are not addressed, the fragmented processing of multi-source monitoring data will prevent risk assessment models from identifying the evolution trend of overall slope stability, delaying optimal emergency response and engineering measures. Rigid feature extraction mechanisms will hinder the system from capturing critical early warning signals under multi-field coupling effects, easily issuing erroneous safety signals due to a single indicator not exceeding its limit. The lack of multi-source data fusion capabilities will also cause a disconnect between local monitoring anomalies and the overall stable state, reducing the reliability of landslide early warning models and ultimately creating a negative feedback loop that undermines the scientific validity of disaster prevention and mitigation decisions.

[0032] When faced with the aforementioned problems, traditional systems use isolated thresholds to process multi-source data, leading to the neglect of key geological evolution parameters and an inability to reflect the true health status of slopes. To address this, this application attempts to extract rate-of-change features using a sliding window and least squares regression, and constructs a multi-dimensional feature matrix by combining rainfall and environmental data. A Gaussian membership function is then used to transform heterogeneous data into risk confidence vectors. Further analysis reveals that relying solely on single-sensor data or simple weighting is insufficient to handle conflicts and uncertainties among monitoring data. Therefore, the analytic hierarchy process (AHP) is introduced to determine the relative weights of different monitoring dimensions, and a weighted evidence reasoning model is constructed. Through a linkage mechanism between the risk vector construction module, the risk weighting calculation module, and the slope health assessment module, evidence reasoning algorithms are used to eliminate data conflicts, calculate the global membership probability, and generate health scores. This enables adaptive adjustment of slope stability assessment from local perception to overall judgment, solving the problems of assessment lag and multi-source data mismatch.

[0033] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0034] Example 1:

[0035] Please see Figure 1-2 As shown, a smart slope health assessment method based on multi-source data fusion is proposed, which includes:

[0036] It receives multi-source monitoring data of slopes collected by multiple monitoring terminals. The multi-source monitoring data of slopes includes displacement time series and rainfall environment data.

[0037] In one specific embodiment, multiple monitoring terminals deployed at key runoff points and areas of severe deformation on the slope perform synchronous data acquisition tasks, receiving displacement time series data acquired by a high-precision GNSS receiver and rainfall environmental data acquired by a tipping bucket rain gauge. The initial sampling frequency of the displacement time series is preset to an adjustable range of once every 15 to 60 minutes, preferably once every 30 minutes. This frequency setting logic ensures that minute-level creep characteristics of the slope induced by rainfall can be captured, while effectively filtering out random high-frequency noise generated by GNSS satellite positioning.

[0038] A sliding window with a predetermined step size is used to segment the multi-source monitoring data of the slope. The rate of change characteristics of the displacement time series are extracted using least squares regression calculation. Combined with rainfall and environmental data, a multi-dimensional feature matrix is ​​formed for each monitoring terminal, including:

[0039] The sliding window moves along the time axis according to a predetermined step size. The sliding window is used to extract the displacement time series in the multi-source monitoring data of the slope and obtain the displacement time data segment within the time range covered by the current window.

[0040] By using a sliding window, rainfall environmental data from multi-source slope monitoring data are captured synchronously to obtain rainfall environmental data segments that are consistent with the time points of the displacement time data segments.

[0041] A univariate linear fitting equation is established using each time point in the displacement time data segment as the independent variable and the displacement acquisition value corresponding to each time point as the dependent variable.

[0042] The slope of the univariate linear fitting equation that minimizes the sum of squared residuals of all displacement data is calculated using the least squares regression algorithm. The slope of the univariate linear fitting equation is then set as the rate of change characteristic of the displacement time series.

[0043] Extract the rainfall values ​​at each monitoring time point in the rainfall environment data segment, and sum all the rainfall values ​​within the rainfall environment data segment to obtain the cumulative rainfall characteristics corresponding to the sliding window;

[0044] The rate of change features and cumulative rainfall features are dimensionally aligned and array-concatenated according to the physical identifier of each monitoring terminal to construct a multi-dimensional feature matrix for each monitoring terminal.

[0045] In a specific embodiment, the system slides a window along the time axis according to a predetermined step size S. This predetermined step size S is recommended to be set to an adjustable range of 1 to 6 hours, preferably 2 hours, to balance the sensitivity of real-time monitoring with the consumption of computing resources. The length W of the sliding window is preset to 24 to 72 hours, preferably 48 hours. Its setting logic is based on the physical laws of the slope soil permeability coefficient and the lag effect of rainfall infiltration, ensuring that the window coverage can completely encompass a physical cycle from rainfall occurrence to deformation response.

[0046] Next, the sliding window is used to extract the displacement time series from the multi-source monitoring data of the slope, obtaining the displacement time data segment within the current window's coverage time range. ,in The displacement observation value of the i-th sampling point is expressed in millimeters, and n is the total number of sampling points within the window; rainfall environmental data segments within the same timestamp range are extracted synchronously. ,in This represents the rainfall value at the corresponding time, in millimeters. To quantify the deformation trend, the system uses the relative time of each sampling moment in the displacement-time data segment relative to the start point of the window as the independent variable, and the corresponding displacement acquisition value as the independent variable. As the dependent variable, a univariate linear fitting equation is established. .

[0047] In the above univariate linear fitting equation, d represents the observed displacement value, i.e., the slope displacement at a certain moment within the current window. t represents relative time (unit: h), i.e., the time offset of this moment relative to the starting point of the window. v represents the rate of change characteristic (unit: mm / h), i.e., the slope of the fitted line, reflecting how fast the displacement changes with time; b represents the intercept of the fitted line (unit: mm), i.e., the estimated displacement value when t=0, usually corresponding to the displacement baseline at the starting point of the window.

[0048] The system uses the least squares regression algorithm to solve for the slope of the equation, which is used as the rate of change characteristic v of the displacement time series. The calculation formula is as follows: In the formula, v represents the rate of change, measured in millimeters per hour. This represents the time offset of the i-th data point, in hours. The displacement value is collected for the i-th data point; and These are the arithmetic mean of the time offset and displacement values ​​within the window, respectively. This calculation method effectively filters out random jumps in the sensor caused by environmental interference by minimizing the sum of squared residuals between all observation points and the fitted line.

[0049] Simultaneously, the system performs scalar accumulation on the rainfall values ​​within the rainfall environmental data segment to extract the cumulative rainfall characteristics. ,in The data represents cumulative rainfall, expressed in millimeters. The time interval corresponding to the monitoring frequency; The preset rainfall infiltration correction coefficient is a dimensionless coefficient, whose value is preset based on the vegetation cover and slope of the slope surface, and typically ranges from 0.6 to 0.95. This embodiment directly uses the periodic cumulative rainfall data recorded by a tipping bucket rain gauge. The formula then simplifies to .

[0050] Finally, the system will calculate the rate of change characteristic v and the cumulative rainfall characteristic. Dimensional mapping is performed based on the physical ID of the monitoring terminal to construct a multi-dimensional feature matrix for a single monitoring terminal. For example, if the displacement monitored within the window increases uniformly and linearly from 10 mm to 22 mm over 48 hours, then v is calculated as (22-10) / 48 = 0.25 mm / h; if the total rainfall during the same period is 50 mm, then the matrix... .

[0051] Each eigenvalue in the multidimensional feature matrix is ​​mapped to a preset Gaussian membership function to calculate the initial reliability distribution of the monitoring terminal under different risk levels, thus constructing a risk reliability vector.

[0052] Preferably, the construction logic of the risk confidence vector includes:

[0053] Extract the rate of change features and cumulative rainfall features from the multidimensional feature matrix to obtain the preset level center value corresponding to each risk level;

[0054] Calculate the squared deviations between the rate of change characteristics, cumulative rainfall characteristics, and the preset center values ​​corresponding to each risk level;

[0055] The squared deviation value is scaled using a preset standard deviation, and the scaled value is used as the exponent of the natural logarithm base to perform a negative operation to obtain the membership score of each feature for different risk levels.

[0056] Based on the numerical range of the rate of change characteristics and cumulative rainfall characteristics, the membership scores at both ends of the risk level sequence are saturated and smoothed to obtain the corrected membership scores of each feature under different risk levels.

[0057] The membership scores of each feature of the same monitoring terminal under the same risk level are averaged to generate the initial reliability distribution of the monitoring terminal under different risk levels.

[0058] The initial confidence distributions corresponding to each risk level are arranged in a predetermined order from low to high risk to construct a risk confidence vector that can characterize the risk status of a single monitoring point.

[0059] Preferably, the calculation process for the modified membership scores of each feature under different risk levels includes the following steps:

[0060] Determine the lowest risk level at the beginning and the highest risk level at the end of the risk level sequence;

[0061] The change rate feature and cumulative rainfall feature are compared with the preset level center value corresponding to the lowest risk level. If the feature value is less than the preset level center value corresponding to the lowest risk level, the membership score corresponding to the lowest risk level is updated to the predetermined constant peak value.

[0062] The change rate feature and cumulative rainfall feature are compared with the preset level center value corresponding to the highest risk level. If the feature value is greater than the preset level center value corresponding to the highest risk level, the membership score corresponding to the highest risk level is updated to the predetermined constant peak value.

[0063] Keeping the membership scores of intermediate risk levels (excluding the lowest and highest risk levels) unchanged in the risk level sequence, the updated membership scores are combined with the unchanged membership scores to generate modified membership scores for each feature under different risk levels.

[0064] In a specific embodiment, the rate of change feature v and the cumulative rainfall feature are extracted from the multidimensional feature matrix. It also retrieves a preset risk level center value matrix, which contains a sequence of center values ​​for displacement rates. and the central value sequence of rainfall Where k is the risk level index, ranging from 1 to 5, representing five levels from extremely low risk to extremely high risk; the selection logic of these central values ​​is based on the slope engineering geological investigation specifications and historical landslide statistics, for example... This threshold is typically set as the upper limit for issuing a landslide warning for this type of slope.

[0065] For each feature, the system calculates the squared deviation from the corresponding level center value and uses a preset standard deviation. Scaling is performed using the formula. Calculate the preliminary membership score. Among them, The feature components representing the input, Let the center position parameter of this feature be at level k. The preset risk coverage bandwidth parameter (with dimensions consistent with the corresponding characteristics) is designed to control the overlap between adjacent risk levels, and is typically set to the center value. The membership percentage is 15% to 25%, preferably 20%, to ensure a smooth transition of membership during risk state transitions.

[0066] For the lowest risk level, if the characteristic value Less than the center value Then, force the correction of its membership score. Updated to a predetermined constant peak value of 1.0; similarly, for the highest risk level, if the eigenvalue... Greater than the center value Then adjust its membership score. Updated to 1.0. The necessity of this correction logic lies in the fact that when the displacement rate far exceeds the critical slip threshold, the system should maintain full confidence in the highest risk level, rather than decreasing due to deviation from the Gaussian center, thus ensuring the reliability of the early warning.

[0067] Subsequently, the system averaged the modified membership scores of each feature for the same monitoring terminal at the same risk level to obtain the initial reliability distribution. This process eliminates the risk of false alarms that may arise from a single indicator by equally weighting the inducing factor (rainfall) and the response performance (displacement rate). Finally, the initial confidence distributions corresponding to the five risk levels are vectorized and encapsulated in ascending order of risk to construct a risk confidence vector. represents the membership score for the k-th risk level after saturation smoothing correction of the rate of change characteristic; it represents the result after boundary correction for the lowest or highest risk level based on the original Gaussian membership calculation, ensuring that the membership score can still reasonably reflect the risk status when the characteristic value exceeds the normal range. This represents the membership score of the cumulative rainfall characteristics after saturated smoothing correction for the k-th risk level. Similarly, it is the membership score after Gaussian calculation and correction of the rainfall characteristics.

[0068] Retrieve the pre-stored weight configuration information, which includes the relative weight values ​​assigned to sensors of different monitoring dimensions and locations using the analytic hierarchy process.

[0069] The preset logic for weight configuration information is as follows:

[0070] Establish a reference sample set containing multiple historical slope instability events. For each historical instability event, extract the displacement time series, rainfall environment data, and corresponding slope stability state labels within a predetermined time period before instability.

[0071] The reference sample set is divided into a deformation feature layer, a rainfall environment layer, and a location distribution layer according to the monitoring dimensions. Based on the correlation contribution of each monitoring dimension to the slope stability state label, an initial judgment matrix for pairwise comparison is constructed.

[0072] The maximum eigenvalue and consistency ratio of the matrix are set as iteration parameters. The power method is used to iteratively calculate the initial judgment matrix to solve the eigenvector of the initial judgment matrix at the current iteration number.

[0073] The reference sample set is weighted using the components of each dimension in the feature vector to obtain the sample risk assessment value, and the mean square error between the sample risk assessment value and the slope stability status label is calculated.

[0074] The elements in the initial judgment matrix are adjusted using the mean square error value, and the consistency ratio value in the current iteration parameters is calculated simultaneously.

[0075] If the consistency ratio is less than the preset consistency threshold and the mean square error reaches the global minimum distribution, stop the iteration and extract the current feature vector as the target weight vector.

[0076] The target weight vector is mapped and encapsulated according to the physical index of sensors in different monitoring dimensions and locations to generate pre-stored weight configuration information that uniquely corresponds to the monitoring terminal.

[0077] In one specific embodiment, the system retrieves pre-stored weight configuration information. This weight configuration information includes relative weight values ​​assigned to different monitoring dimensions and sensors at different locations using the analytic hierarchy process (AHP). Its construction logic is based on an iterative optimization process using historical data. First, a reference sample set containing N historical slope instability events is established, where N is a positive integer representing the total number of samples. For each historical instability event, the displacement time series, rainfall environment data, and corresponding slope stability state label y are extracted within a predetermined time period (e.g., 48 hours) before the instability occurs. Here, y is a normalized dimensionless value, ranging from [0,1]. A larger value indicates a higher probability of instability, and this value originates from expert assessments of historical disaster reports. Subsequently, the reference sample set is divided into a deformation feature layer, a rainfall environment layer, and a location distribution layer according to the monitoring dimensions. Based on the correlation contribution of each monitoring dimension to the slope stability state label, an initial judgment matrix A for pairwise comparisons is constructed, with dimensions [missing information]. Where n is the total number of monitoring dimensions, and the matrix elements are... This represents the importance ratio of the i-th dimension to the j-th dimension. It is a dimensionless coefficient, and the initial value is set in the range of 1 to 9 based on expert experience.

[0078] Next, the largest eigenvalue of the matrix is ​​set. The consistency ratio CR is used as the iteration parameter. The power method is used to iteratively calculate the initial judgment matrix, and the eigenvector of the initial judgment matrix at the current iteration number is solved. The calculation formula is: .in, Let be the eigenvector obtained in the k-th iteration, and its components be denoted as '(k-th)'. The weight value of the i-th dimension is a dimensionless coefficient; A is the judgment matrix for the current iteration. The 1-norm of the vector, representing the sum of the absolute values ​​of its elements, is used to normalize the feature vector, ensuring that the sum of all weights is 1; k is the iteration index. The risk assessment value of the sample is obtained by weighting the reference sample set using the dimensional components of the feature vector. ,in Let be the normalized feature value of the i-th dimension. These are weighted risk prediction values, all dimensionless. The mean square error L between the sample risk assessment values ​​and the slope stability status labels is calculated using the following formula: ,in and are the predicted value and the true label value of the j-th sample, respectively, and L is a dimensionless coefficient used to characterize the accuracy of the model's prediction.

[0079] Based on the mean square error value L, the elements in the initial judgment matrix are... Perform gradient adjustment, the adjustment formula is as follows: .in, The preset learning rate is a dimensionless coefficient, and its value range is recommended to be set to 0.001 to 0.05, preferably 0.01. The preset logic is to balance the step size of weight updates with convergence stability and avoid oscillation and divergence caused by excessive step size. The partial derivative of the loss function with respect to the matrix elements reflects the degree to which changes in these elements affect the prediction error. During the adjustment process, the consistency ratio CR in the current iteration parameters is calculated simultaneously using the following formula: .in, The largest eigenvalue of matrix A is obtained by... The calculated value is a dimensionless coefficient; RI is a preset average random consistency index, a dimensionless constant whose value is obtained by looking up a table based on the matrix order n, used to correct for deviations in random consistency. When the consistency ratio CR is less than the preset consistency threshold (recommended value is 0.1) and the mean square error L reaches a globally minimized distribution, the iteration stops and the current feature vector w is extracted as the target weight vector. Finally, the target weight vector is mapped and encapsulated according to the physical indices of different monitoring dimensions and different location sensors to generate pre-stored weight configuration information uniquely corresponding to the monitoring terminal.

[0080] The risk confidence vector is weighted by applying relative weight values ​​to generate a weighted evidence term that represents the risk contribution of a local area.

[0081] In a specific embodiment, the system extracts the relative weight value for a specific monitoring terminal from the storage module. This weight value is a dimensionless coefficient, and its magnitude reflects the reliability and importance of the monitoring terminal in the overall monitoring deployment. Extract the risk confidence vector corresponding to this monitoring terminal. ,in The initial confidence distribution of the i-th monitoring terminal at the k-th risk level, calculated in the aforementioned steps, is a dimensionless percentage value ranging from [0,1].

[0082] The system calculates the weighted reliability quality corresponding to each risk level based on the discount operator logic of evidence theory. The calculation formula is as follows: .in, Let be the weighted reliability quality of the i-th monitoring terminal at the k-th risk level, and be a dimensionless coefficient. This formula linearly scales the original reliability using weighting coefficients, transforming the local monitoring information acquired by the sensor into evidence components with global reference value. To characterize the information uncertainty caused by sensor weight limitations, the system simultaneously calculates the residual quality items not assigned to a clearly defined risk level, using the following formula: .in, Let be the remaining quality term corresponding to the i-th monitoring terminal, reflecting the uncertainty space left due to the monitoring weight discount, and be a dimensionless coefficient.

[0083] Finally, the system will weight the reliability quality. With remaining mass item The arrays are concatenated and aligned according to the risk level sequence to generate weighted evidence items representing the risk contribution of local areas. This weighted evidence not only includes the supporting strength of each risk level, but also includes margin information reflecting the weight of the measurement points, eliminating risk assessment bias caused by the limited accuracy of a single sensor.

[0084] Weighted evidence items from multiple monitoring terminals are input into a weighted evidence inference model. Within this model, evidence conflict identification and joint probability synthesis are performed. An iterative synthesis algorithm eliminates data conflicts between different monitoring dimensions, ultimately calculating the overall global membership probability of the slope at each risk assessment level, including:

[0085] The reliability assignment values ​​corresponding to different risk assessment levels in each weighted evidence item are extracted, and the basic reliability quality assigned to each risk assessment level and the residual uncertainty quality not assigned to a specific level are calculated based on the relative weight of each monitoring terminal.

[0086] By performing a cross-product summation operation on the basic reliability quality among different monitoring terminals, the mutual exclusivity of different monitoring terminals under the same risk assessment level is identified, and a conflict interference factor used to characterize the degree of evidence contradiction is calculated.

[0087] The recursive synthesis algorithm is adopted, taking the basic reliability quality of the first monitoring terminal as the initial fusion state, and introducing the basic reliability quality of subsequent monitoring terminals one by one. In each iteration, the cross-product result is adjusted by the conflict interference factor to generate the joint probability distribution of the intermediate state.

[0088] After completing the iterative synthesis of all monitoring terminals, the final generated joint probability distribution and the synthesized residual uncertainty mass are extracted. By performing normalization mapping, the residual uncertainty mass is proportionally redistributed to each risk assessment level.

[0089] The assigned values ​​are defined as the global membership probability of the slope as a whole at each risk assessment level.

[0090] Preferably, the basic reliability quality assigned to each risk assessment level and the residual uncertainty quality not assigned to a specific level are calculated, including:

[0091] For the weighted evidence items, the corresponding reliability allocation values ​​are extracted according to the five preset risk assessment levels;

[0092] The reliability allocation value corresponding to each risk assessment level is multiplied by the relative weight of the monitoring terminal to generate the basic reliability quality for each risk assessment level.

[0093] The difference between the numerical value 1 and the relative weight is calculated to obtain the weight discount quality component, which characterizes the dilution of the trust level of the monitoring terminal by the weight.

[0094] Perform a cumulative summation operation on the basic reliability quality corresponding to all risk assessment levels, and calculate the difference between the value 1 and the cumulative summation result to obtain the inherent uncertainty quality component that characterizes the fuzziness of the monitoring data itself;

[0095] The weighted discounted mass component and the inherently uncertain mass component are linearly summed to generate the residual uncertainty mass that is not assigned to a specific level.

[0096] Preferably, a recursive synthesis algorithm is used, taking the basic reliability quality of the first monitoring terminal as the initial fusion state, and successively introducing the basic reliability quality of subsequent monitoring terminals. In each iteration, a conflict interference factor is used to adjust the gain of the cross-product result, generating a joint probability distribution of the intermediate state, including:

[0097] Assign the basic reliability quality of the first monitoring terminal at each risk assessment level to the current fusion probability vector as the initial fusion state for recursive calculation;

[0098] Extract the components in the current fusion probability vector and the basic reliability quality of the next monitoring terminal to be fused at the corresponding risk assessment level, and calculate the cross product value of the two.

[0099] Calculate the difference between the value 1 and the conflict interference factor, and set the reciprocal of this difference as the gain correction coefficient to characterize the degree of amplification of the evidence consistency part;

[0100] The cross-product values ​​are amplified by a gain correction factor to obtain the updated current fusion probability vector, thus completing the evidence synthesis for the current round.

[0101] The extraction, calculation, and amplification steps are performed cyclically according to the order of the monitoring terminals until the basic reliability quality of all monitoring terminals has been fused. The final fused probability vector is then determined as the joint probability distribution of the intermediate state.

[0102] The preferred construction logic of the weighted evidence reasoning model includes:

[0103] Establish a historical review sample set containing multiple slope evolution cycles, and extract historical displacement time series, historical rainfall environmental data, and corresponding risk level expert annotation values ​​from the historical review sample set;

[0104] Historical change rate characteristics are calculated using historical displacement time series data, and historical cumulative rainfall characteristics are extracted by combining historical rainfall environmental data, thus constructing a historical feature matrix that can reflect historical risk status.

[0105] Set the initial iteration parameters for the weighted evidence reasoning model. The iteration parameters include evidence discount weight, conflict adjustment coefficient, and membership degree correction compensation value.

[0106] The historical feature matrix is ​​input into the weighted evidence reasoning model to be trained. By performing confidence quality conversion and evidence recursion synthesis operations, the simulated risk probability distribution corresponding to the historical samples is obtained.

[0107] Calculate the sum of squared residuals between the simulated risk probability distribution and the expert-labeled risk level values, and define it as the loss function value characterizing the model bias;

[0108] The gradient descent algorithm is used to update the evidence discount weights, and the backpropagation mechanism is used simultaneously to dynamically adjust the conflict adjustment coefficient and the membership correction compensation value until the loss function value converges to the preset global optimum range.

[0109] Extract the evidence discount weight, conflict adjustment coefficient, and membership degree correction compensation value when the preset convergence range is reached, encapsulate them as parameters, and generate a weighted evidence reasoning model.

[0110] In one specific embodiment, the model extracts confidence assignment values ​​for five preset risk assessment levels (from low risk to extremely high risk) from each weighted evidence item. Where i is the monitoring terminal index and k is the risk level index. The coefficients are dimensionless coefficients within the interval [0,1]; the system incorporates the relative weight values ​​of each monitoring terminal. The basic reliability quality assigned to each risk assessment level was calculated. This value represents the degree of certainty with respect to the k-th risk level by the i-th terminal.

[0111] To characterize the uncertainties during the monitoring process, the system synchronously calculates the residual uncertainty mass that has not been assigned to a defined level. It is composed of weighted discounted quality components. With inherent uncertain mass components It is formed by linear addition, that is ;in This reflects a lack of trust due to sensor weight dilution, and This reflects the inherent ambiguity of the monitoring data. Subsequently, the system performs evidence conflict identification by calculating the conflict interference factor through cross-product summation of the basic reliability quality among different monitoring terminals. This parameter is used to quantify the degree of mutual exclusion between different terminals in the same risk assessment, and is a dimensionless value. This represents the basic reliability quality of the i-th terminal at the k-th risk level.

[0112] Next, the system employs a recursive synthesis algorithm, using the basic reliability quality of the first monitoring terminal as the initial fusion state, and then introducing data from subsequent terminals one by one. In each iteration, the gain correction coefficient is calculated using the conflict interference factor. The underlying logic of this coefficient is to eliminate data conflicts between different monitoring dimensions by amplifying the consistency of evidence. (Joint probability distribution of intermediate states) Through formula The calculation shows that, among which The fusion probability for the current round remains dimensionless. After iterating through all terminals, the final joint probability distribution and the total residual uncertainty mass after synthesis are extracted. Through normalization mapping, The risk levels are then proportionally redistributed, and the resulting value represents the overall global membership probability of the slope at each risk level. .

[0113] In the recursive composition algorithm, This represents the residual uncertainty quality in the current fusion result that has not been assigned to any specific risk level; that is, the "uncertainty" shared by all the monitoring terminal evidence that has been fused. Its purpose is to preserve the reliability gaps caused by evidence weight discounting and the inherent ambiguity of the data, and to incorporate it into subsequent synthesis along with the residual uncertainty quality of new evidence. To integrate and update the overall level of uncertainty.

[0114] The source and update rules are as follows:

[0115] When synthesizing evidence for the first monitoring terminal, the residual uncertainty mass of that terminal is directly taken. (Obtained by linear summation of the weighted discount component and the inherent uncertainty component).

[0116] In each round of introducing new evidence, the remaining mass of the i-th terminal is... After that, the new remaining mass Calculated using the residual mass update rule in the evidence-based synthesis formula:

[0117] Where K is the conflict interference factor calculated in the current round, reflecting the degree of contradiction between the old and new evidence. This updated formula ensures the reasonable transmission and normalization of uncertainty during the fusion process. Updated That is, to become the next round of synthesis This continues until all evidence is fully integrated. The final total remaining mass... It will participate in the normalized allocation of global membership probability.

[0118] To ensure the accuracy of the aforementioned inference model, its internal parameters underwent rigorous training with historical samples. The system established a historical review sample set containing multiple slope evolution cycles, extracted historical displacement rate features and cumulative rainfall features to construct a historical feature matrix, and set initial iteration parameters including evidence discount weights. Conflict adjustment coefficient and membership degree correction compensation value (All coefficients are dimensionless). The historical feature matrix is ​​input into the model to solve for the simulated risk probability distribution, and the sum of squared residuals between this distribution and the expert-labeled risk level values ​​is calculated, defined as the loss function value. Gradient descent is used to update the parameters, with a recommended learning rate ranging from 0.005 to 0.02, preferably 0.01, to ensure smooth convergence of the loss function during iteration. When L reaches the globally minimized distribution, the result at this point is extracted. The parameters are encapsulated to generate the final weighted evidence reasoning model.

[0119] The global membership probability is quantitatively weighted and summed to generate a health score that represents the overall stability of the slope. The health score is then sent to the early warning terminal to trigger the corresponding safety control logic.

[0120] In one specific embodiment, a risk level severity mapping vector is preset. Each component corresponds to a basic state score ranging from low risk to extremely high risk, and is a dimensionless value. The preset logic of this score sequence follows a non-linear decreasing principle, that is, as the risk level increases, the slope of the score decrease gradually increases, aiming to enhance the sensitivity of the early warning by widening the score gap in the high-risk range.

[0121] Next, the system calculates the health score by performing an inner product operation between the global membership probability and the severity mapping vector, and by combining the environmental sensitivity correction coefficient. The calculation formula is: .in, The overall health score of the slope is typically between 0 and 100. The global membership probability of the slope as a whole at the k-th risk assessment level is output by the aforementioned weighted evidence reasoning model, and is a dimensionless coefficient that satisfies... ; This corresponds to the score in the aforementioned severity mapping vector; This is the preset environmental sensitivity correction coefficient, which is a dimensionless coefficient with a dimension defined as 1. The acquisition logic is dynamically adjusted based on the current climate season or the complexity of the geological environment of the slope: in the non-rainy season or in areas with simple geological structures, The default value is 1.0; if the system detects that it is currently in a preset rainy season (such as June to August each year) or that there is a known fault in the area, then it will... The value is adjusted to an adjustable range of 0.85 to 0.95, preferably 0.90.

[0122] After generating a health score, the system sends it to the early warning terminal, which then triggers the corresponding logic based on a preset tiered control strategy. The specific control logic is preset as follows: When... When the score is greater than or equal to 85, it is judged as a green safety state, and the regular monitoring frequency of once every 2 hours is maintained; when When the location is within the [60, 85) range, it is determined to be in a yellow alert state, triggering an SMS notification to the relevant responsible person, and the monitoring frequency is increased to once every 1 hour; when When the location is within the [40, 60) range, it is determined to be an orange warning state. The system automatically plays a warning voice through the loudspeaker terminal and requests on-site personnel to stop non-essential operations; when If the score is less than 40 minutes, it is determined to be a red emergency state, and the automated gate leading to the slope impact area is immediately closed, triggering the personnel evacuation order.

[0123] Example 2:

[0124] like Figure 3As shown, the intelligent slope health assessment system based on multi-source data fusion includes a multi-source data acquisition module, a risk vector construction module, a risk weighted calculation module, and a slope health assessment module.

[0125] The various modules are connected via wired and / or wireless connections to enable data transmission between them;

[0126] Multi-source data acquisition module: Receives multi-source slope monitoring data collected by multiple monitoring terminals. The multi-source slope monitoring data includes displacement time series and rainfall environmental data.

[0127] Risk vector construction module: The multi-source monitoring data of the slope is segmented using a sliding window with a predetermined step size. The rate of change of displacement time series is extracted by least squares regression calculation. Combined with rainfall environmental data, a multi-dimensional feature matrix for a single monitoring terminal is formed. Each feature value in the multi-dimensional feature matrix is ​​mapped to a preset Gaussian membership function to calculate the initial confidence distribution of the monitoring terminal under different risk levels, and a risk confidence vector is constructed.

[0128] Risk weighting calculation module: retrieves pre-stored weight configuration information, which includes relative weight values ​​assigned to sensors of different monitoring dimensions and locations using the analytic hierarchy process; applies the relative weight values ​​to perform weighted mapping on the risk confidence vector to generate weighted evidence items characterizing the risk contribution of a local area;

[0129] The slope health assessment module inputs weighted evidence items from multiple monitoring terminals into a weighted evidence inference model. Within the model, evidence conflict identification and joint probability synthesis are performed. An iterative synthesis algorithm eliminates data conflicts between different monitoring dimensions, and the global membership probability of the slope at each risk assessment level is calculated. The global membership probability is then quantitatively weighted and summed to generate a health score that characterizes the overall stability of the slope. This health score is then sent to the early warning terminal to trigger the corresponding safety control logic.

[0130] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using 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.

[0131] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0132] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations 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. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0133] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent assessment of slope health based on multi-source data fusion, characterized in that, include: It receives multi-source slope monitoring data collected by multiple monitoring terminals, which includes displacement time series and rainfall environmental data; The multi-source monitoring data of the slope is segmented by a sliding window with a predetermined step size. The rate of change of displacement time series is extracted by least squares regression calculation. Combined with rainfall environmental data, a multi-dimensional feature matrix is ​​formed for a single monitoring terminal. Each eigenvalue in the multidimensional feature matrix is ​​mapped to a preset Gaussian membership function to calculate the initial confidence distribution of the monitoring terminal under different risk levels, and a risk confidence vector is constructed. Retrieve the pre-stored weight configuration information, which includes the relative weight values ​​assigned to sensors of different monitoring dimensions and locations using the analytic hierarchy process (AHP). The risk confidence vector is weighted by applying relative weight values ​​to generate a weighted evidence term that represents the risk contribution of a local area; The weighted evidence items corresponding to multiple monitoring terminals are input into the weighted evidence reasoning model. Evidence conflict identification and joint probability synthesis are performed inside the weighted evidence reasoning model. Data conflicts between different monitoring dimensions are eliminated through iterative synthesis algorithm, and the global membership probability of the slope as a whole at each risk assessment level is calculated. The global membership probability is quantitatively weighted and summed to generate a health score that represents the overall stability of the slope, and the health score is sent to the early warning terminal to trigger the corresponding safety control logic; The global membership probability of the slope as a whole at each risk assessment level is calculated, including: The reliability assignment values ​​corresponding to different risk assessment levels in each weighted evidence item are extracted, and the basic reliability quality assigned to each risk assessment level and the residual uncertainty quality not assigned to a specific level are calculated based on the relative weight of each monitoring terminal. By performing a cross-product summation operation on the basic reliability quality among different monitoring terminals, the mutual exclusivity of different monitoring terminals under the same risk assessment level is identified, and a conflict interference factor used to characterize the degree of evidence contradiction is calculated. The recursive synthesis algorithm is adopted, taking the basic reliability quality of the first monitoring terminal as the initial fusion state, and introducing the basic reliability quality of subsequent monitoring terminals one by one. In each iteration, the cross-product result is adjusted by the conflict interference factor to generate the joint probability distribution of the intermediate state. After completing the iterative synthesis of all monitoring terminals, the final generated joint probability distribution and the synthesized residual uncertainty mass are extracted. By performing normalization mapping, the residual uncertainty mass is proportionally redistributed to each risk assessment level. The allocated values ​​are defined as the global membership probability of the slope as a whole at each risk assessment level. The basic reliability quality assigned to each risk assessment level and the residual uncertainty quality not assigned to a specific level are calculated, including: For the weighted evidence items, the corresponding reliability allocation values ​​are extracted according to the five preset risk assessment levels; The reliability allocation value corresponding to each risk assessment level is multiplied by the relative weight of the monitoring terminal to generate the basic reliability quality for each risk assessment level. The difference between the numerical value 1 and the relative weight is calculated to obtain the weight discount quality component, which characterizes the dilution of the trust level of the monitoring terminal by the weight. Perform a cumulative summation operation on the basic reliability quality corresponding to all risk assessment levels, and calculate the difference between the value 1 and the cumulative summation result to obtain the inherent uncertainty quality component that characterizes the fuzziness of the monitoring data itself; The weighted discounted mass component and the inherently uncertain mass component are linearly summed to generate the residual uncertainty mass that is not assigned to a specific level.

2. The intelligent slope health assessment method based on multi-source data fusion according to claim 1, characterized in that, The steps for obtaining the multidimensional feature matrix of a single monitoring terminal are as follows: The sliding window moves along the time axis according to a predetermined step size. The sliding window is used to extract the displacement time series in the multi-source monitoring data of the slope and obtain the displacement time data segment within the time range covered by the current window. By using a sliding window, rainfall environmental data from multi-source slope monitoring data are captured synchronously to obtain rainfall environmental data segments that are consistent with the time points of the displacement time data segments. A univariate linear fitting equation is established using each time point in the displacement time data segment as the independent variable and the displacement acquisition value corresponding to each time point as the dependent variable. The slope of the univariate linear fitting equation that minimizes the sum of squared residuals of all displacement data is calculated using the least squares regression algorithm. The slope of the univariate linear fitting equation is then set as the rate of change characteristic of the displacement time series. Extract the rainfall values ​​at each monitoring time point in the rainfall environment data segment, and sum all the rainfall values ​​within the rainfall environment data segment to obtain the cumulative rainfall characteristics corresponding to the sliding window; The rate of change features and cumulative rainfall features are dimensionally aligned and array-concatenated according to the physical identifier of each monitoring terminal to construct a multi-dimensional feature matrix for each monitoring terminal.

3. The intelligent slope health assessment method based on multi-source data fusion according to claim 1, characterized in that, The construction logic of the risk confidence vector includes: Extract the rate of change features and cumulative rainfall features from the multidimensional feature matrix to obtain the preset level center value corresponding to each risk level; Calculate the squared deviations between the rate of change characteristics, cumulative rainfall characteristics, and the preset center values ​​corresponding to each risk level; The squared deviation value is scaled using a preset standard deviation, and the scaled value is used as the exponent of the natural logarithm base to perform a negative operation to obtain the membership score of each feature for different risk levels. Based on the numerical range of the rate of change characteristics and cumulative rainfall characteristics, the membership scores at both ends of the risk level sequence are saturated and smoothed to obtain the corrected membership scores of each feature under different risk levels. The membership scores of each feature of the same monitoring terminal under the same risk level are averaged to generate the initial reliability distribution of the monitoring terminal under different risk levels. The initial confidence distributions corresponding to each risk level are arranged in a predetermined order from low to high risk to construct a risk confidence vector.

4. The intelligent slope health assessment method based on multi-source data fusion according to claim 3, characterized in that, The calculation process for the adjusted membership scores of each feature under different risk levels includes the following steps: Determine the lowest risk level at the beginning and the highest risk level at the end of the risk level sequence; The change rate feature and cumulative rainfall feature are compared with the preset level center value corresponding to the lowest risk level. If the feature value is less than the preset level center value corresponding to the lowest risk level, the membership score corresponding to the lowest risk level is updated to the predetermined constant peak value. The change rate feature and cumulative rainfall feature are compared with the preset level center value corresponding to the highest risk level. If the feature value is greater than the preset level center value corresponding to the highest risk level, the membership score corresponding to the highest risk level is updated to the predetermined constant peak value. Keeping the membership scores of intermediate risk levels (excluding the lowest and highest risk levels) unchanged in the risk level sequence, the updated membership scores are combined with the unchanged membership scores to generate modified membership scores for each feature under different risk levels.

5. The intelligent slope health assessment method based on multi-source data fusion according to claim 1, characterized in that, The default logic for weight configuration information is as follows: Establish a reference sample set containing multiple historical slope instability events. For each historical instability event, extract the displacement time series, rainfall environment data, and corresponding slope stability state labels within a predetermined time period before instability. The reference sample set is divided into a deformation feature layer, a rainfall environment layer, and a location distribution layer according to the monitoring dimensions. Based on the correlation contribution of each monitoring dimension to the slope stability state label, an initial judgment matrix for pairwise comparison is constructed. The maximum eigenvalue and consistency ratio of the matrix are set as iteration parameters. The power method is used to iteratively calculate the initial judgment matrix to solve the eigenvector of the initial judgment matrix at the current iteration number. The reference sample set is weighted using the components of each dimension in the feature vector to obtain the sample risk assessment value, and the mean square error between the sample risk assessment value and the slope stability status label is calculated. The elements in the initial judgment matrix are adjusted using the mean square error value, and the consistency ratio value in the current iteration parameters is calculated simultaneously. If the consistency ratio is less than the preset consistency threshold and the mean square error reaches the global minimum distribution, stop the iteration and extract the current feature vector as the target weight vector. The target weight vector is mapped and encapsulated according to the physical index of sensors at different monitoring dimensions and locations to generate weight configuration information.

6. The intelligent slope health assessment method based on multi-source data fusion according to claim 1, characterized in that, A recursive synthesis algorithm is adopted, using the basic reliability quality of the first monitoring terminal as the initial fusion state, and then introducing the basic reliability quality of subsequent monitoring terminals one by one. In each iteration, the cross-product result is adjusted for gain using a conflict interference factor to generate the joint probability distribution of the intermediate state, including: Assign the basic reliability quality of the first monitoring terminal at each risk assessment level to the current fusion probability vector as the initial fusion state for recursive calculation; Extract the components in the current fusion probability vector and the basic reliability quality of the next monitoring terminal to be fused at the corresponding risk assessment level, and calculate the cross product value of the two. Calculate the difference between the value 1 and the interference factor, and set the reciprocal of this difference as the gain correction coefficient; The cross-product values ​​are amplified by a gain correction factor to obtain the updated current fusion probability vector, thus completing the evidence synthesis for the current round. The extraction, calculation, and amplification steps are performed cyclically according to the order of the monitoring terminals until the basic reliability quality of all monitoring terminals has been fused. The final fused probability vector is then determined as the joint probability distribution of the intermediate state.

7. A slope health intelligent assessment system based on multi-source data fusion, characterized in that, It is implemented based on the intelligent slope health assessment method based on multi-source data fusion as described in any one of claims 1-6, and includes: Multi-source data acquisition module: Receives multi-source slope monitoring data collected by multiple monitoring terminals. The multi-source slope monitoring data includes displacement time series and rainfall environmental data. Risk vector construction module: The multi-source monitoring data of the slope is segmented using a sliding window with a predetermined step size. The rate of change of displacement time series is extracted by least squares regression calculation. Combined with rainfall environmental data, a multi-dimensional feature matrix for a single monitoring terminal is formed. Each feature value in the multi-dimensional feature matrix is ​​mapped to a preset Gaussian membership function to calculate the initial confidence distribution of the monitoring terminal under different risk levels, and a risk confidence vector is constructed. Risk weighting calculation module: retrieves pre-stored weight configuration information, which includes relative weight values ​​assigned to sensors of different monitoring dimensions and locations using the analytic hierarchy process; applies the relative weight values ​​to perform weighted mapping on the risk confidence vector to generate weighted evidence items characterizing the risk contribution of a local area; The slope health assessment module inputs weighted evidence items from multiple monitoring terminals into a weighted evidence inference model. Within the model, evidence conflict identification and joint probability synthesis are performed. An iterative synthesis algorithm eliminates data conflicts between different monitoring dimensions, and the global membership probability of the slope at each risk assessment level is calculated. The global membership probability is then quantitatively weighted and summed to generate a health score that characterizes the overall stability of the slope. This health score is then sent to the early warning terminal to trigger the corresponding safety control logic.

Images