Intelligent analysis system for slope state based on multi-source monitoring data fusion

Abstract

The present application relates to the technical field of electric digital data processing, in particular to a slope state intelligent analysis system based on multi-source monitoring data fusion. The system acquires multi-source time series monitoring data of displacement, rainfall and soil moisture content; constructs a space-time heterogeneous graph, defines different types of sensors as heterogeneous nodes, the node attributes are corresponding time series characteristics, and the edge types between nodes reflect the water-rock coupling physical correlation mechanism; adopts a graph attention network based on physical prior constraints, takes the water-rock coupling evolution law as a bias term of attention coefficient calculation, and aggregates the heterogeneous node features; extracts the dynamic evolution characteristics of the heterogeneous nodes under physical constraints through a time series convolution network; and inputs the graph-level representation vector into a fully connected layer to output the slope state classification result. The present application overcomes the feature interference and loss of physical meaning caused by direct splicing of heterogeneous data, improves the expression accuracy of the fused features on the physical evolution process of the slope, and reduces the state misjudgment caused by time series misalignment.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically to an intelligent slope condition analysis system based on the fusion of multi-source monitoring data. Background Technology

[0002] Slope condition monitoring is a crucial means of geological disaster prevention. Existing technologies typically deploy various monitoring devices in slope areas, such as displacement gauges, rain gauges, and soil moisture sensors, to continuously collect multi-source time-series monitoring data. In the data processing stage, conventional solutions generally employ feature-level stitching or simple decision-level fusion of multi-source data, inputting the physical parameters collected by each sensor as independent feature channels into a deep neural network for feature extraction and state classification; or they use graph-based data processing, treating each sensor as a homogeneous node in a graph structure, and aggregating data through a homogeneous graph attention mechanism between nodes. This conventional approach assumes that the attributes of each sensor node are homogeneous and that the physical relationships between nodes are symmetrical and undifferentiated, directly performing fusion calculations at the data level or shallow feature level.

[0003] In the actual geological evolution of slopes, different physical parameters play different roles in the disaster-causing process, and their spatiotemporal evolution scales are fundamentally different. Rainfall exhibits rapid, pulse-like characteristics, soil moisture content reflects the water infiltration process and lags behind rainfall in time, while displacement exhibits slow, cumulative characteristics. Existing technologies directly stitch together features from multiple sources or treat them as isomorphic node aggregations, failing to consider the temporal misalignment between rapidly changing rainfall signals and slowly changing displacement signals, and also failing to distinguish the heterogeneity and directionality of rainfall infiltration and matrix suction reduction in the physical correlation of water-rock coupling. Conventional isomorphic map structures cannot characterize the causal time-lag chain of rainfall driving moisture content changes, moisture content reducing soil strength, and ultimately triggering displacement, causing the fusion process of multi-source data to deviate from the actual physical evolution mechanism.

[0004] Because existing technologies neglect the physical mechanisms of water-rock coupling and time-delay effects in multi-source heterogeneous data, the spatiotemporal scale mismatch and physical mechanism disconnect between heterogeneous sensor data lead to the failure of fusion feature extraction. Direct fusion of rapidly changing pulse signals and slowly changing cumulative signals without physical alignment results in the features of the rapidly changing signals being easily averaged by the slowly changing signals, or the trends of the slowly changing signals being submerged by the rapidly changing signals, causing feature masking and interference. This physical mechanism disconnect prevents subsequent feature extraction networks from capturing the dynamic evolution of the disaster-causing forces, leading to distorted slope condition analysis results. The spatiotemporal scale mismatch and physical mechanism disconnect between heterogeneous multi-source data, resulting in fusion failure, constitute the core technical problem currently faced by the technology. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent slope condition analysis system based on the fusion of multi-source monitoring data, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The intelligent slope condition analysis system based on multi-source monitoring data fusion includes: a data interface for acquiring multi-source time-series monitoring data containing displacement, rainfall, and soil moisture content; a graph builder for constructing a spatiotemporal heterogeneous graph based on the multi-source time-series monitoring data, defining different types of sensors as heterogeneous nodes, with the attributes of the heterogeneous nodes representing corresponding time-series features, and the edge types between heterogeneous nodes reflecting the physical correlation mechanism of water-rock coupling; a feature aggregator for aggregating the features of heterogeneous nodes in the spatiotemporal heterogeneous graph using a graph attention network based on physical prior constraints, taking the water-rock coupling evolution law as a bias term in the attention coefficient calculation; a feature extractor for extracting the dynamic evolution features of the heterogeneous nodes under physical constraints through a temporal convolutional network; and a classifier for inputting the aggregated graph-level representation vector into a fully connected layer and outputting the slope condition classification result.

[0007] Preferably, when the graph builder constructs the edge types reflecting the physical correlation mechanism of water-rock coupling in the spatiotemporal heterogeneous graph, it constructs directed heterogeneous edges between heterogeneous nodes based on the time lag effect of rainfall infiltration and displacement response in geomechanics and the reduction law of soil strength due to water content changes. The directed heterogeneous edges include infiltration edges from rainfall nodes to water content nodes and softening edges from water content nodes to displacement nodes. The graph builder dynamically configures the initial values ​​of the edge weights of the directed heterogeneous edges according to the lag time of the extreme points of mutual information between different physical parameters in the multi-source time-series monitoring data.

[0008] Preferably, when the feature aggregator uses the water-rock coupling evolution law as a bias term in the attention coefficient calculation, it derives the functional relationship between water content and matrix suction based on the unsaturated soil seepage mechanics equation and maps the functional relationship into a prior bias matrix in the attention calculation; when calculating the attention coefficient between the target node and its neighboring nodes in the spatiotemporal heterogeneous graph, it performs a Hadamard product operation on the prior bias matrix and the original attention score calculated based on node features to reconstruct the feature transfer weight distribution of the neighboring nodes to the target node.

[0009] Preferably, when the feature extractor extracts dynamic evolution features through a temporal convolutional network, it adopts a multi-layer dilated causal convolutional structure. Based on the slow variation characteristics of displacement changes and the fast variation characteristics of rainfall pulses in the multi-source temporal monitoring data, it configures differentiated dilation coefficients for different types of heterogeneous nodes. For displacement nodes, a large-scale dilation coefficient is configured to expand the temporal receptive field, and for rainfall nodes, a small-scale dilation coefficient is configured to capture instantaneous pulse features. The output results of each layer of dilated causal convolution are fused through residual connections.

[0010] Preferably, before inputting the graph-level representation vector into the fully connected layer, the classifier performs a global pooling operation by introducing heterogeneous node importance weights to aggregate node-level features into a graph-level representation vector; the importance weights are jointly determined based on the mean of the attention coefficients output by the graph attention network and the degree centrality index of the heterogeneous nodes; the fully connected layer uses a flexible maximum activation function to map the graph-level representation vector to the state space of slope stability, slope creep, and slope instability, and outputs the slope state classification probability distribution.

[0011] Preferably, after the graph builder dynamically configures the initial values ​​of the edge weights of the directed heterogeneous edges, an adaptive edge weight update mechanism is introduced during the feature aggregation iteration of the graph attention network; based on the gradient variance of the feature transmission of neighboring nodes in the current iteration round, the edge weights of the directed heterogeneous edges are dynamically attenuated or enhanced; when the gradient variance exceeds a preset threshold, it is determined that the corresponding directed heterogeneous edge is affected by environmental noise, and an attenuation factor is applied to the edge weights to suppress noise transmission, thereby realizing the dynamic evolution of the topology.

[0012] Preferably, when calculating the attention coefficient, a data-driven residual term is introduced to adaptively correct the prior bias matrix; an abnormal coupling mode deviating from the unsaturated soil seepage mechanics equation is captured in the multi-source time-series monitoring data through a learnable parameterized matrix; the parameterized matrix is ​​added to the prior bias matrix to obtain a comprehensive bias matrix that integrates the physical mechanism and data residuals; the comprehensive bias matrix is ​​used to replace the prior bias matrix in the Hadamard product operation to correct the feature transfer weight distribution under local abnormal conditions.

[0013] Preferably, when the feature extractor fuses the output results of each layer of dilated causal convolution, a cross-scale time alignment fusion mechanism is adopted; for the output sequences corresponding to different dilation coefficients, the time dimension is pruned and aligned according to the core scale benchmark; a channel attention mechanism is applied to the aligned multi-scale feature sequence, and the channel weights are redistributed based on the response sensitivity of each channel feature to the slope state change, highlighting the feature channels with high response sensitivity and suppressing redundant channels, and the aligned and weighted multi-scale features are spliced ​​into dynamic evolution features.

[0014] Preferably, after the data interface acquires the multi-source time-series monitoring data, it introduces a graph-based data missing compensation mechanism; when a data missing is detected in the target sensor node within the target time window, the set of neighboring nodes of the target sensor node in the spatiotemporal heterogeneous graph is located; the time series features of the set of neighboring nodes within the target time window are extracted, and combined with the edge type between the target sensor node and each neighboring node, the missing data of the target sensor node is reconstructed through a bidirectional long short-term memory network.

[0015] Preferably, after the classifier outputs the slope state classification probability distribution, an early warning decision mechanism based on cognitive uncertainty is introduced; the Monte Carlo random deactivation method is used to perform multiple forward propagation inferences on the feature extractor and the classifier to obtain the state category variance of the multiple inference results; if the state category variance is greater than the cognitive uncertainty threshold, a sensor verification instruction is triggered to reacquire multi-source time-series monitoring data; if the state category variance is less than the cognitive uncertainty threshold, an early warning decision is executed based on the maximum value in the slope state classification probability distribution.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a spatiotemporal heterogeneous graph, defining different types of sensors as heterogeneous nodes. Based on the time-delay effect of rainfall infiltration and displacement response, and the reduction law of soil strength by water content, it constructs directed heterogeneous edges and initial values ​​of edge weights that reflect the physical correlation mechanism of water-rock coupling. This overcomes the feature interference and loss of physical meaning caused by direct splicing of heterogeneous data, and improves the accuracy of the fused features in expressing the physical evolution process of slopes. Through a graph attention network based on physical prior constraints, the water-rock coupling evolution law is mapped to a prior bias matrix and subjected to Hadamard product with the original attention scores, realizing deep interaction of multi-source heterogeneous data at the physical mechanism level, reducing state misjudgments caused by temporal misalignment due to differences in sensor sampling frequencies. By configuring differentiated expansion coefficients for different types of heterogeneous nodes through a temporal convolutional network, instantaneous rainfall pulse features and slowly varying cumulative displacement features are captured respectively. Global pooling is performed in combination with the importance weights of heterogeneous nodes, improving the robustness of multi-source data coupling feature extraction under extreme rainfall conditions, and overcoming the technical bottleneck of limited generalization ability of traditional black-box fusion models under complex geological conditions.

[0017] 2. In the feature aggregation iteration process, this invention dynamically updates the edge weights of directed heterogeneous edges based on the gradient variance of neighbor node feature propagation. An attenuation factor is applied to edge weights affected by environmental noise, suppressing noise propagation and realizing dynamic evolution of the topology, thus improving the noise resistance of the analysis model. By introducing a learnable parameterized matrix to capture abnormal coupling patterns deviating from the physical equations, the feature propagation weight distribution under local abnormal conditions is corrected, overcoming the limitations of pure physical priors under abnormal conditions. A cross-scale time alignment fusion mechanism and a channel attention mechanism are employed to suppress redundant feature channels, highlighting features with high sensitivity to slope state changes. A graph-based data missing compensation mechanism reconstructs missing data using the time-series features of neighbor node sets and edge types, ensuring the integrity of multi-source time-series monitoring data. A cognitive uncertainty-based early warning decision-making mechanism obtains the state category variance through a random discarding method. When the variance exceeds a threshold, a sensor verification command is triggered, avoiding false alarms caused by model cognitive uncertainty and improving the system's decision-making reliability under complex conditions. Attached Figure Description

[0018] Figure 1 This is a flowchart of the overall end-to-end data processing of the system according to the present invention; Figure 2 This is a flowchart of the spatiotemporal heterogeneous graph construction and directed heterogeneous edge weight initialization process of the present invention. Figure 3 This is a flowchart of the attention bias matrix generation and coefficient calculation for physical prior superposition data residual correction in this invention; Figure 4 This is a flowchart of the differentiated dilated convolution configuration and cross-scale feature alignment and fusion processing of the present invention; Figure 5 This is a flowchart illustrating the node-weighted global pooling and fully connected layer slope state classification output of the present invention. Detailed Implementation

[0019] refer to Figure 1In one embodiment, the intelligent slope condition analysis system based on multi-source monitoring data fusion includes a data interface, a graph builder, a feature aggregator, a feature extractor, and a classifier. The data interface establishes communication connections with displacement gauges, rain gauges, and soil moisture sensors deployed in the slope area to acquire multi-source time-series monitoring data including displacement, rainfall, and soil moisture content. The graph builder receives the multi-source time-series monitoring data output from the data interface, constructs a spatiotemporal heterogeneous graph, defines different types of sensors as heterogeneous nodes, assigns attributes to corresponding time-series features, and reflects the physical correlation mechanism of water-rock coupling between heterogeneous nodes. The feature aggregator uses a graph attention network based on physical prior constraints, employing the water-rock coupling evolution law as a bias term in the attention coefficient calculation to aggregate the features of heterogeneous nodes in the spatiotemporal heterogeneous graph. The feature extractor extracts the dynamic evolution features of heterogeneous nodes under physical constraints through a temporal convolutional network. The classifier inputs the aggregated graph-level representation vector into a fully connected layer and outputs the slope condition classification result.

[0020] In this embodiment, the data interface establishes a connection with various sensors in the slope monitoring network via wired or wireless communication, acquiring multi-source time-series monitoring data according to a preset sampling frequency. Displacement gauges collect displacement data from the slope surface or interior, with a sampling frequency set to once per hour; rain gauges collect rainfall data from the slope area, with a sampling frequency set to once per minute; soil moisture sensors collect moisture content data from soil layers at different depths, with a sampling frequency set to once every ten minutes. The data interface performs preliminary preprocessing on the acquired raw monitoring data, including data format conversion, outlier removal, and timestamp unification. Outlier removal employs… The criteria are as follows: data exceeding three standard deviations from the mean are marked as outliers and removed. Timestamp unification aligns sensor data from different sampling frequencies onto a unified time axis, forming a multi-source time-series data sequence with ten-minute intervals.

[0021] The graph builder receives preprocessed multi-source time-series monitoring data and constructs a spatiotemporal heterogeneous graph. The spatiotemporal heterogeneous graph is defined as G=(V,E,A), where V represents the set of nodes, E represents the set of edges, and A represents the node attribute matrix. The node set V contains three types of heterogeneous nodes: displacement nodes... Rainfall nodes and moisture content nodes Displacement nodes All displacement gauges deployed in the corresponding slope area, each displacement gauge corresponds to a displacement node; rainfall nodes All rain gauges deployed in the corresponding slope area, each rain gauge corresponds to a rainfall node; moisture content node All soil moisture sensors deployed in the corresponding slope area are represented, with each sensor corresponding to a moisture content node. Each row of the node attribute matrix A corresponds to an attribute vector for a node, where the attribute vector is the time-series data sequence collected by the corresponding sensor over T consecutive time steps. For displacement nodes, the attribute vector is... ,in This represents the displacement measurement value at time step t; for the rainfall node, the attribute vector is... ,in This represents the rainfall measurement at time step t; for the moisture content node, the attribute vector is... ,in This represents the moisture content measurement value at time step t.

[0022] Edge set E contains directed heterogeneous edges reflecting the physical correlation mechanism of water-rock coupling. This mechanism manifests as rainfall infiltration leading to increased soil moisture content, which in turn reduces soil strength, ultimately causing slope displacement. Based on this mechanism, the graph builder constructs two types of directed heterogeneous edges: infiltration edges (where rainfall nodes point to moisture content nodes) and softening edges (where moisture content nodes point to displacement nodes). Infiltration edges indicate that rainfall affects soil moisture content through infiltration, while softening edges indicate that changes in soil moisture content affect slope displacement through soil softening. Edge set E does not contain edges where displacement nodes point to other nodes, nor does it contain edges where rainfall nodes directly point to displacement nodes, to conform to the causal relationship of water-rock coupling.

[0023] The feature aggregator employs a graph attention network based on physical prior constraints to aggregate the features of heterogeneous nodes in a spatiotemporally heterogeneous graph. The graph attention network contains multiple attention layers. Each layer calculates the attention coefficients between the target node and its neighbors, and then performs a weighted summation of the features of the neighboring nodes to obtain the aggregated features of the target node. When calculating the attention coefficients, the feature aggregator uses the water-rock coupling evolution law as a bias term to correct the original attention score. The water-rock coupling evolution law is derived from the unsaturated soil seepage mechanics equations and reflects the functional relationship between water content and matrix suction. The feature aggregator maps this functional relationship to a prior bias matrix in the attention calculation. When calculating the attention coefficients between the target node and its neighbors, it performs a Hadamard product operation between the prior bias matrix and the original attention score calculated based on node features to reconstruct the feature transfer weight distribution from neighboring nodes to the target node.

[0024] The feature extractor extracts the dynamic evolution features of heterogeneous nodes under physical constraints using a temporal convolutional network. This network employs a multi-layer dilated causal convolutional structure, which expands the temporal receptive field without increasing the number of parameters. Causal convolution ensures that the output at time step t depends only on the input at time step t and earlier, conforming to the causal relationship of temporal data. Dilated convolution expands the effective receptive field of the convolutional kernel by inserting zero values ​​between kernel elements. The feature extractor configures differentiated dilation coefficients for different types of heterogeneous nodes to adapt to the spatiotemporal evolution characteristics of different physical parameters. For displacement nodes, since displacement changes have a slowly cumulative characteristic, a large-scale dilation coefficient is configured to expand the temporal receptive field and capture long-term displacement trends; for rainfall nodes, since rainfall exhibits a rapidly changing pulse characteristic, a small-scale dilation coefficient is configured to capture instantaneous rainfall pulse features. The feature extractor fuses the outputs of each layer of dilated causal convolutions through residual connections to obtain the dynamic evolution features of the heterogeneous nodes.

[0025] The classifier receives the dynamically evolving features output by the feature extractor, aggregates them into a graph-level representation vector, and inputs it into a fully connected layer for state classification. The classifier first aggregates node-level features into a graph-level representation vector through a global pooling operation. This global pooling operation introduces heterogeneous node importance weights, performing a weighted summation of features from different node types. The heterogeneous node importance weights are jointly determined based on the mean of the attention coefficients output by the graph attention network and the degree centrality index of the heterogeneous nodes. The classifier inputs the graph-level representation vector into the fully connected layer, which uses a flexible maximum activation function to map the graph-level representation vector to the state spaces of slope stability, slope creep, and slope instability, outputting a slope state classification probability distribution.

[0026] In this embodiment, the system workflow is as follows: The data interface acquires multi-source time-series monitoring data and performs preliminary preprocessing; the graph builder constructs a spatiotemporal heterogeneous graph based on the preprocessed multi-source time-series monitoring data, defining three types of heterogeneous nodes and two types of directed heterogeneous edges; the feature aggregator uses a graph attention network based on physical prior constraints to aggregate the features of heterogeneous nodes in the spatiotemporal heterogeneous graph; the feature extractor extracts the dynamic evolution features of heterogeneous nodes through a temporal convolutional network; the classifier aggregates the dynamic evolution features into a graph-level representation vector and inputs it into a fully connected layer to output the slope state classification result.

[0027] refer to Figure 2In a preferred embodiment, when constructing edge types reflecting the physical correlation mechanism of water-rock coupling in a spatiotemporal heterogeneous graph, the graph builder constructs directed heterogeneous edges between heterogeneous nodes based on the time lag effect of rainfall infiltration and displacement response in geomechanics, as well as the reduction law of soil strength due to water content changes. These directed heterogeneous edges include infiltration edges from rainfall nodes to water content nodes, and softening edges from water content nodes to displacement nodes. The graph builder dynamically configures the initial values ​​of the edge weights of the directed heterogeneous edges based on the lag time of extreme points of mutual information between different physical parameters in multi-source time-series monitoring data.

[0028] In this embodiment, the time lag effect of rainfall infiltration and displacement response manifests as follows: after a change in rainfall, the soil moisture content needs a certain amount of time to reach its peak, and the slope displacement response also needs a certain amount of time to reach its peak. The reduction law of soil strength due to changes in moisture content shows that the higher the soil moisture content, the lower the soil shear strength, and the more easily the slope will displace. Based on these two geomechanical laws, the graph builder constructs directed heterogeneous edges to clarify the causal relationship and direction of action between different physical parameters.

[0029] The infiltration edge connects the rainfall node and the moisture content node, pointing from the rainfall node to the moisture content node, indicating that changes in rainfall cause changes in moisture content. The existence of the infiltration edge reflects the process of rainfall infiltration, where rainwater infiltrates into the soil through the surface, leading to an increase in soil moisture content. The softening edge connects the moisture content node and the displacement node, pointing from the moisture content node to the displacement node, indicating that changes in moisture content cause changes in displacement. The existence of the softening edge reflects the process of soil softening, where increased soil moisture content leads to a decrease in soil shear strength, causing slope displacement under gravity.

[0030] The graph builder dynamically configures the initial edge weights of directed heterogeneous edges based on the lag time of extreme points of mutual information between different physical parameters in multi-source time-series monitoring data. Mutual information measures the degree of dependence between two random variables; a higher mutual information value indicates a stronger dependence between the two variables. For a rainfall sequence R and a moisture content sequence W, different lag times are calculated. Mutual information under The formula for calculating mutual information is:

[0031] in, This indicates that the rainfall sequence R takes the value r at time t and the water content sequence W takes the value r at time t. The joint probability of taking the value w is given by , and p(r) represents the marginal probability of the rainfall sequence R taking the value r. This indicates that the moisture content sequence W represents the time... The marginal probability is w.

[0032] Calculate different lag times Mutual information under Find the lag time corresponding to the maximum mutual information. This lag time is the average time delay between rainfall infiltration and the resulting change in water content. Similarly, for the water content sequence W and the displacement sequence D, different lag times are calculated. Mutual information under Find the lag time corresponding to the maximum mutual information. This lag time is the average time delay of the displacement change caused by the change in water content.

[0033] Table 1 shows the statistical results of the mutual information extreme point lag time for different slope regions.

[0034] Table 1. Statistics on the lag time of mutual information extreme points in different slope regions.

[0035] The graph builder dynamically configures the initial values ​​of edge weights for directed heterogeneous edges based on the lag time of mutual information extrema. For infiltration edges, the initial values ​​of edge weights are... With lag time The effect is inversely proportional; that is, the shorter the lag time, the more direct the impact of rainfall on water content, and the larger the initial value of the edge weight. The formula for calculating the initial value of the edge weight of an infiltration edge is:

[0036] in, To prevent small constants with zero denominators, the value is set to 0.1.

[0037] For softened edges, the initial values ​​of the edge weights are... With lag time The effect is inversely proportional; that is, the shorter the lag time, the more direct the influence of water content on displacement, and the larger the initial value of the edge weight. The formula for calculating the initial value of the edge weight of a softened edge is:

[0038] in, To prevent small constants with zero denominators, the value is set to 0.1.

[0039] In this embodiment, the graph builder dynamically configures the initial values ​​of the edge weights of directed heterogeneous edges by calculating the extreme point lag time of mutual information between different physical parameters. This allows the edge weights to reflect the actual coupling strength and time delay characteristics between different physical parameters. This dynamic configuration method avoids the problem that fixed edge weights cannot adapt to different slope geological conditions and improves the ability of spatiotemporal heterogeneous graphs to express actual physical processes.

[0040] In a preferred embodiment, after the graph builder dynamically configures the initial values ​​of the edge weights of directed heterogeneous edges, an adaptive edge weight update mechanism is introduced during the feature aggregation iteration of the graph attention network. Based on the gradient variance of the feature propagation of neighboring nodes in the current iteration, the edge weights of the directed heterogeneous edges are dynamically attenuated or enhanced. When the gradient variance exceeds a preset threshold, it is determined that the corresponding directed heterogeneous edge is affected by environmental noise, and an attenuation factor is applied to the edge weight to suppress noise propagation, thereby realizing the dynamic evolution of the topology.

[0041] In this embodiment, the edge weight adaptive update mechanism is executed after each training iteration of the graph attention network. For each directed heterogeneous edge e, the gradient variance of the neighboring nodes connected to that edge when transmitting features to the target node in the current iteration is calculated. Gradient variance reflects the degree of fluctuation in the feature transfer process. The larger the gradient variance, the more severe the noise interference in the feature transfer process.

[0042] The gradient variance is calculated as follows: During the backpropagation process of the current iteration, record the sequence of gradient values ​​transmitted from neighboring nodes to the target node via edge e. , where n is the number of samples in the current batch. Calculate the mean of this gradient value sequence. and variance :

[0043]

[0044] Preset gradient variance threshold When the gradient variance of edge e Greater than the threshold When an edge is deemed severely affected by environmental noise, an attenuation factor needs to be applied to its weight to suppress noise propagation. Attenuation factor With gradient variance The attenuation factor is directly proportional to the gradient variance; that is, the larger the gradient variance, the smaller the attenuation factor, and the greater the attenuation of the edge weights. The formula for calculating the attenuation factor is:

[0045] When the gradient variance of edge e Less than or equal to the threshold When it is determined that the edge is less affected by environmental noise, it is not necessary to attenuate the edge weight. The attenuation factor is... The value is 1.

[0046] The formula for updating edge weights is:

[0047] in, This represents the edge weight after the k-th iteration. This represents the edge weight at the start of the (k+1)th iteration.

[0048] In this embodiment, the edge weight adaptive update mechanism can dynamically adjust the edge weights based on the gradient variance during feature propagation, and perform weight attenuation on edges severely affected by noise to suppress the propagation of noise in the graph structure. This mechanism enables the dynamic evolution of the spatiotemporally heterogeneous graph topology, allowing the graph structure to adapt to noise levels under different operating conditions and improving the system's noise resistance.

[0049] refer to Figure 3 In a preferred embodiment, when the feature aggregator uses the water-rock coupling evolution law as a bias term in the attention coefficient calculation, it derives the functional relationship between water content and matrix suction based on the unsaturated soil seepage mechanics equation and maps this functional relationship into a prior bias matrix in the attention calculation. When calculating the attention coefficient between the target node and its neighboring nodes in the spatiotemporal heterogeneous graph, the prior bias matrix and the original attention score calculated based on node features are subjected to a Hadamard product to reconstruct the feature transfer weight distribution of the neighboring nodes to the target node.

[0050] In this embodiment, the Van Genuchten model is used as the seepage mechanics equation for unsaturated soil. This model describes the relationship between water content and matrix suction in unsaturated soil. The expression for the Van Genuchten model is:

[0051] in, This refers to the volumetric moisture content. Residual moisture content saturated moisture content For matrix suction, , n, and m are the parameters of the VanGenuchten model, and m = 1 - 1 / n.

[0052] Matric suction refers to the attractive force exerted on water in unsaturated soil. The greater the matric suction, the higher the shear strength of the soil. When the water content increases, matric suction decreases, the shear strength of the soil decreases, and the slope is more prone to displacement. Therefore, the functional relationship between water content and matric suction reflects the fundamental law of water-rock coupling evolution.

[0053] The feature aggregator maps the VanGenuchten model to a prior bias matrix in the attention computation. First, the VanGenuchten model parameters are determined based on geological survey data of the slope area. , , Then, the matrix suction values ​​at different moisture contents are calculated, and a correspondence table between moisture content and matrix suction is established. Finally, this correspondence table is converted into a prior bias matrix B, where the elements of B are... This indicates the degree of influence of the matrix suction corresponding to the moisture content value of node i on the displacement node j.

[0054] Elements of the prior bias matrix B The calculation formula is:

[0055] in, Let i be the current moisture content of node i. The corresponding matrix suction value is calculated based on the VanGenuchten model. This represents the maximum matrix suction value in the slope area.

[0056] When calculating the attention coefficient between a target node and its neighboring nodes in a spatiotemporal heterogeneous graph, the feature aggregator first calculates the raw attention score based on the node features. The formula for calculating the raw attention score is:

[0057] in, Let i be the feature vector of the target node i. Let be the feature vector of neighbor node j, W be the learnable weight matrix, and a be the learnable attention vector. This represents a vector concatenation operation, and LeakyReLU is a linear rectified activation function with leakage.

[0058] Then, the feature aggregator combines the prior bias matrix B with the original attention scores. Perform the Hadamard product operation to obtain the corrected attention score. :

[0059] in, This represents the Hadamard product operation.

[0060] Finally, the feature aggregator normalizes the corrected attention score using a flexible maximum function to obtain the final attention coefficients. :

[0061] in, Let i be the set of neighboring nodes of the target node i.

[0062] In this embodiment, the feature aggregator maps the water-rock coupling evolution law into a prior bias matrix and performs a Hadamard product operation with the original attention scores, enabling the attention coefficients to reflect the actual physical coupling mechanism. This attention mechanism based on physical prior constraints avoids the possibility that a purely data-driven attention mechanism might learn feature transfer weights that do not conform to physical laws, thus improving the physical interpretability and reliability of the feature aggregation process.

[0063] In a preferred embodiment, a data-driven residual term is introduced to adaptively correct the prior bias matrix when calculating the attention coefficient. Anomaly coupling patterns deviating from the unsaturated soil seepage mechanics equations in multi-source time-series monitoring data are captured using a learnable parameterized matrix. The parameterized matrix is ​​added to the prior bias matrix to obtain a comprehensive bias matrix that integrates the physical mechanism and data residuals. This comprehensive bias matrix is ​​then used to replace the prior bias matrix in the Hadamard product operation to correct the feature transfer weight distribution under local anomalous conditions.

[0064] In this embodiment, the seepage mechanics equations for unsaturated soil are derived based on ideal geological conditions. In actual slope engineering, due to the complexity and uncertainty of geological conditions, abnormal coupling modes deviating from the ideal equations may exist. For example, in slope areas with fissures, the rainfall infiltration rate may be much greater than the rate predicted by the ideal equations, causing the relationship between water content and matrix suction to deviate from the Van Genuchten model. To compensate for the limitations of purely physical priors under abnormal conditions, a data-driven residual term is introduced to adaptively correct the prior bias matrix.

[0065] The data-driven residual term uses a learnable parameterized matrix R, whose dimension is the same as the prior bias matrix B. The elements of the parameterized matrix R... This indicates the degree to which the coupling relationship between the water content node i and the displacement node j deviates from the ideal physical equation. The parameterized matrix R is learned through the backpropagation algorithm during model training, which can automatically capture abnormal coupling patterns in multi-source time-series monitoring data.

[0066] Integrated bias matrix combining physical mechanisms and data residuals The calculation formula is:

[0067] Where B is the prior bias matrix and R is the data-driven residual matrix.

[0068] When calculating the attention coefficient, the comprehensive bias matrix is ​​used. The corrected attention score is obtained by replacing the prior bias matrix B in the Hadamard product operation. :

[0069] in, The original attention score is calculated based on node features. This represents the Hadamard product operation.

[0070] In this embodiment, by introducing a data-driven residual term, the system can adaptively correct the feature transfer weight distribution under local abnormal conditions while adhering to basic physical laws. This attention mechanism, which integrates physical priors and data-driven approaches, retains the interpretability and reliability of the physical model while possessing the flexibility and adaptability of a data-driven model, enabling it to better handle slope condition analysis problems under complex geological conditions.

[0071] Table 2 shows the comparison results of the prior bias matrix and the comprehensive bias matrix under different operating conditions.

[0072] Table 2 Comparison of prior bias matrix and combined bias matrix under different operating conditions

[0073] As shown in Table 2, under normal operating conditions, the residual values ​​are relatively small, and the overall bias value is close to the prior bias value. However, under fracture and karst conditions, the residual values ​​are relatively large, and the overall bias value differs significantly from the prior bias value. This indicates that data-driven residual terms can effectively capture coupling patterns under abnormal operating conditions and correct the bias of the prior bias matrix.

[0074] refer to Figure 4 In a preferred embodiment, when the feature extractor extracts dynamic evolution features through a temporal convolutional network, it employs a multi-layer dilated causal convolutional structure. Based on the slow-changing characteristics of displacement changes and the fast-changing characteristics of rainfall pulses in multi-source temporal monitoring data, differentiated dilation coefficients are configured for different types of heterogeneous nodes. A large-scale dilation coefficient is configured for displacement nodes to expand the temporal receptive field, while a small-scale dilation coefficient is configured for rainfall nodes to capture instantaneous pulse features. The output results of each layer of dilated causal convolution are then fused through residual connections.

[0075] In this embodiment, the temporal convolutional network comprises three dilated causal convolutional layers, each with a kernel size of 3. The output calculation formula for the dilated causal convolutional layer is as follows:

[0076] in, This is the output at time step t. The k-th element of the convolution kernel For the first The input at each time step is d, where d is the dilation coefficient and K is the kernel size.

[0077] The feature extractor configures differentiated dilation coefficients for different types of heterogeneous nodes. For displacement nodes, since displacement changes have a slow-cumulative characteristic, a larger temporal receptive field is needed to capture long-term displacement trends. Therefore, the dilation coefficient of the temporal convolutional network corresponding to the displacement node is set to [1,2,4], and the effective receptive fields of the three dilated causal convolutional layers are 3, 5, and 9 time steps, respectively. For rainfall nodes, since rainfall exhibits a fast-changing pulse characteristic, a smaller temporal receptive field is needed to capture instantaneous rainfall pulses. Therefore, the dilation coefficient of the temporal convolutional network corresponding to the rainfall node is set to [1,1,1], and the effective receptive field of each of the three dilated causal convolutional layers is 3 time steps. For water content nodes, since the time scale of water content changes is between that of displacement and rainfall, the dilation coefficient of the temporal convolutional network corresponding to the water content node is set to [1,2,2], and the effective receptive fields of the three dilated causal convolutional layers are 3, 5, and 5 time steps, respectively.

[0078] Table 3 shows the expansion coefficient configuration and effective receptive field of different types of heterogeneous nodes.

[0079] Table 3. Expansion coefficient configuration and effective receptive field of different types of heterogeneous nodes

[0080] The feature extractor fuses the outputs of each dilated causal convolutional layer through residual connections. Residual connections directly pass the input features to the output layer, adding them to the output features of the convolutional layers, which helps alleviate the vanishing gradient problem during deep neural network training. For each type of heterogeneous node, the output of the temporal convolutional network is the residual fusion result of the output features from three dilated causal convolutional layers.

[0081] In this embodiment, by configuring differentiated dilation coefficients for different types of heterogeneous nodes, the temporal convolutional network can simultaneously capture dynamic evolution features at different time scales. Large-scale dilation coefficients can capture the slow, cumulative trend of displacement, while small-scale dilation coefficients can capture the fast, pulse-like features of rainfall. This differentiated dilation coefficient configuration solves the problem of spatiotemporal scale mismatch in multi-source heterogeneous data, avoiding feature masking and interference between fast and slow-changing signals.

[0082] In a preferred embodiment, when the feature extractor fuses the outputs of each layer of dilated causal convolution, a cross-scale time-aligned fusion mechanism is employed. For the output sequences corresponding to different dilation coefficients, the time dimension is pruned and aligned based on the core scale benchmark. A channel attention mechanism is applied to the aligned multi-scale feature sequences, redistributing channel weights based on the response sensitivity of each channel feature to changes in slope state, highlighting feature channels with high response sensitivity, suppressing redundant channels, and concatenating the aligned and weighted multi-scale features into dynamically evolving features.

[0083] In this embodiment, a cross-scale temporal alignment fusion mechanism is used to address the issue of inconsistent output sequence lengths corresponding to different dilation coefficients. Due to the different dilation coefficients, the effective length of the output sequence of each dilated causal convolutional layer differs in the temporal dimension. For example, the first four time steps of the output sequence of a convolutional layer with a dilation coefficient of 4 are invalid because they rely on data that did not exist before the input sequence.

[0084] The cross-scale time alignment fusion mechanism first determines the core scale benchmark, which is the maximum value among all expansion coefficients. For displacement nodes, the core scale benchmark is 4; for water content nodes, the core scale benchmark is 2; and for rainfall nodes, the core scale benchmark is 1. Then, for the output sequences corresponding to different expansion coefficients, the time dimension is pruned according to the core scale benchmark, removing invalid time steps at the beginning of the output sequences to ensure that the effective length of all output sequences is consistent.

[0085] For example, for a temporal convolutional network corresponding to a displacement node, the input sequence length is T. The dilation coefficient of the first dilated causal convolutional layer is 1, and the effective length of the output sequence is T-2; the dilation coefficient of the second dilated causal convolutional layer is 2, and the effective length of the output sequence is T-4; the dilation coefficient of the third dilated causal convolutional layer is 4, and the effective length of the output sequence is T-8. The core scale baseline is 4. Therefore, the first 6 and first 4 time steps of the output sequences of the first and second layers are pruned respectively, so that the effective length of all output sequences is T-8.

[0086] A channel attention mechanism is applied to the aligned multi-scale feature sequence. This mechanism learns the importance weights of each channel's features and performs a weighted summation of features from different channels. The computation process is as follows: First, global average pooling is performed on the aligned multi-scale feature sequence to obtain channel-level feature vectors; then, these vectors are input into two fully connected layers to learn channel importance weights; finally, the learned channel importance weights are multiplied by the original multi-scale feature sequence to obtain a weighted multi-scale feature sequence.

[0087] The formula for calculating channel importance weights is:

[0088] Where z is the channel-level feature vector obtained by global average pooling. and The weight matrix is ​​a learnable matrix. It is the ReLU activation function. It is the Sigmoid activation function. This is the channel importance weight vector.

[0089] The feature extractor concatenates aligned and weighted multi-scale features into dynamically evolving features. The concatenation operation is performed along the channel dimension, combining feature sequences from different scales into a unified feature sequence, which serves as the dynamic evolutionary features of heterogeneous nodes.

[0090] In this embodiment, the cross-scale time alignment fusion mechanism solves the problem of inconsistent output sequence lengths corresponding to different expansion coefficients, ensuring that multi-scale features can be accurately aligned in the time dimension. The channel attention mechanism can automatically learn the importance of each channel feature, highlighting feature channels that are highly sensitive to changes in slope state, suppressing redundant channels, and improving the expressive power of dynamic evolution features.

[0091] refer to Figure 5 In a preferred embodiment, before inputting the graph-level representation vector into the fully connected layer, the classifier performs a global pooling operation by introducing heterogeneous node importance weights to aggregate node-level features into a graph-level representation vector. The importance weights are jointly determined based on the mean of the attention coefficients output by the graph attention network and the degree centrality index of the heterogeneous nodes. The fully connected layer uses a flexible maximum activation function to map the graph-level representation vector to the state space of slope stability, slope creep, and slope instability, outputting a slope state classification probability distribution.

[0092] In this embodiment, the global pooling operation aggregates the features of all nodes in the spatiotemporally heterogeneous graph into a unified graph-level representation vector. Traditional global average pooling assigns the same weight to the features of all nodes, ignoring the differences in importance between different types of nodes during slope state evolution. To address this issue, a weighted sum of the features of different node types is introduced, based on the heterogeneous node importance.

[0093] The importance weights of heterogeneous nodes are jointly determined based on the mean attention coefficients output by the graph attention network and the degree centrality index of the heterogeneous nodes. The mean attention coefficient reflects the importance of a node in the feature aggregation process; the larger the mean attention coefficient, the greater the influence of the node on other nodes. The degree centrality index reflects the degree of connectivity of a node in the graph structure; the larger the degree centrality, the more connections the node has with other nodes, and the more important its position in the graph structure.

[0094] For node i, its importance weight The calculation formula is:

[0095] in, Let be the mean attention coefficient of node i, which is the average of the attention coefficients of all nodes with node i as a neighbor. Let be the degree centrality index of node i, which is the sum of the in-degree and out-degree of node i; This is a balancing parameter used to adjust the relative importance of the mean attention coefficient and the degree centrality index, and its value ranges from [0,1].

[0096] The formula for calculating global pooling operations is:

[0097] in, Let V be the graph-level representation vector, and V be the set of nodes. Let i be the dynamic evolution feature vector of node i.

[0098] The classifier represents the vector at the graph level. The input is a fully connected layer, which contains two hidden layers and one output layer. The first hidden layer has 128 neurons and uses the ReLU activation function; the second hidden layer has 64 neurons and uses the ReLU activation function; the output layer has 3 neurons, corresponding to the three state categories of slope stability, slope creep, and slope instability, and uses the flexible maximum activation function.

[0099] The formula for calculating the flexible maximum activation function is:

[0100] in, Let be the classification probability of the k-th state category. This is the output value of the k-th neuron in the output layer.

[0101] The slope condition classification probability distribution output by the classifier is as follows: ,in Let be the probability of slope stability. Let be the probability of slope creep. This represents the probability of slope instability.

[0102] In this embodiment, by introducing heterogeneous node importance weights to perform global pooling, the differences in importance of different types of nodes during the slope state evolution process can be fully considered, enabling the graph-level representation vector to more accurately reflect the overall state of the slope. The fully connected layer uses a flexible maximum activation function, which can output the probability distribution of each state category, providing richer information for subsequent early warning decisions.

[0103] In a preferred embodiment, after acquiring multi-source time-series monitoring data through the data interface, a graph-based data loss compensation mechanism is introduced. When a data loss is detected in a target sensor node within a target time window, the set of neighboring nodes of the target sensor node in the spatiotemporal heterogeneous graph is located. The time-series features of the neighboring node set within the target time window are extracted, and combined with the edge types between the target sensor node and each neighboring node, the missing data of the target sensor node is reconstructed through a bidirectional long short-term memory network.

[0104] In this embodiment, a data loss compensation mechanism is used to address the problem of missing multi-source time-series monitoring data caused by sensor failure or communication interruption. After acquiring multi-source time-series monitoring data, the data interface first checks the data integrity of each sensor node within each time window. When the target sensor node i is detected within the target time window... When data is missing in memory, a data loss compensation mechanism is triggered.

[0105] The execution process of the data missing compensation mechanism is as follows: First, locate the set of neighboring nodes of the target sensor node i in the spatiotemporal heterogeneous graph. Neighbor node set Includes all nodes connected to the target node i by an edge. For a rainfall node, the set of neighboring nodes is all water content nodes connected to it by an infiltration edge; for a water content node, the set of neighboring nodes is all rainfall nodes connected to it by an infiltration edge and all displacement nodes connected to it by a softening edge; for a displacement node, the set of neighboring nodes is all water content nodes connected to it by a softening edge.

[0106] Then, extract the set of neighboring nodes. Within the target time window The time series features within the target time window are extracted. These features include statistical characteristics such as mean, variance, maximum, minimum, slope, and autocorrelation coefficient. Simultaneously, historical time series features of the target sensor node i prior to the target time window are extracted as a reference for reconstructing missing data.

[0107] Next, an input feature vector is constructed by combining the edge types between the target sensor node i and its neighboring nodes. Different feature weights are assigned to different edge types. The feature weight for infiltration edges is set to 0.6, and the feature weight for softening edges is set to 0.4. The input feature vector includes the time-series features of neighboring nodes, the historical time-series features of the target node, and the edge type weights.

[0108] Finally, the input feature vector is fed into a bidirectional long short-term memory network to reconstruct the target sensor node i within the target time window. The bidirectional long short-term memory (LSTM) network contains forward and backward LSM layers, enabling it to simultaneously capture both forward and backward dependencies in the time series. The output of the bidirectional LSM network is reconstructed time series data, filling in the missing data for the target sensor nodes.

[0109] In this embodiment, the graph-based data missing compensation mechanism utilizes the physical relationships between nodes in a spatiotemporally heterogeneous graph to reconstruct the missing data of the target node through the time-series features of neighboring nodes. This mechanism fully considers the coupling relationships between different physical parameters, enabling more accurate reconstruction of missing data and ensuring the integrity of multi-source time-series monitoring data.

[0110] In a preferred embodiment, after the classifier outputs the slope state classification probability distribution, an early warning decision mechanism based on cognitive uncertainty is introduced. The Monte Carlo random deactivation method is used to perform multiple forward propagation inferences on the feature extractor and classifier, obtaining the state class variance of the multiple inference results. If the state class variance is greater than the cognitive uncertainty threshold, a sensor verification command is triggered to reacquire multi-source time-series monitoring data; if the state class variance is less than the cognitive uncertainty threshold, an early warning decision is executed based on the maximum value in the slope state classification probability distribution.

[0111] In this embodiment, cognitive uncertainty refers to the uncertainty of prediction results caused by the uncertainty of model parameters. When the model lacks sufficient understanding of the input data, the variance of the prediction results will be large. The early warning decision-making mechanism based on cognitive uncertainty improves the reliability of early warning decisions by quantifying the cognitive uncertainty of the model.

[0112] Monte Carlo random deactivation is an effective method for quantifying cognitive uncertainty in neural networks. This method keeps the random deactivation layer on during the inference phase, obtains multiple predictions through multiple forward propagation inferences, and calculates the variance of these predictions as a measure of cognitive uncertainty.

[0113] The execution process of the early warning decision-making mechanism based on cognitive uncertainty is as follows: First, a random deactivation layer is inserted between the fully connected layers of the feature extractor and the classifier, with a random deactivation rate set to 0.2. Then, during the inference phase, the random deactivation layer remains on, and M forward propagation inferences are performed on the same input data to obtain M slope state classification probability distributions. , where M is the number of inferences, and its value is 50.

[0114] Next, the variance of each state category is calculated. For the k-th state category, its variance is... The calculation formula is:

[0115] in, Let be the probability of the k-th state category obtained from the m-th inference. Let be the mean probability of the k-th state category obtained from M inferences.

[0116] Then, calculate the maximum value of the state category variance. Preset cognitive uncertainty threshold ,when Greater than the threshold When the model's understanding of the current input data is insufficient, and the reliability of the prediction results is low, a sensor verification command is triggered to reacquire multi-source time-series monitoring data. Less than or equal to the threshold When the model has a sufficient understanding of the current input data and the prediction results are highly reliable, it makes an early warning decision based on the maximum value in the probability distribution of slope state classification.

[0117] The rules for implementing early warning decisions are as follows: if the state category corresponding to the maximum value in the slope state classification probability distribution is slope stability, no early warning will be issued; if the state category corresponding to the maximum value is slope creep, a yellow early warning will be issued; if the state category corresponding to the maximum value is slope instability, a red early warning will be issued.

[0118] Table 4 shows the early warning decision-making results under different levels of cognitive uncertainty.

[0119] Table 4. Early warning decision-making results under different levels of cognitive uncertainty

[0120] In this embodiment, the early warning decision-making mechanism based on cognitive uncertainty quantifies the cognitive uncertainty of the model using the Monte Carlo random deactivation method. When the uncertainty is high, it triggers a sensor verification command, avoiding false alarms caused by insufficient model cognition. This mechanism improves the reliability of the system's decision-making under complex operating conditions and reduces the occurrence of false alarms and missed alarms.

Claims

1. A slope condition intelligent analysis system based on multi-source monitoring data fusion, characterized in that, include: The data interface acquires multi-source time-series monitoring data including displacement, rainfall, and soil moisture content; The graph builder constructs a spatiotemporal heterogeneous graph based on the multi-source time-series monitoring data, defines different types of sensors as heterogeneous nodes, the attributes of the heterogeneous nodes are the corresponding time-series features, and the edge types between heterogeneous nodes reflect the physical correlation mechanism of water-rock coupling. The feature aggregator employs a graph attention network based on physical prior constraints, using the water-rock coupling evolution law as a bias term in the attention coefficient calculation to aggregate the features of heterogeneous nodes in the spatiotemporal heterogeneous graph. The feature extractor extracts the dynamic evolution features of the heterogeneous nodes under physical constraints through a temporal convolutional network; The classifier takes the aggregated graph-level representation vectors and inputs them into the fully connected layer, outputting the slope state classification results. When the graph builder constructs edge types that reflect the physical correlation mechanism of water-rock coupling in the spatiotemporal heterogeneous graph, it constructs directed heterogeneous edges between heterogeneous nodes based on the time lag effect of rainfall infiltration and displacement response in geomechanics and the reduction law of soil strength due to water content changes. The directed heterogeneous edge includes an infiltration edge from the rainfall node to the moisture content node, and a softening edge from the moisture content node to the displacement node. The graph builder dynamically configures the initial values ​​of the edge weights of the directed heterogeneous edges based on the lag time of the extreme points of mutual information between different physical parameters in the multi-source time-series monitoring data. When the feature aggregator uses the water-rock coupling evolution law as the bias term for the attention coefficient calculation, it derives the functional relationship between water content and matrix suction based on the unsaturated soil seepage mechanics equation, and maps the functional relationship into the prior bias matrix in the attention calculation. Based on the geological survey data of the slope area, the parameters of the VanGenuchten model were determined. Then, the matrix suction values ​​at different moisture contents were calculated, and a correspondence table between moisture content and matrix suction was established. This correspondence table was then converted into a priori bias matrix B, where the elements of B... This indicates the degree of influence of the matrix suction corresponding to the moisture content value of node i on the displacement node j; When calculating the attention coefficients between the target node and its neighboring nodes in the spatiotemporal heterogeneous graph, the prior bias matrix and the original attention scores calculated based on node features are subjected to a Hadamard product operation to reconstruct the feature transfer weight distribution of the neighboring nodes to the target node.

2. The intelligent slope condition analysis system based on multi-source monitoring data fusion according to claim 1, characterized in that, When the feature extractor extracts dynamic evolution features through a temporal convolutional network, it adopts a multi-layer dilated causal convolutional structure. Based on the slow-changing characteristics of displacement changes and the fast-changing characteristics of rainfall pulses in the multi-source temporal monitoring data, it configures differentiated dilation coefficients for different types of heterogeneous nodes. A large-scale dilation coefficient is configured for displacement nodes to expand the temporal receptive field, and a small-scale dilation coefficient is configured for rainfall nodes to capture instantaneous pulse features. The output results of each layer of dilated causal convolution are then fused through residual connections.

3. The intelligent slope condition analysis system based on multi-source monitoring data fusion according to claim 1, characterized in that, Before the classifier inputs the graph-level representation vector into the fully connected layer, it performs a global pooling operation by introducing heterogeneous node importance weights to aggregate node-level features into a graph-level representation vector. The importance weights are determined jointly based on the mean of the attention coefficients output by the graph attention network and the degree centrality index of the heterogeneous nodes. The fully connected layer uses a flexible maximum activation function to map the graph-level representation vector to the state space of slope stability, slope creep, and slope instability, and outputs the slope state classification probability distribution.

4. The intelligent slope condition analysis system based on multi-source monitoring data fusion according to claim 1, characterized in that, After the graph builder dynamically configures the initial values ​​of the edge weights of the directed heterogeneous edges, an adaptive edge weight update mechanism is introduced during the feature aggregation iteration process of the graph attention network. Based on the gradient variance of the neighbor node features in the current iteration round, the edge weights of the directed heterogeneous edges are dynamically reduced or enhanced. When the gradient variance exceeds a preset threshold, it is determined that the corresponding directed heterogeneous edge is affected by environmental noise. An attenuation factor is applied to the edge weight to suppress noise transmission and realize the dynamic evolution of the topology.

5. The intelligent slope condition analysis system based on multi-source monitoring data fusion according to claim 1, characterized in that, When calculating the attention coefficient, a data-driven residual term is introduced to adaptively correct the prior bias matrix. Anomalous coupling patterns deviating from the seepage mechanics equations of unsaturated soil in the multi-source time-series monitoring data are captured by a learnable parameterized matrix. The parameterized matrix is ​​added to the prior bias matrix to obtain a comprehensive bias matrix that integrates the physical mechanism and the data residual. The comprehensive bias matrix is ​​used to replace the prior bias matrix in the Hadamard product operation to correct the feature transfer weight distribution under local abnormal conditions.

6. The intelligent slope condition analysis system based on multi-source monitoring data fusion according to claim 2, characterized in that, When the feature extractor fuses the output results of each layer of dilated causal convolution, it adopts a cross-scale time alignment fusion mechanism. For the output sequences corresponding to different expansion coefficients, the time dimension is pruned and aligned based on the core scale benchmark; A channel attention mechanism is applied to the aligned multi-scale feature sequence. Based on the response sensitivity of each channel feature to changes in slope state, the channel weights are redistributed to highlight the feature channels with high response sensitivity and suppress redundant channels. The aligned and weighted multi-scale features are then spliced ​​into dynamic evolution features.

7. The intelligent slope condition analysis system based on multi-source monitoring data fusion according to claim 1, characterized in that, After the data interface acquires the multi-source time-series monitoring data, it introduces a data missing compensation mechanism based on a graph structure. When a data gap is detected in the target sensor node within the target time window, the set of neighboring nodes of the target sensor node in the spatiotemporal heterogeneous graph is located. Extract the time series features of the neighbor node set within the target time window, and combine them with the edge types between the target sensor node and each neighbor node to reconstruct the missing data of the target sensor node through a bidirectional long short-term memory network.

8. The intelligent slope condition analysis system based on multi-source monitoring data fusion according to claim 1, characterized in that, After the classifier outputs the slope state classification probability distribution, an early warning decision-making mechanism based on cognitive uncertainty is introduced. The Monte Carlo random deactivation method is used to perform multiple forward propagation inferences on the feature extractor and the classifier to obtain the state class variance of the multiple inference results. If the variance of the state category is greater than the cognitive uncertainty threshold, a sensor verification instruction is triggered to reacquire multi-source time-series monitoring data; If the variance of the state category is less than the cognitive uncertainty threshold, then an early warning decision is made based on the maximum value in the slope state classification probability distribution.

Images