A Transformer-based Slope Knowledge Graph Risk Early Warning Method

By using a Transformer-based slope knowledge graph method, which combines multi-source time-series data with a slope risk knowledge graph, the problems of high false alarm rate from a single data source and lack of prior geological knowledge in slope risk early warning are solved, enabling timely identification and accurate early warning of progressive risks coupled with multiple factors.

CN122925922BActive Publication Date: 2026-11-17UNIV OF ELECTRONICS SCI & TECH OF CHINA
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611425736.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-15
Publication Date
2026-11-17
Estimated Expiration
2046-09-15

AI Technical Summary

Technical Problem

Existing technologies for slope risk early warning suffer from high false alarm rates due to single data sources and lack explicit integration of prior geological knowledge, making it difficult to identify hidden risks in a timely manner from the progressive evolution of multiple coupled factors.

Method used

A Transformer-based slope knowledge graph method is adopted. By collecting and preprocessing multi-source time-series data and combining it with the slope risk knowledge graph, the spatiotemporal characteristics of multi-source physical fields, knowledge evolution characteristics, and knowledge-guided spatiotemporal joint characteristics are extracted. Long-range dependency coding, progressive risk state evolution learning, and deep risk state distillation are then carried out to finally assess the risk state and determine the level.

Benefits of technology

It effectively reduces the false alarm rate caused by fluctuations in a single data source, improves the timeliness of identification and the accuracy of early warning for multi-factor coupled progressive risks, and can capture the continuous development of internal damage when there are no obvious anomalies in surface monitoring data, taking into account both the overall average level and the response sensitivity to rapid deterioration before disaster.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122925922B_ABST
    Figure CN122925922B_ABST
Patent Text Reader

Abstract

This invention provides a slope risk early warning method based on a Transformer-based knowledge graph, belonging to the field of slope stability monitoring and early warning technology. By introducing a slope risk knowledge graph, this invention explicitly integrates structured knowledge such as geological conditions, historical disaster cases, and the causal logic between influencing factors into the early warning process. This overcomes the shortcomings of purely data-driven methods, such as shallow knowledge intervention and limited versatility in multi-factor coupled scenarios. It effectively reduces the false alarm rate caused by fluctuations in a single data source and improves the timeliness and accuracy of identifying progressive risks coupled with multiple factors. Furthermore, this invention utilizes deep internal states to adaptively calibrate surface features, ensuring that even if surface monitoring data does not show obvious anomalies, the continuous development of internal damage can be reflected in the feature expression, enabling the early warning results to capture early signals of risk evolution even when there are no significant changes in monitoring data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of slope stability monitoring and early warning technology, specifically involving a slope knowledge graph risk early warning method based on Transformer. Background Technology

[0002] Slope risk early warning refers to the process of acquiring slope status information using various sensors and monitoring methods, and making advance judgments on potential instability and failure. It is a crucial link in geological disaster prevention and control. Compared with general structures, slopes are located in complex geological environments. Their deformation and failure are affected by multiple factors, including stratum lithology, geological structure, hydrological conditions, and engineering disturbances. Moreover, the failure process often exhibits both gradual accumulation and sudden release characteristics, making it difficult to directly identify hidden hazards from a single signal before a disaster occurs. For example, after continuous rainfall, the surface displacement rate of a high slope may still fluctuate within the safe threshold range, but cracks in the internal rock mass structure are already slowly expanding. This phenomenon of relatively stable surface data while internal damage continues to accumulate is not uncommon in engineering sites, placing higher demands on the timeliness and accuracy of early warning.

[0003] In engineering practice, a common early warning method is based on threshold judgment of key indicators. This involves setting multiple warning values ​​for monitoring points, such as displacement rate, crack propagation, or daily rainfall. An alarm is triggered when a certain indicator exceeds a preset limit. This method is simple to deploy and responds directly; however, due to the coupled nature of slope systems, single-point or single-factor anomalies are often caused by non-hazardous factors such as environmental fluctuations and sensor disturbances, leading to frequent false alarms. Furthermore, the threshold method mainly reflects the current data exceeding the limit and lacks the ability to characterize the dynamic interaction between displacement, seepage, and stress fields. It is difficult to reveal the gradual risks driven by the slow accumulation of multiple adverse factors, and it cannot capture hidden precursors of disasters before apparent anomalies appear, resulting in insufficient lead time for early warning decisions and a short window for on-site response.

[0004] With the development of sensing and data acquisition technologies, deep learning-based early warning methods have been gradually introduced to automatically learn deep features from large amounts of monitoring data. In a typical process, multi-source time-series data such as displacement, rainfall, and groundwater levels are first normalized, then input into a Long Short-Term Memory (LSTM) network to extract temporal dependencies, or a three-layer convolutional kernel is used to extract spatial features of the sensor array, and a fully connected layer outputs the risk probability. Compared to thresholding methods, these methods offer improved tolerance to data fluctuations and pattern recognition accuracy. However, existing deep learning solutions primarily focus on the direct driving force of monitoring time-series data, essentially replacing physical mechanisms with statistical data patterns, resulting in a shallow level of knowledge intervention. Due to the lack of explicit fusion of structured knowledge such as slope geological conditions, historical disaster cases, potential slip surface locations, and causal logic among influencing factors, the model's versatility is limited when facing multi-factor coupled scenarios outside the training distribution. Especially in the stage where there are no obvious changes in monitoring data but internal damage has been developing, purely data-driven methods are less able to infer the development trend of hidden risks from limited surface information, and the timeliness of early warning and the interpretability of causality still have certain limitations. Summary of the Invention

[0005] To address the problems existing in the background technology, the purpose of this invention is to provide a slope knowledge graph-based risk early warning method based on Transformer, so as to solve the problems of high false alarm rate of single data source in traditional methods, lack of explicit fusion of geological prior knowledge in pure data-driven methods, and difficulty in timely identification of hidden risks from the progressive evolution of multi-factor coupling.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A slope knowledge graph-based risk early warning method based on Transformer includes the following steps:

[0008] Step 1: Obtain a slope risk knowledge graph;

[0009] Time series data of surface displacement, deep displacement, groundwater level, and daily rainfall were collected from the slope and preprocessed to obtain preprocessed time series data of surface displacement, deep displacement, groundwater level, and daily rainfall.

[0010] Step 2: Based on the preprocessed surface displacement time series data, preprocessed deep displacement time series data, preprocessed groundwater level time series data, preprocessed daily rainfall time series data, and slope risk knowledge graph, extract the spatiotemporal features of multi-source physical fields, knowledge evolution features, and knowledge-guided spatiotemporal joint features in sequence.

[0011] Step 3: Based on the knowledge-guided spatiotemporal joint characteristics, knowledge evolution characteristics, and slope risk knowledge graph, perform long-range dependency encoding, progressive risk state evolution learning, and deep risk state distillation in sequence to obtain knowledge-enhanced long-range dependency characteristics, risk state evolution characteristics, and deep risk states.

[0012] Step 4: Based on the knowledge-guided spatiotemporal joint features and deep risk status, perform multi-granularity risk feature extraction and fusion, nonlinear risk status update modeling, global risk correlation modeling, and comprehensive risk quantification in sequence to obtain multi-granularity calibration features, risk evolution features, attention-enhanced risk features, and comprehensive risk index.

[0013] Step 5: Based on the comprehensive risk index, conduct risk status assessment and risk level determination in sequence to obtain a quantitative risk score vector and predict the risk level.

[0014] Furthermore, surface displacement time-series data were acquired using a GNSS receiver, deep displacement time-series data were acquired using an inclinometer, groundwater level time-series data were acquired using a water level gauge, and daily rainfall time-series data for the slope area were acquired using a rain gauge; all four types of data are time-series numerical data.

[0015] Furthermore, the data type of the slope risk knowledge graph is structured knowledge graph data; the entity types in the slope risk knowledge graph include displacement monitoring points, slope structure type, slope ratio, dip angle, slope height, soil type, groundwater depth, and rainfall events; the relationship types in the slope risk knowledge graph include monitoring association, attribute affiliation, spatial bearing capacity, hydrological constraints, time-series induction, deformation response, parameter constraints, and coupling effects.

[0016] Furthermore, the specific process of time-series numerical data preprocessing is as follows:

[0017] First, the collected time-series numerical data is cleaned: outliers are removed using the three-standard-deviation method, missing values ​​are filled in using linear interpolation, and finally Z-score standardization is used to obtain the cleaned time-series numerical data.

[0018] The cleaned time-series numerical data is timestamped and the sampling frequency is standardized to obtain preprocessed time-series numerical data.

[0019] Furthermore, the specific process of step 2 is as follows:

[0020] Step 2.1: Based on the preprocessed surface displacement time series data, preprocessed deep displacement time series data, preprocessed groundwater level time series data, and preprocessed daily rainfall time series data, parallel spatiotemporal feature extraction is performed to obtain the spatiotemporal features of the multi-source physical field. The calculation method is as follows:

[0021] ,

[0022] ,

[0023] ,

[0024] ,

[0025] in, For multi-source monitoring time series splicing tensor, This is a channel-level splicing operation. This is the preprocessed surface displacement time series data. This is the preprocessed deep displacement time series data. This is the preprocessed groundwater level time series data. This is the preprocessed time-series data of daily rainfall. It is a time-dependent feature. It is a bidirectional gated loop unit. It is a one-dimensional convolutional network. It is a multi-scale physical field characteristic. This is a one-dimensional dilated convolution operation. For the Softmax function, For feature dimension, The spatiotemporal characteristics of multi-source physical fields For layer normalization operation, For transpose operation, For element-wise addition, For Hadama accumulation, Use the Sigmoid activation function;

[0026] Step 2.2: Based on the slope risk knowledge graph, semantic knowledge features are extracted to obtain knowledge evolution features. The calculation method is as follows:

[0027] ,

[0028] ,

[0029] ,

[0030] ,

[0031] in, For a symmetric normalized adjacency matrix, The square root of the inverse of the degree matrix. Given an adjacency matrix with self-loops, This is the physical distance attenuation weight matrix. It is an exponential function. This function calculates the physical distance between entities. For entities without spatial coordinates, the distance between them returns a default value of 0. It only calculates the actual physical distance for entities with coordinates. This is a set of entity nodes in the slope risk knowledge graph, including displacement monitoring points, slope structure type, slope ratio, dip angle, slope height, soil type, groundwater depth, and rainfall event entities. To integrate the characteristics of risk propagation path sensing nodes constrained by physical distance, For graph attention network layers, For embedding layer, As a characteristic of knowledge evolution, For Transformer encoders, This is the distance attenuation coefficient;

[0032] Step 2.3: Based on the spatiotemporal characteristics of multi-source physical fields and knowledge evolution characteristics, feature fusion is performed to obtain knowledge-guided spatiotemporal joint features. The calculation method is as follows:

[0033] ,

[0034] ,

[0035] ,

[0036] in, Cross-attention features guided by knowledge. This is a multi-head cross-attention mechanism. The weights are for the first adaptive mode gating. It is a multilayer perceptron. Knowledge guides the spatiotemporal joint characteristics.

[0037] Furthermore, in the entity node set of the slope risk knowledge graph, entities with spatial coordinates are displacement monitoring points.

[0038] Furthermore, the specific process of step 3 is as follows:

[0039] Step 3.1: Based on the knowledge-guided spatiotemporal joint features, knowledge evolution features, and slope risk knowledge graph, perform long-range dependency encoding to obtain knowledge-enhanced long-range dependency features. The calculation method is as follows:

[0040] ,

[0041] ,

[0042] ,

[0043] ,

[0044] in, As a feature of knowledge structure, This is a set of relationship types in the slope risk knowledge graph. Let r be the symmetric normalized adjacency matrix corresponding to relation type r, where r is the relation type index. This is a relation type weight matrix. This is a self-loop weight matrix. For attention knowledge pooling vectors, For linear layers, Enhance attentional features for knowledge. To enhance knowledge with long-range dependency features, For the replication function along the time dimension, Copy along the timeline to match the time sequence length; ⊙ represents the Hadamard product, and ⊕ represents element-wise addition.

[0045] Step 3.2: Based on the long-range dependency characteristics and knowledge evolution characteristics of knowledge reinforcement, perform progressive risk state evolution learning to obtain the risk state evolution characteristics. The calculation method is as follows:

[0046] ,

[0047] ,

[0048] ,

[0049] ,

[0050] in, The knowledge propagation feature sequence encoded by the graph attention network. Aggregate vectors for attention-based knowledge risks. For the second adaptive modal gating weights, It is the hyperbolic tangent function. Characteristics of risk state evolution;

[0051] Step 3.3: Based on the risk state evolution characteristics and knowledge evolution characteristics, perform deep risk state distillation to obtain the deep risk state. The calculation method is as follows:

[0052] ,

[0053] ,

[0054] ,

[0055] ,

[0056] ,

[0057] in, To illustrate the characteristics of attention propagation, This is a feature of multi-hop knowledge graph reasoning. For multi-hop feature fusion projection matrix, For cross-modal attention features, For the third adaptive modal gating weights, This indicates a state of deep risk.

[0058] Furthermore, the specific process of step 4 is as follows:

[0059] Step 4.1: Based on the knowledge-guided spatiotemporal joint features and deep risk state sequences, multi-granularity risk features are extracted and fused through multi-scale dilated temporal convolution and cross-source dimensional adaptive calibration gating mechanisms to obtain multi-granularity calibration features. The calculation method is as follows:

[0060] ,

[0061] ,

[0062] ,

[0063] ,

[0064] in, This represents the initial fusion characteristics across multiple scales. To use a one-dimensional convolutional layer with a first convolutional kernel, To employ a one-dimensional dilated convolutional layer with a second convolutional kernel, To employ a one-dimensional dilated convolutional layer with a third convolutional kernel, For multi-scale deep fusion features, To adaptively calibrate the gate weights, This is a multi-granularity calibration feature;

[0065] Step 4.2: Based on the multi-granularity calibration characteristics, perform nonlinear risk state update modeling to obtain risk evolution characteristics. The calculation method is as follows:

[0066] ,

[0067] ,

[0068] ,

[0069] in, It features bidirectional gating fusion. This represents the risk state offset. Characteristics of risk evolution;

[0070] Step 4.3: Based on the risk evolution characteristics, a global risk association model is performed using a temporal multi-head self-attention enhancement and adaptive confidence weighted fusion mechanism to obtain attention-enhanced risk features. The calculation method is as follows:

[0071] ,

[0072] ,

[0073] ,

[0074] in, As a characteristic of attention risk, To enhance attention to risk characteristics, For element-wise confidence weights, It is a soft positive activation function;

[0075] Step 4.4: Based on the risk characteristics of attention enhancement, a comprehensive risk quantification is performed through temporal global pooling and composite risk regression mapping to obtain a comprehensive risk index. The calculation method is as follows:

[0076] ,

[0077] ,

[0078] in, This is a one-dimensional global average pooling operation. The original risk regression value, To comprehensively assess risk characteristics, This is a numerical clipping operation. This is the risk composite adjustment coefficient. These are the lower and upper bounds for cropping, respectively.

[0079] Furthermore, the specific process of step 5 is as follows:

[0080] Step 5.1: Conduct a risk status assessment based on comprehensive risk characteristics to obtain a multi-level risk response vector. The calculation method is as follows:

[0081] ,

[0082] ,

[0083] in, For risk response gating vector, In response to the weight matrix, In response to the bias vector, This is a quantitative risk scoring vector;

[0084] Step 5.2: Based on the quantified risk scoring vector, the risk level is determined through threshold comparison and probabilistic mapping to obtain the predicted risk level. The calculation method is as follows:

[0085] ,

[0086] ,

[0087] in, For the threshold response vector, This is the temperature scaling factor. For the level mapping weight matrix, For the level mapping bias vector, To predict the risk level, This is the index for retrieving the maximum value.

[0088] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0089] This invention introduces a slope risk knowledge graph, explicitly integrating structured knowledge such as geological conditions, historical disaster cases, and causal logic among influencing factors into the early warning process. This overcomes the shortcomings of purely data-driven methods, such as shallow knowledge intervention and limited versatility in multi-factor coupled scenarios. It effectively reduces the false alarm rate caused by fluctuations in a single data source and improves the timeliness of identification and accuracy of early warning for progressive risks coupled with multiple factors. In addition, this invention uses deep internal states to adaptively calibrate surface features, so that even if there are no obvious anomalies in surface monitoring data, the continuous development of internal damage can be reflected in the feature expression. Finally, the risk quantification takes into account both the overall average level and the sensitivity to rapid deterioration before disaster, so that the early warning results can still capture early signals of risk evolution even when there are no obvious changes in monitoring data. Attached Figure Description

[0090] Figure 1 The visualized slope risk knowledge graph provided by this invention;

[0091] Figure 2 The image shows the results of slope risk early warning and monitoring data provided by this invention. Detailed Implementation

[0092] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0093] Example 1

[0094] A slope knowledge graph-based risk early warning method based on Transformer includes the following steps:

[0095] Step 1: Collect slope risk knowledge graph; collect time-series data on slope surface displacement, deep displacement, groundwater level, and daily rainfall, and preprocess them respectively to obtain preprocessed time-series data on surface displacement, deep displacement, groundwater level, and daily rainfall; the specific process is as follows:

[0096] Step 1.1: Collect surface displacement time series data of the slope using a GNSS receiver. The data type is time series numerical data, including the horizontal and vertical displacements of several monitoring points on the slope. Use the three-standard-deviation method to remove outliers from the surface displacement time series data, then use linear interpolation to fill in missing values, and finally use Z-score standardization to obtain the cleaned surface displacement time series data.

[0097] Step 1.2: Collect deep displacement time series data of the slope using an inclinometer. The data type is time series numerical data, including the horizontal displacement of measuring points at different depths. Use the three-standard-deviation method to remove outliers from the deep displacement time series data, then use linear interpolation to fill in missing values, and finally use Z-score standardization to obtain the cleaned deep displacement time series data.

[0098] Step 1.3: Collect time-series groundwater level data of the slope using a water level gauge and daily rainfall time-series data of the area where the slope is located using a rain gauge. Both data types are time-series numerical data. Outliers in the groundwater level time-series data and daily rainfall time-series data are removed using the three-standard-deviation method. Missing values ​​in the groundwater level time-series data are filled using linear interpolation. Abnormal single-day extreme values ​​in the daily rainfall time-series data are replaced with the mean of the nearest date. Then, both are standardized using Z-score to obtain cleaned groundwater level time-series data and cleaned daily rainfall time-series data.

[0099] Step 1.4: Timestamp alignment and sampling frequency unification are performed on the cleaned surface displacement time series data, cleaned deep displacement time series data, cleaned groundwater level time series data, and cleaned daily rainfall time series data to obtain preprocessed surface displacement time series data, preprocessed deep displacement time series data, preprocessed groundwater level time series data, and preprocessed daily rainfall time series data.

[0100] Step 1.5: Collect slope risk knowledge graphs, with the data type being structured knowledge graph data; the entity types in the slope risk knowledge graph include displacement monitoring points, slope structure type, slope ratio, dip angle, slope height, soil type, groundwater depth, and rainfall events; the relationship types in the slope risk knowledge graph include monitoring association, attribute affiliation, spatial bearing capacity, hydrological constraints, time-series induction, deformation response, parameter constraints, and coupling effects.

[0101] Step 2: Based on the preprocessed surface displacement time series data, preprocessed deep displacement time series data, preprocessed groundwater level time series data, preprocessed daily rainfall time series data, and slope risk knowledge graph, extract the multi-source physical field spatiotemporal features, knowledge evolution features, and knowledge-guided spatiotemporal joint features sequentially; the specific process is as follows:

[0102] Step 2.1: Based on the preprocessed surface displacement time series data, preprocessed deep displacement time series data, preprocessed groundwater level time series data, and preprocessed daily rainfall time series data, parallel spatiotemporal feature extraction is performed to obtain the spatiotemporal features of the multi-source physical field. The calculation method is as follows:

[0103]

[0104]

[0105]

[0106]

[0107] in, For multi-source monitoring time series splicing tensor, This is a channel-level splicing operation. This is the preprocessed surface displacement time series data. This is the preprocessed deep displacement time series data. This is the preprocessed groundwater level time series data. This is the preprocessed time-series data of daily rainfall. It is a time-dependent feature. It is a bidirectional gated loop unit. It is a one-dimensional convolutional network. It is a multi-scale physical field characteristic. This is a one-dimensional dilated convolution operation. For the Softmax function, For feature dimension, The spatiotemporal characteristics of multi-source physical fields For layer normalization operation, For transpose operation, For element-wise addition, For Hadama accumulation, Use the Sigmoid activation function;

[0108] Step 2.2: Based on the slope risk knowledge graph, semantic knowledge features are extracted to obtain knowledge evolution features. The calculation method is as follows:

[0109]

[0110]

[0111]

[0112]

[0113] in, For a symmetric normalized adjacency matrix, The square root of the inverse of the degree matrix. Given an adjacency matrix with self-loops, This is the physical distance attenuation weight matrix. It is an exponential function. This function calculates the physical distance between entities. For entities without spatial coordinates, the distance between them returns a default value of 0. It only calculates the actual physical distance for entities with coordinates. This is a set of entity nodes in the slope risk knowledge graph, including displacement monitoring points, slope structure type, slope ratio, dip angle, slope height, soil type, groundwater depth, and rainfall event entities. To integrate the characteristics of risk propagation path sensing nodes constrained by physical distance, For graph attention network layers, For embedding layer, As a characteristic of knowledge evolution, For Transformer encoders, This is the distance attenuation coefficient;

[0114] Step 2.3: Based on the spatiotemporal characteristics of multi-source physical fields and knowledge evolution characteristics, feature fusion is performed to obtain knowledge-guided spatiotemporal joint features. The calculation method is as follows:

[0115]

[0116]

[0117]

[0118] in, Cross-attention features guided by knowledge. This is a multi-head cross-attention mechanism. The weights are for the first adaptive mode gating. It is a multilayer perceptron. Knowledge guides the spatiotemporal joint characteristics.

[0119] It should be further explained that slope deformation and failure are influenced by multiple factors, including stratum lithology, hydrological conditions, and engineering disturbances. A single monitoring indicator can only reflect the state of a certain dimension, and hidden hazards are difficult to identify directly from a single signal before a disaster occurs. For example, after continuous rainfall, the groundwater level has risen significantly, but the surface displacement rate is still fluctuating within the safe threshold range. At this time, cracks at the deep rock mass structure may already be slowly expanding. If only surface displacement data is relied upon for judgment, it is very easy to miss the internal risks that are brewing. At the same time, the geological endowment conditions of the slope—including slope structure type, slope ratio, dip angle, slope height, soil type, groundwater depth, etc.—constitute an indispensable prior knowledge framework for engineers to assess risks: soil slopes and rock slopes have different instability modes under the same rainfall conditions, and steep slopes and gentle slopes also have different responsiveness to changes in groundwater level. To address the above issues, this invention adopts a dual-path parallel architecture that complements data-driven and knowledge-driven approaches, achieving deep integration of multi-source physical fields and geological knowledge through a layered approach from raw monitoring signals to high-level semantics.

[0120] On the data processing side, the first layer stitches together four types of monitoring data—surface displacement, deep displacement, groundwater level, and daily rainfall—along the channel dimension, so that different physical monitoring quantities are uniformly expressed at the same time section. Next, considering that slope deformation exhibits both short-term rapid responses (such as a sharp rise in groundwater level within hours after a rainstorm, leading to accelerated shallow displacement) and long-term slow evolution (such as a continuous decline in groundwater level over several months during a dry season, resulting in continuous deformation of the slope at a slow creep rate), a single time-series model cannot adequately capture these two distinct dynamic characteristics. Therefore, a bidirectional gated cyclic unit traverses the stitched sequence in both forward and reverse time directions to capture the forward and backward temporal dependencies of each monitoring indicator—for example, the rise and fall of groundwater level requires a complete hydrological response cycle, and bidirectional modeling can simultaneously utilize historical information from the rising phase and future information from the falling phase to determine which stage of the cycle the current water level is in. Simultaneously, a self-attention mechanism is used to calculate the correlation strength between any two time steps in the sequence, and the feature table of each time step is... The model is represented as a weighted aggregation of features across all time steps. This allows the model to directly correlate daily displacement data with a key rainfall event from several weeks prior, unaffected by long-distance information attenuation during the time-series recursion. Building upon this, dilated convolution is introduced to perform multi-scale scanning of the attention output sequence under receptive fields of different dilation rates: small dilation rates capture local fluctuation patterns between adjacent days (e.g., accelerated displacement over two or three consecutive days), while large dilation rates capture trend changes across weeks or even months (e.g., cumulative deformation within a rainy season cycle). Finally, a gating fusion mechanism dynamically adjusts the contribution ratio of the two at the element-wise level. When the data exhibits a clear periodic pattern, the gating enhances the weight of the time-dependent features; when cross-scale anomalies occur, the gating enhances the weight of the multi-scale physical field features. This forms a multi-source spatiotemporal physical field feature that simultaneously reflects both short-term hydrological responses and long-term creep trends.

[0121] On the knowledge side, the geological priors relied upon for slope risk assessment are structured into graph data in the form of a slope risk knowledge graph. Entities in the knowledge graph cover key elements such as displacement monitoring points, slope structure types, slope ratio, dip angle, slope height, soil type, groundwater depth, and rainfall events. Relationships between entities are categorized into eight types: monitoring association, attribute affiliation, spatial bearing capacity, hydrological constraints, temporal induction, deformation response, parameter constraints, and coupling effects. For example, a rainfall event points to groundwater level changes through a "temporal induction" relationship; groundwater level changes are associated with a specific soil type through a "hydrological constraint" relationship; and this soil type, in turn, influences the deformation response pattern of the monitoring point through a "parameter constraint" relationship. A graph attention network layer calculates attention weights along these eight types of relationship edges and propagates information, ensuring that the feature updates of each entity node depend not only on the information of its neighbors but also on the type of connecting edge—the same rainfall information is transmitted to the groundwater level node through a "temporal induction" edge and through a "coupling" edge. When the "influence" edge is passed to the displacement monitoring point node, its transmission intensity and method are different. At the same time, the adjacency matrix incorporates attenuation constraints based on the physical distance between entities: for entities with spatial coordinates, such as the spatial distance between displacement monitoring points, the edge weight decreases exponentially as the distance increases, which is consistent with the practical understanding in slope engineering that "the closer the monitoring points are in space, the more similar their geological units and stress states are." For attribute entities without spatial coordinates, such as soil type and slope structure type, the distance attenuation is zero by default to ensure that the entity membership relationship edge is not weakened, so that the semantic constraints of the attribute entity on the monitoring point are fully preserved. On this basis, the Transformer encoder is used to globally encode the node features after propagation through the graph attention network layer, so that any two entities—even if they are not directly connected in the knowledge graph, such as between different monitoring points, or between rainfall events and slope rate—can establish indirect high-order interaction relationships through the self-attention mechanism, and finally output the knowledge evolution features of global semantic association.

[0122] On the fusion side, a multi-head cross-attention mechanism is used to perform cross-modal retrieval alignment by using the spatiotemporal features of multi-source physical fields as queries and knowledge evolution features as keys and values. This can be understood as the multi-source monitoring status at each moment searching the knowledge graph for the geological background information that best matches the current physical field pattern. For example, when monitoring data shows a pattern of "rapidly rising groundwater levels accompanied by slow increases in deep displacement," the cross-attention mechanism will retrieve soil types, slope ratios, and other attributes associated with this pattern from the knowledge graph, as well as historical deformation response records under similar rainfall conditions. To ensure that the two are truly complementary rather than simply superimposed, adaptive modal gating is introduced: the physical field... After the features and cross-attention features are concatenated along the channel dimension, a fusion weight at the element-wise granularity is learned through a multilayer perceptron. When the monitoring data signal is clear and the data from multiple sensors corroborate each other, the gating weight tends to retain more physical field features to make full use of real-time monitoring information. When the monitoring data signal is weak or there are contradictions between the sensors (e.g., the groundwater level rises but the surface displacement does not respond synchronously), the gating weight tends to rely more on the geological priors in the knowledge evolution features for auxiliary inference. Finally, after layer normalization, a knowledge-guided spatiotemporal joint feature that carries both real-time multi-physics coupling dynamics and embeds geological causal structures is obtained.

[0123] Compared to existing single-source driven judgment methods, the advantages of this invention are as follows: First, the parallel spatiotemporal feature extraction of multi-source monitoring data allows for the explicit modeling and preservation of the temporal response order and coupling relationship between multiple physical fields such as groundwater level fluctuations, deep rock deformation, and surface displacement response, solving the problem that a single indicator cannot determine the risk situation under multi-field feedback conditions; Second, the slope risk knowledge graph systematically injects prior knowledge scattered in survey reports, engineering experience, and geological data into the feature representation in a structured form, and through physical distance attenuation constraints and cross-attention retrieval, the knowledge guidance has spatial rationality and... Semantic targeting enables the model to make reasonable inferences using causal logic chains in the knowledge graph when encountering rainfall intensity-geological condition combinations not covered in the training data, thus alleviating the problem of insufficient generalization ability of pure data-driven methods in out-of-distribution scenarios. Thirdly, the introduction of adaptive modal gating allows physical field features and knowledge features to be dynamically matched at the element-by-element granularity according to the current data quality, avoiding the insufficient adaptability of fixed weights in existing fusion methods when facing data quality fluctuations. This allows the fused features to achieve the dual goals of ensuring accuracy by primarily using data-driven approaches when monitoring data is normal, and ensuring robustness by primarily using knowledge-driven approaches when monitoring data is abnormal.

[0124] Step 3: Based on the knowledge-guided spatiotemporal joint characteristics, knowledge evolution characteristics, and slope risk knowledge graph, perform long-range dependency encoding, progressive risk state evolution learning, and deep risk state distillation sequentially to obtain knowledge-enhanced long-range dependency characteristics, risk state evolution characteristics, and deep risk states; the specific process is as follows:

[0125] Step 3.1: Based on the knowledge-guided spatiotemporal joint features, knowledge evolution features, and slope risk knowledge graph, perform long-range dependency encoding to obtain knowledge-enhanced long-range dependency features. The calculation method is as follows:

[0126] ,

[0127] ,

[0128] ,

[0129] ,

[0130] in, As a feature of knowledge structure, This is a set of relationship types in the slope risk knowledge graph. Let r be the symmetric normalized adjacency matrix corresponding to relation type r, where r is the relation type index. This is a relation type weight matrix. This is a self-loop weight matrix. For attention knowledge pooling vectors, For linear layers, Enhance attentional features for knowledge. To enhance knowledge with long-range dependency features, For the replication function along the time dimension, Copy along the timeline to match the time sequence length; ⊙ represents the Hadamard product, and ⊕ represents element-wise addition.

[0131] Step 3.2: Based on the long-range dependency characteristics and knowledge evolution characteristics of knowledge reinforcement, perform progressive risk state evolution learning to obtain the risk state evolution characteristics. The calculation method is as follows:

[0132] ,

[0133] ,

[0134] ,

[0135] ,

[0136] in, The knowledge propagation feature sequence encoded by the graph attention network. Aggregate vectors for attention-based knowledge risks. For the second adaptive modal gating weights, It is the hyperbolic tangent function. Characteristics of risk state evolution;

[0137] Step 3.3: Based on the risk state evolution characteristics and knowledge evolution characteristics, perform deep risk state distillation to obtain the deep risk state. The calculation method is as follows:

[0138] ,

[0139] ,

[0140] ,

[0141] ,

[0142] ,

[0143] in, To illustrate the characteristics of attention propagation, This is a feature of multi-hop knowledge graph reasoning. For multi-hop feature fusion projection matrix, For cross-modal attention features, For the third adaptive modal gating weights, This indicates a state of deep risk.

[0144] Furthermore, the relationship types in step 3.1 include monitoring association, attribute affiliation, spatial carrying capacity, hydrological constraints, time-series induction, deformation response, parameter constraints, and coupling effects.

[0145] It needs to be further explained that the characteristic that distinguishes slope instability from general structural failure is that its evolution process is not triggered instantaneously by a single factor, but is the result of gradual transmission along a multi-hop causal chain. Taking rainfall-induced slope instability as an example: rainfall infiltration first causes the groundwater level to rise. The change in groundwater level leads to a decrease in effective stress by changing the pore water pressure distribution inside the slope. The attenuation of effective stress further weakens the shear strength of the potential sliding surface. When the shear strength is insufficient to resist the sliding force, deformation begins to accelerate and eventually develops into macroscopic instability. Each hop in this causal chain corresponds to a different physical mechanism and a different type of knowledge graph relationship—the relationship between rainfall events and groundwater level changes belongs to "time-series induction," the relationship between groundwater level and soil mechanical response belongs to "hydrological constraints" and "parameter constraints," and the relationship between mechanical response and deformation acceleration belongs to "deformation response." To address the above issues, based on the knowledge-guided spatiotemporal joint features and knowledge evolution features output in step 2, this invention uses a three-stage progressive modeling approach to gradually make the causal transmission path of risk explicit in the feature space.

[0146] The first stage involves long-range dependency encoding, addressing the issue of how to inject temporal features into the global risk situation of the knowledge graph. First, convolutional encoding of the eight types of relationships in the slope risk knowledge graph is performed: an independent weight matrix is ​​assigned to each type of relationship (monitoring association, attribute affiliation, spatial bearing capacity, hydrological constraints, temporal induction, deformation response, parameter constraints, and coupling effects), ensuring that the information aggregation method of the same entity node differs under different types of relationships—for example, a groundwater level entity uses one transformation method when receiving information from rainfall events through the "temporal induction" edge, and another transformation method when transmitting information to soil types through the "hydrological constraint" edge, thus explicitly distinguishing the information flow of different physical mechanisms at the feature level. After completing the relationship-specific encoding, a weighted pooling of the node features across the entire graph is performed using an attention mechanism to automatically identify the nodes in the current knowledge graph that are most critical for risk assessment. —For example, during periods of concentrated rainfall, attention weights may be tilted towards rainfall event nodes and groundwater depth nodes, while during the dry season they may be tilted towards slope structure type and slope rate nodes—and the key information of the entire graph is aggregated into a compact knowledge pooling vector; at the same time, multi-head self-attention is applied to the knowledge-guided spatiotemporal joint features output in step 2, so that the joint features of any two time steps in the sequence can interact directly, capturing long-range dependencies across time steps; subsequently, the knowledge pooling vector is copied to a time axis of the same length as the temporal features, concatenated with the temporal self-attention features and then convolved and gated, so that the global risk situation at the knowledge graph level acts as an auxiliary condition signal to regulate the expression of temporal features step by step—equivalent to the model understanding what the most concerned risk element of the entire knowledge graph is when analyzing the monitoring status at each moment, ensuring that the encoding of temporal features is always constrained by the global risk context;

[0147] The second stage is progressive risk state evolution learning, addressing the issue of how the risk state of the knowledge graph itself is updated step-by-step under multi-hop propagation. First, a first-order information propagation is performed on the knowledge graph using a graph attention network layer, obtaining a knowledge propagation feature sequence for each entity node after incorporating information from its direct neighbors. Next, the knowledge propagation features are globally aggregated using the same attention pooling method as in the first stage, resulting in an attention-based knowledge risk aggregation vector for the current graph state—this vector represents the overall risk posture of each entity in the knowledge graph after completing the first-order neighbor information exchange. On the temporal side, convolution and nonlinear transformations are applied to the long-range dependency features output from the first stage, extracting them as a representation reflecting the current temporal pattern characteristics. Then, the attention knowledge risk convergence vector is copied to the time axis and concatenated with the temporal feature. After convolution and sigmoid activation, a second adaptive modality gating weight is generated. This gating weight adjusts the fusion ratio between the temporal feature after convolution and the temporal offset after hyperbolic tangent function processing on an element-wise basis. When the risk convergence signal of the knowledge graph indicates that each entity is in a low-risk state, the gating tends to suppress the offset to maintain the stability of the feature. When the risk convergence signal of the knowledge graph indicates that multiple entity nodes enter a high-risk state at the same time, the gating tends to enhance the offset to strengthen the sensitivity of the feature to risk changes. Thus, the evolution of the risk state is gradual and driven by the knowledge graph, rather than a simple temporal recursion.

[0148] The third stage is deep risk state distillation, addressing the further extraction and refinement of risks in the second-order causal chain. First, based on the first-order knowledge propagation features encoded by the graph attention network in the previous stage, a second neighborhood aggregation is performed using the same graph attention network layer. At this point, the information received by each entity node not only comes from its direct neighbors but also indirectly includes information from its second-order neighbors, equivalent to implementing second-order reasoning on the knowledge graph. For example, the first-order propagation of a rainfall event node directly incentivizes its groundwater level neighbors, while the second-order propagation further incentivizes soil type nodes and displacement monitoring point nodes connected to the groundwater level entity with indirect information, thus fully covering the second-order causal chain of "rainfall → groundwater level → soil response → monitoring point deformation" in the feature space. The system first-order causal chain is constructed by concatenating first-order and second-order propagation features and then fusing them through a projection matrix to obtain multi-hop knowledge graph reasoning features. These features retain both first-order direct causal and second-order indirect causal information. Next, a multi-head cross-attention mechanism is used to perform cross-modal interaction between the time-series risk state evolution features as queries and the multi-hop knowledge graph reasoning features as keys and values. This means that when the time-series features are updating their risk states, they can simultaneously retrieve the corresponding single-hop and double-hop causal evidence in the knowledge graph, thus giving the time-series feature evolution causal traceability. Finally, a third adaptive modal gating weight is introduced to adjust the fusion ratio of the risk state evolution features and the cross-modal cross-attention features at the element-wise granularity, and the deep risk state is obtained after layer normalization.

[0149] Compared to existing technologies that rely on single-layer graph convolution or direct temporal prediction, this approach offers several advantages: First, the differential encoding of various relationships by relational graph convolution allows for the explicit separation and modeling of causal signals from different types of physical mechanisms in the feature space, avoiding information confusion that occurs when graph networks with single-sided weights process physical relationships with vastly different properties under the same type of attention coefficients. Second, the stage-by-stage extraction and injection of attention knowledge pooling vectors and attention knowledge risk convergence vectors enable the global risk situation at the knowledge graph level to serve as a conditional signal continuously guiding the encoding and evolution of temporal features, rather than just as a one-time input feature. This allows the model to make forward-looking inferences using the causal structure of the knowledge graph even when there are no obvious anomalies in the temporal data. Third, the second-order inference mechanism of the double-hop graph attention network directly corresponds to the multi-hop causal chain in slope risk assessment, ensuring that the deep risk state output by the model has causal traceability from cause to effect in physical logic, rather than relying solely on black-box inference based on statistical correlation.

[0150] Step 4: Based on the knowledge-guided spatiotemporal joint features and deep risk state, perform multi-granularity risk feature extraction and fusion, nonlinear risk state update modeling, global risk correlation modeling, and comprehensive risk quantification in sequence to obtain multi-granularity calibration features, risk evolution features, attention-enhanced risk features, and a comprehensive risk index; the specific process is as follows:

[0151] Step 4.1: Based on the knowledge-guided spatiotemporal joint features and deep risk state sequences, multi-granularity risk features are extracted and fused through multi-scale dilated temporal convolution and cross-source dimensional adaptive calibration gating mechanisms to obtain multi-granularity calibration features. The calculation method is as follows:

[0152] ,

[0153] ,

[0154] ,

[0155] ,

[0156] in, This represents the initial fusion characteristics across multiple scales. To use a one-dimensional convolutional layer with a first convolutional kernel, To employ a one-dimensional dilated convolutional layer with a second convolutional kernel, To employ a one-dimensional dilated convolutional layer with a third convolutional kernel, For multi-scale deep fusion features, To adaptively calibrate the gate weights, This is a multi-granularity calibration feature;

[0157] Step 4.2: Based on the multi-granularity calibration characteristics, perform nonlinear risk state update modeling to obtain risk evolution characteristics. The calculation method is as follows:

[0158] ,

[0159] ,

[0160] ,

[0161] in, It features bidirectional gating fusion. This represents the risk state offset. Characteristics of risk evolution;

[0162] Step 4.3: Based on the risk evolution characteristics, a global risk association model is performed using a temporal multi-head self-attention enhancement and adaptive confidence weighted fusion mechanism to obtain attention-enhanced risk features. The calculation method is as follows:

[0163] ,

[0164] ,

[0165] ,

[0166] in, As a characteristic of attention risk, To enhance attention to risk characteristics, For element-wise confidence weights, It is a soft positive activation function;

[0167] Step 4.4: Based on the risk characteristics of attention enhancement, a comprehensive risk quantification is performed through temporal global pooling and composite risk regression mapping to obtain a comprehensive risk index. The calculation method is as follows:

[0168] ,

[0169] ,

[0170] in, This is a one-dimensional global average pooling operation. The original risk regression value, As a comprehensive risk index, This is a numerical clipping operation. This is the risk composite adjustment coefficient. These are the lower and upper bounds for cropping, respectively.

[0171] It is important to further explain that the key difference between slope deformation monitoring and general time-series prediction tasks lies in the fact that slope displacement signals are essentially the superposition of deformation components at multiple time scales, and there is a decoupling phenomenon between surface monitoring data and the actual internal risk state. Specifically, after a rainstorm, slope displacement may respond rapidly within hours to days—this is deformation caused by a sudden increase in the water content of shallow soil. Over a longer time span, seasonal fluctuations in groundwater levels caused by the alternation of rainy and dry seasons will cause the slope to undergo reciprocating deformation on a monthly cycle. At a more macroscopic scale, the slope itself also exhibits continuous creep on an annual basis, the rate of which is low but the direction is constant and irreversible. The deformation components at different time scales are superimposed on the same condition monitoring curve in real time, making it impossible for single receptive field models—whether they focus only on short-window convolutions that focus on the most recent sampling points or global pooling that only captures trend changes—to effectively decouple the contributions of components at different scales. Narrow-window models are prone to misjudging short-term acceleration caused by a rainfall event as a dangerous precursor to entering an accelerated creep stage, while wide-window models may drown out the slow trend of interannual creep with the noise of seasonal fluctuations. To address the above problems, this invention, based on the knowledge-guided spatiotemporal joint features output in step 2 and the deep risk status output in step 3, completes the entire link from multi-granularity feature extraction to comprehensive risk quantification through four progressive stages.

[0172] The first stage begins with a three-way parallel convolutional branch to perform multi-scale scanning of the knowledge-guided spatiotemporal joint features: a conventional one-dimensional convolution with the first kernel is used to capture local deformation patterns between adjacent days—for example, whether fluctuations in displacement rate over three to five consecutive days constitute a local acceleration trend; a one-dimensional dilated convolution with the second kernel is used to skip sampling in the sequence with a larger dilation rate to capture seasonal responses spanning weeks to months; and a one-dimensional dilated convolution with the third kernel is used to cover the entire monitoring period with an even larger dilation rate. The outputs of the three branches are concatenated along the channel dimension and then fused by convolution to obtain a preliminary multi-scale fused feature that simultaneously contains short, medium, and long-scale features. Based on this, the preliminary multi-scale fused feature is concatenated again with the original knowledge-guided spatiotemporal joint features, and then processed by a multilayer perceptron and layer normalization to obtain a deep multi-scale fused feature—this back-and-forth design ensures that the fine-grained information in the original time series is not diluted by multiple convolutions during the expansion of the receptive field. At this point, the multi-scale feature extraction on the time series side is completed, and the key question is how to utilize the output of step 3. The deep risk state is calibrated by the following method: the deep risk state is a feature representation reflecting the degree of internal damage, obtained through double-hop causal reasoning and stepwise gating distillation on a knowledge graph, while the multi-scale deep fusion feature is the representational encoding of the current monitoring data—there may be a disconnect between the two. Therefore, an adaptive calibration gating is introduced. The multi-scale deep fusion feature and the deep risk state are linearly mapped and then added element-wise. The gating weight is generated by multilayer perceptron and sigmoid activation. This weight is dynamically calibrated element-wise between the two: when the deep risk state indicates a high degree of internal damage while the multi-scale deep fusion feature is still within the normal range, the gating tends to enhance the expression weight of the deep risk state, so that the output feature can still reflect the gradual deviation of internal risk under the condition of stable data representation; when the risk levels indicated by the two are consistent, the gating tends to achieve balanced fusion. The calibrated multi-granularity calibration feature realizes the adaptive alignment of the surface deformation pattern of multiple time scales with the internal risk state obtained by deep causal reasoning at the element-wise granularity.

[0173] In the second stage, multi-granularity calibration features are fed into a bidirectional gated recurrent unit, allowing the calibrated features to undergo state traversal and fusion along both forward and reverse time directions to obtain bidirectional gated fused features. Based on this, the features are processed using two parallel convolutional paths: one path is activated by convolution and the hyperbolic tangent function to generate candidate offsets with values ​​in the positive and negative intervals—positive offsets correspond to increased risk, and negative offsets correspond to decreased risk; the other path is activated by convolution and the sigmoid function to generate a gating signal with values ​​in the (0,1) interval—controlling whether the candidate offset is activated at the current position; both... After performing the Hadamard product, a risk state offset is formed. This offset is superimposed on the original multi-granularity calibration feature only at the position and extent allowed by the gating signal. This candidate offset + gating modulation design means that the risk state update does not occur at the same amplitude in every time step, but is triggered only when the gating determines that the current context supports the change of risk state. When the monitoring data of multiple consecutive time steps show a consistent deterioration trend and deviate from the deep risk state in the same direction, the gating continues to be open, and the risk offset accumulates step by step, which is finally manifested in the feature space as a leap from quantitative to qualitative change in risk state.

[0174] The third stage applies multi-head self-attention to the risk evolution features, enabling direct interaction between the risk evolution features of any two time steps in the sequence. This means that the increment of the risk signal at a certain moment is not only determined by the features of that moment itself, but also by the risk status of other moments in the entire monitoring period. If the anomaly at that moment is isolated in the entire sequence (e.g., an occasional jump in a single sensor), the self-attention mechanism will naturally weaken its influence due to its low correlation with other time steps. Subsequently, a multilayer perceptron and a soft positive activation function are applied to the self-attention output to generate a positive confidence weight for each feature element. The soft positive activation ensures that the weight is always non-negative and the output curve is smooth, avoiding feature loss caused by zero or negative confidence. The self-attention features are then weighted element by element with this confidence and superimposed with residual connections to obtain attention-enhanced risk features. The effect of this confidence weighting is that time periods with clear patterns and sufficient confirmation in the global context are given higher confidence, and their risk signals are enhanced. Time periods with ambiguous patterns and lack of consistency in the global context are given lower confidence, and their risk signals are suppressed, thereby reducing the probability of false alarms and false negatives overall.

[0175] The fourth stage applies one-dimensional global average pooling to the attention-enhanced risk features, compressing the features across the entire time axis into a fixed-dimensional representation vector. This vector is then processed by a multilayer perceptron and activated by a sigmoid function to obtain the original risk regression value, which represents the risk estimate based on the average state across all monitoring periods. Considering that the smoothing effect of global average pooling on extreme values ​​may make the model insensitive to risk signals that deteriorate rapidly during the pre-disaster phase, a parallel branch is established: one-dimensional convolution is first applied to the attention-enhanced risk features to strengthen the local anomaly response, followed by one-dimensional global average pooling and hyperbolic tangent activation for risk composite adjustment. The coefficients are weighted element-wise and added to the original risk regression value. Finally, the numerical values ​​are clipped to a preset lower and upper bound to obtain the comprehensive risk index. This bi-branch composite regression design allows the comprehensive risk index to reflect the overall average risk level within the monitoring period while maintaining responsiveness to periods of rapid deterioration before a disaster. When the monitoring data is stable throughout, the contribution of the adjustment branch is close to zero, and the comprehensive risk index mainly depends on the original risk regression value. When there is a sharp acceleration in the pre-disaster stage, the adjustment branch, after being activated by Tanh, outputs a large positive value, pushing the comprehensive risk index higher in the dangerous direction.

[0176] Compared to existing technologies that directly output risk values ​​using convolutional networks with fixed receptive fields or single recurrent networks, the technical advantages of this invention are as follows: First, the three-branch parallel design of multi-scale dilated temporal convolution enables deformation components of three different time scales—short-term hydrological response, seasonal fluctuations, and long-term creep trends—to be perceived and explicitly fused separately, avoiding misjudgments by single receptive field models in aliased multi-scale signals. Second, adaptive calibration gating introduces deep risk states as calibration signals into the multi-granularity fusion process of temporal features, allowing the model to calibrate the expression of surface features by using deviations from deep risk states when surface monitoring data shows no obvious anomalies but internal damage has been continuously developing, solving the problem of missed judgments in scenarios where pure data-driven methods do not match the surface and internal conditions. Third, the risk state offset adopts a candidate offset plus gating modulation mechanism, ensuring that the evolution of risk is not a simple temporal recursion, but is activated and accumulated only when the gating signal confirms a favorable context, which is closer to the physical law of long-term quantitative accumulation → short-term qualitative change triggering in slope creep damage.

[0177] Step 5: Based on the comprehensive risk index, conduct risk status assessment and risk level determination sequentially to obtain a quantitative risk scoring vector and predict the risk level; the specific process is as follows:

[0178] Step 5.1: Assess the risk status based on the comprehensive risk index to obtain a multi-level risk response vector. The calculation method is as follows:

[0179] ,

[0180] ,

[0181] in, For risk response gating vector, In response to the weight matrix, In response to the bias vector, This is a quantitative risk scoring vector;

[0182] Step 5.2: Based on the quantified risk scoring vector, the risk level is determined through threshold comparison and probabilistic mapping to obtain the predicted risk level. The calculation method is as follows:

[0183] ,

[0184] ,

[0185] in, For the threshold response vector, This is the temperature scaling factor. For the level mapping weight matrix, For the level mapping bias vector, To predict the risk level, This is the index for retrieving the maximum value.

[0186] Example 2

[0187] In this embodiment, the specific parameter settings for various neural networks are as follows:

[0188] For initial feature extraction of multi-source monitoring data, one-dimensional convolution is used to map the concatenated sequences of four types of monitoring data—surface displacement, deep displacement, groundwater level, and daily rainfall—along the channel dimension to a 128-dimensional feature space, with a convolution kernel length of 3. The hidden state dimension of the bidirectional gated recurrent unit is set to 128. The multi-scale dilated convolution uses three parallel branches with dilation rates of 1, 7, and 60, respectively, and a convolution kernel length of 3, so that the receptive fields of each branch cover daily-scale local fluctuations, weekly-to-monthly-scale seasonal responses, and cross-monthly to inter-yearly trend changes, respectively. The number of heads in the multi-head attention mechanism is uniformly set to 8 at each stage. On the knowledge graph side, the dimension of the entity embedding vector is set to 128. The hidden dimension of the graph attention network layer is 128, and the number of attention heads is 8. The Transformer encoder contains 4 stacked layers, with 8 attention heads in each layer. The distance decay coefficient λ is set to 0.01; the relation graph convolution sets independent transformation weights for each of the eight relation types, and the hidden dimension of each relation type is 128; the multilayer perceptron in each stage adopts a two-layer structure: the first layer maps the input to 256 dimensions, the second layer maps it back to 128 dimensions, and the ReLU activation function is used in between; the multi-hop feature fusion projection matrix fuses the concatenated first-order and second-order knowledge propagation features from 256 dimensions to 128 dimensions; in the multi-granularity risk feature extraction stage, the number of output channels of the three convolution branches is set to 64, which is 192 dimensions after channel concatenation, and then fused to 128 dimensions through convolution; in the risk state update stage, the hidden state dimension of the bidirectional gated recurrent unit is 128, and the kernel length of the two parallel convolution branches that generate candidate offsets and gate signals is 3; in the comprehensive risk quantification stage, the risk composite adjustment coefficient is... Set the value to 0.2 and crop the lower bound. Crop top In the risk level determination process, the temperature scaling factor τ is set to 1.0, and the output includes four levels: low risk, medium risk, high risk, and extremely high risk.

[0189] In terms of training methods, this embodiment uses the Adam optimizer with an initial learning rate of 0.001 and a cosine annealing strategy for learning rate decay. The batch size is set to 32, and the maximum number of training epochs is set to 200. The loss function consists of two parts: mean squared error loss for regression tasks of the comprehensive risk index and cross-entropy loss for classification tasks of risk level determination. The two are weighted and summed in a ratio of 0.7:0.3 to form the total loss function for joint optimization. An early stopping strategy is introduced during training. If the total loss of the validation set does not decrease within 20 consecutive epochs, training is terminated early to prevent overfitting.

[0190] This embodiment selects two typical slopes (slope A and slope B) of a highway in a mountainous area of ​​southern China as a joint monitoring and assessment scenario. Both slopes are key protection slopes of the road section. Slope A is a soil slope mainly composed of silty clay, with a slope height of 45 meters, an overall slope ratio of 1:1.50, and a slope angle of 33°. A 2-4m thick layer of gravelly soil exists in the middle layer of the slope. The natural moisture content of the soil is 22%~28%, the cohesion is 18~25 kPa, the internal friction angle is 12°~18°, and the groundwater depth in the area is 3.2~5.8m. Slope B is a strongly weathered sandstone slope, with a slope height of 80 meters, an overall slope ratio of 1:0.75, and a slope angle of 53°. The rock mass has well-developed joints and fissures, an RQD value of 25%~50%, and a uniaxial compressive strength of 5~15 MPa. The groundwater depth in the area is 8.5~12.1m. The slope risk knowledge graphs constructed based on slopes A and B in this embodiment are as follows: Figure 1 As shown.

[0191] Figure 2 This is a graph showing the results of slope risk early warning and monitoring data provided by the present invention. Figure 2 As shown, the monitoring and evaluation period is 123 days from June 1 to September 30, 2024, and daily risk warnings are issued for slope A (soil slope) and slope B (rock slope).

[0192] Slope A had a comprehensive risk index of only 0.08 at the end of the dry season in early June, indicating a low-risk status. After the heavy rain on June 15, the groundwater level dropped from 5.72m to 5.05m, and the comprehensive risk index rose to 0.22, then jumped to 0.31 the following day, entering the medium-risk range. During the continuous rainfall from July 20 to 22, the comprehensive risk index climbed from 0.48 to 0.62, triggering a high-risk warning for the first time. On August 10, the day of the extreme rainstorm (120.7mm), the comprehensive risk index reached 0.85, classifying it as extremely high-risk, and peaked at 0.92 on August 15. Subsequently, as the rainfall weakened, the risk index gradually declined, dropping to 0.48 by September 30, returning to the medium-risk range.

[0193] Due to the low permeability of the bedrock and the large depth of groundwater, the comprehensive risk index of slope B remained below 0.15 from June to July, maintaining a low-risk status. This stable situation was broken by extreme rainstorms on August 10, when the groundwater level dropped sharply by 1.44m in two days. The comprehensive risk index jumped from 0.28 to 0.42, entering the medium-risk category. It reached 0.63 on August 20, entering the high-risk category, and reached a peak of 0.67 on August 25.

[0194] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.

Claims

1. A Transformer-based slope knowledge graph risk early warning method, characterized in that, Includes the following steps: Step 1: Obtain a slope risk knowledge graph; Time series data of surface displacement, deep displacement, groundwater level, and daily rainfall were collected from the slope and preprocessed to obtain preprocessed time series data of surface displacement, deep displacement, groundwater level, and daily rainfall. Step 2: Based on the preprocessed surface displacement time series data, preprocessed deep displacement time series data, preprocessed groundwater level time series data, preprocessed daily rainfall time series data, and slope risk knowledge graph, extract the spatiotemporal features of multi-source physical fields, knowledge evolution features, and knowledge-guided spatiotemporal joint features in sequence. Step 3: Based on the knowledge-guided spatiotemporal joint characteristics, knowledge evolution characteristics, and slope risk knowledge graph, perform long-range dependency encoding, progressive risk state evolution learning, and deep risk state distillation in sequence to obtain knowledge-enhanced long-range dependency characteristics, risk state evolution characteristics, and deep risk states. The specific process is as follows: Step 3.1: Based on the knowledge-guided spatiotemporal joint features, knowledge evolution features, and slope risk knowledge graph, perform long-range dependency encoding to obtain knowledge-enhanced long-range dependency features. The calculation method is as follows: , , , , in, As a feature of knowledge structure, This is a set of relationship types in the slope risk knowledge graph. Let r be the symmetric normalized adjacency matrix corresponding to relation type r, where r is the relation type index. This is a relation type weight matrix. This is a self-loop weight matrix. For attention knowledge pooling vectors, For linear layers, Enhance attentional features for knowledge. To enhance knowledge with long-range dependency features, For the replication function along the time dimension, Copy along the timeline to match the time sequence length; ⊙ represents the Hadamard product, and ⊕ represents element-wise addition. Step 3.2: Based on the long-range dependency characteristics and knowledge evolution characteristics of knowledge reinforcement, perform progressive risk state evolution learning to obtain the risk state evolution characteristics. The calculation method is as follows: , , , , in, The knowledge propagation feature sequence encoded by the graph attention network. Aggregate vectors for attention-based knowledge risks. For the second adaptive modal gating weights, It is the hyperbolic tangent function. Characteristics of risk state evolution; Step 3.3: Based on the risk state evolution characteristics and knowledge evolution characteristics, perform deep risk state distillation to obtain the deep risk state. The calculation method is as follows: , , , , , in, To illustrate the characteristics of attention propagation, This is a feature of multi-hop knowledge graph reasoning. For multi-hop feature fusion projection matrix, For cross-modal attention features, For the third adaptive mode gating weights, This is a state of deep risk. Step 4: Based on the knowledge-guided spatiotemporal joint features and deep risk status, perform multi-granularity risk feature extraction and fusion, nonlinear risk status update modeling, global risk correlation modeling, and comprehensive risk quantification in sequence to obtain multi-granularity calibration features, risk evolution features, attention-enhanced risk features, and comprehensive risk index. Step 5: Based on the comprehensive risk index, conduct risk status assessment and risk level determination in sequence to obtain a quantitative risk score vector and predict the risk level.

2. The slope knowledge graph risk early warning method as described in claim 1, characterized in that, Surface displacement time series data were acquired using a GNSS receiver, deep displacement time series data were acquired using an inclinometer, groundwater level time series data were acquired using a water level gauge, and daily rainfall time series data for the slope area were acquired using a rain gauge; all four types of data are time series numerical data.

3. The slope knowledge graph risk early warning method as described in claim 1, characterized in that, The data type of the slope risk knowledge graph is structured knowledge graph data; the entity types in the slope risk knowledge graph include displacement monitoring points, slope structure type, slope ratio, dip angle, slope height, soil type, groundwater depth, and rainfall events; the relationship types in the slope risk knowledge graph include monitoring association, attribute affiliation, spatial bearing capacity, hydrological constraints, time-series induction, deformation response, parameter constraints, and coupling effects.

4. The slope knowledge graph risk early warning method as described in claim 1, characterized in that, The specific process of time-series numerical data preprocessing is as follows: First, the collected time-series numerical data is cleaned: outliers are removed using the three-standard-deviation method, missing values ​​are filled in using linear interpolation, and finally Z-score standardization is used to obtain the cleaned time-series numerical data. The cleaned time-series numerical data is timestamped and the sampling frequency is standardized to obtain preprocessed time-series numerical data.

5. The slope knowledge graph risk early warning method as described in claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1: Based on the preprocessed surface displacement time series data, preprocessed deep displacement time series data, preprocessed groundwater level time series data, and preprocessed daily rainfall time series data, parallel spatiotemporal feature extraction is performed to obtain the spatiotemporal features of the multi-source physical field. The calculation method is as follows: , , , , in, For multi-source monitoring time series splicing tensor, This is a channel-level splicing operation. This is the preprocessed surface displacement time series data. This is the preprocessed deep displacement time series data. This is the preprocessed groundwater level time series data. This is the preprocessed time-series data of daily rainfall. It is a time-dependent feature. It is a bidirectional gated loop unit. It is a one-dimensional convolutional network. It is a multi-scale physical field characteristic. This is a one-dimensional dilated convolution operation. For the Softmax function, For feature dimension, The spatiotemporal characteristics of multi-source physical fields For layer normalization operation, For transpose operation, For element-wise addition, For Hadama accumulation, Use the Sigmoid activation function; Step 2.2: Based on the slope risk knowledge graph, semantic knowledge features are extracted to obtain knowledge evolution features. The calculation method is as follows: , , , , in, For a symmetric normalized adjacency matrix, The square root of the inverse of the degree matrix. Given an adjacency matrix with self-loops, This is the physical distance attenuation weight matrix. It is an exponential function. This function calculates the physical distance between entities. For entities without spatial coordinates, the distance between them returns a default value of 0. It only calculates the actual physical distance for entities with coordinates. This is a set of entity nodes in the slope risk knowledge graph, including displacement monitoring points, slope structure type, slope ratio, dip angle, slope height, soil type, groundwater depth, and rainfall event entities. To integrate the characteristics of risk propagation path sensing nodes constrained by physical distance, For graph attention network layers, For embedding layer, As a characteristic of knowledge evolution, For Transformer encoders, This is the distance attenuation coefficient; Step 2.3: Based on the spatiotemporal characteristics of multi-source physical fields and knowledge evolution characteristics, feature fusion is performed to obtain knowledge-guided spatiotemporal joint features. The calculation method is as follows: , , , in, Cross-attention features guided by knowledge. This is a multi-head cross-attention mechanism. The weights are for the first adaptive mode gating. It is a multilayer perceptron. Knowledge guides the spatiotemporal joint characteristics.

6. The slope knowledge graph risk early warning method as described in claim 5, characterized in that, The specific process of step 4 is as follows: Step 4.1: Based on the knowledge-guided spatiotemporal joint features and deep risk state sequences, multi-granularity risk features are extracted and fused through multi-scale dilated temporal convolution and cross-source dimensional adaptive calibration gating mechanisms to obtain multi-granularity calibration features. The calculation method is as follows: , , , , in, This represents the initial fusion characteristics across multiple scales. To use a one-dimensional convolutional layer with a first convolutional kernel, To employ a one-dimensional dilated convolutional layer with a second convolutional kernel, To employ a one-dimensional dilated convolutional layer with a third convolutional kernel, For multi-scale deep fusion features, To adaptively calibrate the gate weights, This is a multi-granularity calibration feature; Step 4.2: Based on the multi-granularity calibration characteristics, perform nonlinear risk state update modeling to obtain risk evolution characteristics. The calculation method is as follows: , , , in, It features bidirectional gating fusion. This represents the risk state offset. Characteristics of risk evolution; Step 4.3: Based on the risk evolution characteristics, a global risk association model is performed using a temporal multi-head self-attention enhancement and adaptive confidence weighted fusion mechanism to obtain attention-enhanced risk features. The calculation method is as follows: , , , in, As a characteristic of attention risk, To enhance attention to risk characteristics, For element-wise confidence weights, It is a soft positive activation function; Step 4.4: Based on the risk characteristics of attention enhancement, a comprehensive risk quantification is performed through temporal global pooling and composite risk regression mapping to obtain the comprehensive risk index. The calculation method is as follows: , , in, This is a one-dimensional global average pooling operation. The original risk regression value, To comprehensively assess risk characteristics, This is a numerical clipping operation. This is the risk composite adjustment coefficient. These are the lower and upper bounds for cropping, respectively.

7. The slope knowledge graph risk early warning method as described in claim 6, characterized in that, The specific process of step 5 is as follows: Step 5.1: Conduct a risk status assessment based on comprehensive risk characteristics to obtain a multi-level risk response vector. The calculation method is as follows: , , in, For risk response gating vector, In response to the weight matrix, In response to the bias vector, This is a quantitative risk scoring vector; Step 5.2: Based on the quantified risk scoring vector, the risk level is determined through threshold comparison and probabilistic mapping to obtain the predicted risk level. The calculation method is as follows: , , in, For the threshold response vector, This is the temperature scaling factor. The level mapping weight matrix, For the level mapping bias vector, To predict the risk level, This is the index for retrieving the maximum value.

Citation Information

Patent Citations

  • Railway foreign matter invasion real-time monitoring and alarming system based on SAR (Synthetic Aperture Radar)

    CN120428231A

  • Multi-source monitoring and early warning method for high and steep slope of strip mine based on graph neural network and Transform

    CN121330851A