Xinjiang cold and arid slope risk prevention and control method based on knowledge graph

Abstract

The application discloses a Xinjiang cold and arid slope risk prevention and control method based on a knowledge graph, belongs to the technical field of cold and arid region slope disaster risk prevention and control, and aims at solving the problems of non-uniform data format, weak correlation and insufficient prevention and control measures matching of cold and arid slope disasters. The application realizes the technical effects of improving the risk chain identification accuracy and the measure matching by constructing a heterogeneous space-time knowledge graph, generating multiple types of vectors, adopting a space-time graph transformation coding model to identify the instability chain and search similar cases, outputting a risk level, a treatment priority and recommended measures.

Description

Technical Field

[0001] This invention relates to the field of slope disaster risk prevention and control technology in cold and arid regions, and in particular to a knowledge graph-based method for slope risk prevention and control in Xinjiang's cold and arid regions. Background Technology

[0002] In the cold and arid regions of Xinjiang, slopes along highways, mines, power transmission lines, and other engineering projects are subjected to a complex environment of strong wind erosion, freeze-thaw cycles, snowmelt replenishment, salt crystallization, engineering excavation, and seismic disturbance. Slope instability is often not caused by a single factor, but rather by a chain of disasters formed by geological lithology, slope structure, external factors, engineering activities, and protection status. Current investigation work typically relies on disaster records, monitoring data, inspection conclusions, and treatment data to assess slope risk, and selects treatment measures based on experience, such as drainage and interception, slope toe protection, anchoring and retaining, windbreak and sand fixation, or focused inspections.

[0003] However, the data sources and recording formats differ across regions and engineering stages. There is a lack of unified spatiotemporal indexes and relational expressions among disaster records, monitoring indicators, treatment measures, and treatment effects, making it difficult to quickly reuse disaster chains that have occurred under similar cold and drought conditions. On the other hand, existing risk assessment methods are insufficient in distinguishing between triggering time windows such as freeze-thaw periods, snowmelt periods, strong wind periods, salt crystallization periods, excavation periods, and post-earthquake periods, making it difficult to accurately identify the main controlling factors and their propagation intensity along the disaster chain path.

[0004] In addition, the matching of prevention and control measures often relies on single-point anomalies or human experience, lacking comprehensive utilization of the treatment effects of historical cases, the degree of overlap of links, the consistency of environmental zoning, and the reliability of data. This can easily lead to insufficient correspondence between prevention and control measures and actual triggering factors, and affect the judgment of the governance priority of key road sections and key projects.

[0005] Therefore, a risk prevention and control method for cold and arid slopes in Xinjiang is needed to address the shortcomings of existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose a knowledge graph-based method for risk prevention and control of cold and drought-stricken slopes in Xinjiang. Addressing the problems in existing technologies such as inconsistent data formats, weak correlations, difficulty in identifying similar instability chains, and insufficient matching between prevention and control measures and actual triggering factors, this invention proposes a heterogeneous spatiotemporal knowledge graph for cold and drought-stricken slope disasters. This graph maps target slopes to environmental similarity vectors, monitoring anomaly vectors, and disaster chain path vectors. Candidate instability chains are identified through cold and drought triggering event windows, path constraint attention, and spatiotemporal graph transformation coding models. Furthermore, the invention combines similar historical case retrieval to calculate risk levels, governance priorities, and recommended prevention and control measures. This invention effectively improves the accuracy of risk chain identification under cold and drought triggering factors and enhances the matching between prevention and control measures and the main triggering factors.

[0007] This invention provides a knowledge graph-based method for risk prevention and control of cold and arid slopes in Xinjiang, comprising: S1, converting slope disaster records, monitoring data, inspection conclusions, and treatment data into entities, relationships, and attributes with temporal and spatial attributes, and constructing a heterogeneous spatiotemporal knowledge graph for cold and arid slope disasters; S2, mapping environmental data, monitoring data, and historical disaster chain data of the target slope to the heterogeneous spatiotemporal knowledge graph, generating environmental similarity vectors, monitoring anomaly vectors, and disaster chain path vectors; S3, constructing time window nodes for cold and arid inducing events, calculating the activation coefficients of each event time window node, and inputting the environmental similarity vectors, monitoring anomaly vectors, disaster chain path vectors, and activation coefficients into the spatiotemporal graph. The coding model is changed, and path constraint attention weights are generated based on relation type, time window activation, spatial adjacency, and data reliability. The chain activation vector, main inducing factor contribution, and link confidence of the candidate unstable chain are output. S4. Similar historical cases are retrieved based on graph embedding similarity, link overlap, cold and drought environment zoning consistency, and treatment effect similarity to obtain measure matching correction weights and priority correction coefficients. S5. The risk level and governance priority are calculated based on the chain activation vector, engineering exposure, disaster consequence quantification value, data confidence, and priority correction coefficients. Recommended prevention and control measures and priority treatment order are generated based on the main inducing factor contribution, link confidence, and measure matching correction weights.

[0008] Optionally, S1 includes:

[0009] The slope unit, stratigraphic lithology, slope structure, wind erosion status, freeze-thaw status, snowmelt water supply, salt swelling status, excavation disturbance, seismic disturbance, monitoring indicators, inspection conclusions, disaster morphology, prevention and control measures, and treatment effects are regarded as entity types.

[0010] Consider the following as relation types: located, adjacent, induced, evolved, adopted measures, and formed effects;

[0011] Coordinates, mileage markers, monitoring time, inspection time, and treatment time from different data sources are uniformly encoded to generate a spatiotemporal index key for each entity and relationship;

[0012] Missing fields are marked as unobserved attributes and data source identifiers are recorded to obtain the heterogeneous spatiotemporal knowledge graph.

[0013] Optionally, S2 includes:

[0014] The environmental similarity vector is generated from the cold and arid environment zoning, slope aspect, slope angle, altitude, stratigraphic lithology, structural plane occurrence, wind erosion intensity, freeze-thaw cycle number, snowmelt water replenishment and salinity of the target slope.

[0015] The displacement rate, crack opening amount, groundwater level, slope temperature, water content, and inspection disease type are compared with the corresponding monitoring benchmark interval, and the monitoring anomaly vector is generated according to the magnitude and duration of exceeding the interval boundary.

[0016] The causal entities, process relationships, and disaster forms in the historical disaster chain are encoded according to the order of their occurrence along the path to generate the disaster chain path vector.

[0017] The environmental similarity vector, the monitoring anomaly vector, and the disaster chain path vector are used together as the model input for step S3.

[0018] Optionally, S3 includes:

[0019] The freeze-thaw period, snow melting period, strong wind period, salt crystallization period, excavation period, and post-earthquake period are respectively constructed as event time window nodes;

[0020] The activation coefficients of the corresponding triggers are calculated based on the number of times the temperature crosses zero, the amount of snow melt, the cumulative wind speed, the salt crystallization saturation, the excavation advance, and the peak ground acceleration of the earthquake.

[0021] Write the activation coefficient of the trigger into the time window attribute of adjacent entities and relationships;

[0022] In the spatiotemporal graph transformation coding model, the path constraint attention weight is generated by multiplying relation type embedding, event window node embedding, spatial adjacency weight and data reliability and then normalizing the result.

[0023] Based on the path constraint attention weights and the characteristics of the nodes propagating along the disaster chain path, the chain activation vector, the contribution of the main controlling factor, and the link confidence are obtained.

[0024] Furthermore, the path constraint attention weights are also generated based on path templates: historical disaster chains and expert rules are converted into path templates, each path template including a starting event window node, a sequence of causal relationships, a sequence of intermediate entity types, a terminal disaster form, and an allowed time interval;

[0025] When the entity type, relation type, and event window of a candidate propagation edge fall into any path template, the attention propagation coefficient of that candidate propagation edge is retained.

[0026] When a candidate propagation edge does not fall into any path template, the attention propagation coefficient of the candidate propagation edge is set to zero.

[0027] The path-constrained attention weights are obtained by normalizing the retained attention propagation coefficients;

[0028] Furthermore, the spatiotemporal graph transformation coding model is trained in the following way: a training subgraph with disaster chain annotation, main inducing factor annotation, treatment measure annotation and treatment effect annotation is extracted from the heterogeneous spatiotemporal knowledge graph of historical periods;

[0029] Generate training environment similarity vectors, training monitoring anomaly vectors, and training disaster chain path vectors in the same manner as in step S2, and generate training induced activation coefficients in the same manner as in step S3.

[0030] The spatiotemporal graph transformation coding model is obtained by using the weighted sum of candidate unstable chain classification loss, chain activation regression loss, main inducing factor contribution ranking loss and link confidence calibration loss as the training objective.

[0031] Optionally, S4 includes:

[0032] Graph embedding is performed on the target slope sub-map and the historical case sub-map respectively, and the similarity of the graph embedding is calculated.

[0033] The overlap degree of the target slope is obtained by statistically analyzing the proportion of sequences with the same entity type and the same relationship type in the candidate instability chain and the disaster chain of historical cases.

[0034] The consistency of the cold and arid environment zoning is obtained by setting the values ​​of consistent zoning codes to one and the values ​​of inconsistent zoning codes to zero.

[0035] The similarity of treatment effects was calculated based on the prevention and control measures adopted in historical cases and the changes in monitoring indicators after treatment.

[0036] The similarity of the map embedding, the overlap of the links, the consistency of the cold and arid environment zoning, and the similarity of the treatment effect are weighted and summed. The top m historical cases are selected as similar historical cases, where m is a preset positive integer. The matching correction weight of the measures and the priority correction coefficient are generated from the similar historical cases.

[0037] Furthermore, the measure matching correction weight and the priority correction coefficient are generated in the following manner: the main controlling factors, disaster chain, prevention and control measures already adopted, and treatment effect levels are read from the similar historical cases;

[0038] When the main controlling factors and disaster chains of historical cases are the same as those of the target slope and the treatment effect level reaches the preset effective level, the basic matching weight of the corresponding prevention and control measures will be multiplied by the first correction coefficient, which is a preset value greater than one.

[0039] When the treatment effect level does not reach the preset effective level, the basic matching weight of the corresponding prevention and control measures is multiplied by a second correction coefficient, where the second correction coefficient is a preset value that is greater than zero and less than one.

[0040] The priority correction coefficient is obtained by searching the priority coefficient mapping table based on the weighted results of the map embedding similarity, the link overlap, the consistency of the cold and arid environment zoning, and the treatment effect similarity.

[0041] Optionally, S5 includes:

[0042] The exposure of the project is calculated based on highway traffic volume, project type, distance to surrounding facilities, and construction disturbance status.

[0043] The quantitative value of the disaster consequences is calculated based on the slope size, potential disaster type, protected object category, and impact range.

[0044] The confidence level of the data is calculated based on the reliability of the data source, the time interval between data collection, and the completeness rate of the monitoring items.

[0045] The chain activation vector, the engineering exposure, the disaster consequence quantification value, the data confidence level, and the priority correction coefficient are input into the risk scoring mapping table to obtain the risk score, and the first-level risk, second-level risk, or third-level risk is output according to the preset risk score range.

[0046] Input the contribution of the main controlling factor, the confidence of the link, and the matching and correction weight of the measure into the measure mapping table, and output at least one of the following as recommended prevention and control measures: drainage and water interception, windbreak and sand fixation, anti-salt swelling and water isolation, slope toe protection, anchoring and support, excavation sequence adjustment and key inspection.

[0047] Furthermore, the priority treatment order is generated in the following manner: risk level, risk score, engineering exposure, disaster consequence quantification value, data confidence level, and recommended prevention and control measures are generated for multiple target slopes respectively;

[0048] First, sort them by risk level from Level 3, Level 2 to Level 1, then sort them in descending order by risk score, engineering exposure, and quantified disaster consequences.

[0049] When the data confidence level of any target slope is less than the preset confidence threshold, the recommended prevention and control measures for that target slope will be added to the key inspection, and the priority treatment order of that target slope will be adjusted to that of target slopes within the same risk level that have met the preset confidence threshold.

[0050] After treatment is completed, the post-treatment monitoring indicators and inspection and verification conclusions are written back to the heterogeneous spatiotemporal knowledge graph to update the treatment effect attributes of the corresponding prevention and control measures.

[0051] The beneficial effects of this invention are:

[0052] 1. By unifying slope units, stratigraphic lithology, slope structure, wind erosion status, freeze-thaw status, snowmelt water supply, salt swelling status, excavation disturbance, seismic disturbance, monitoring indicators, inspection conclusions, disaster morphology, prevention and control measures, and treatment effects into entities, relationships, and attributes in a heterogeneous spatiotemporal knowledge graph, a unified spatiotemporal index and relational expression can be established for data from different sources, reducing the impact of data fragmentation on chain identification in the risk investigation of cold and drought slopes.

[0053] 2. By constructing event time window nodes such as freeze-thaw period, snow melting period, strong wind period, salt crystallization period, excavation period, and post-earthquake period, and combining relationship type, time window activation, spatial adjacency, and data reliability in the spatiotemporal graph transformation coding model to generate path constraint attention weights, it can output the chain activation vector, main controlling factor contribution, and link confidence of candidate unstable chains, making the identification of unstable chains under complex cold and drought conditions more targeted.

[0054] 3. By retrieving similar historical cases based on map embedding similarity, link overlap, consistency of cold and drought environment zoning, and treatment effect similarity, and generating measure matching correction weights and priority correction coefficients, historical treatment effects can be used to correct the matching of prevention and control measures and the priority of governance, thereby improving the rationality of the output of treatment order for key road sections and key projects. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 This is a flowchart of a knowledge graph-based method for risk prevention and control of cold and arid slopes in Xinjiang.

[0057] Figure 2 This is a flowchart of step S3 of the present invention, which generates path constraint attention weights based on path templates. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0059] refer to Figures 1-2A knowledge graph-based method for risk prevention and control of cold and arid slopes in Xinjiang includes: S1, converting slope disaster records, monitoring data, inspection conclusions, and treatment data into entities, relationships, and attributes with temporal and spatial attributes to construct a heterogeneous spatiotemporal knowledge graph for cold and arid slope disasters; S2, mapping environmental data, monitoring data, and historical disaster chain data of the target slope to the heterogeneous spatiotemporal knowledge graph to generate environmental similarity vectors, monitoring anomaly vectors, and disaster chain path vectors; S3, constructing time window nodes for cold and arid inducing events, calculating the activation coefficients of each event time window node, and inputting the environmental similarity vectors, monitoring anomaly vectors, disaster chain path vectors, and activation coefficients into the spatiotemporal graph transformation algorithm. The code model generates path constraint attention weights based on relation type, time window activation, spatial adjacency, and data reliability, and outputs the chain activation vector, the contribution of the main controlling factor, and the link confidence of the candidate unstable chain; S4, similar historical cases are retrieved based on graph embedding similarity, link overlap, consistency of cold and drought environment zoning, and treatment effect similarity to obtain measure matching correction weights and priority correction coefficients; S5, risk level and governance priority are calculated based on the chain activation vector, engineering exposure, disaster consequence quantification value, data confidence, and the priority correction coefficient, and recommended prevention and control measures and priority treatment order are generated based on the contribution of the main controlling factor, the link confidence, and the measure matching correction weights.

[0060] In this specific embodiment, S1 includes:

[0061] The process is executed in the following order: data access, object extraction, spatiotemporal encoding, missing data marking, and graph writing.

[0062] The data sources are accessed as four types: slope disaster records, monitoring data, inspection conclusions, and treatment data. The data includes slope number, project name, coordinates, mileage station, occurrence time, monitoring time, inspection time, treatment time, disease description, monitoring index value, treatment measure name, and post-treatment effect description. The coordinates are uniformly converted to the National Geodetic Coordinate System, the mileage station is uniformly converted to a linear mileage position composed of route number, kilometer value, and meter value, and the monitoring time, inspection time, and treatment time are uniformly converted to the year-month-day-hour-minute-second format.

[0063] The slope unit, stratigraphic lithology, slope structure, wind erosion status, freeze-thaw status, snowmelt water supply, salt swelling status, excavation disturbance, seismic disturbance, monitoring indicators, inspection conclusions, disaster morphology, prevention and control measures, and treatment effects are treated as entity types. Each entity is assigned an entity type, entity name, numerical attribute, unit attribute, data source identifier, and collection time attribute. The numerical attributes of the monitoring indicator entity include displacement rate, crack opening, groundwater level, slope temperature, and water content. The numerical attributes of the inspection conclusion entity include disease type and disease level. The numerical attributes of the treatment effect entity include the change in monitoring indicators after treatment and the inspection review conclusion.

[0064] The relationship types are: location, adjacent, induced, evolved, adopted measures, and formed effect. The location relationship is used to connect slope units with strata lithology or engineering location. The adjacent relationship is used to connect slope units that are spatially continuous or whose distance does not exceed the set adjacency distance. The induced relationship is used to connect freeze-thaw state, snowmelt water supply, wind erosion state, salt expansion state, excavation disturbance or seismic disturbance with the corresponding disaster form. The evolved relationship is used to connect disaster process nodes such as cracks, rockfalls, sliding, collapses and landslides. The adopted measures relationship is used to connect disaster form or slope unit with prevention and control measures. The formed effect relationship is used to connect prevention and control measures with treatment effect.

[0065] A spatiotemporal index key is generated for each entity and each relation, and the spatiotemporal index key is calculated according to the formula. Generate, where, Representation Object The spacetime index key, This represents the object to be encoded, and its value can be either an entity object or a relation object. Represents the hash coding function. This represents a string concatenation operation. Representation Object Spatial location coding, which is determined by a unified coordinate grid coding or route mileage coding. Representation Object The time code is determined by the valid time field from the occurrence time, monitoring time, inspection time, or treatment time. Representation Object The type code corresponds to the entity type or relation type. Representation Object Data source identification code;

[0066] For missing content in the coordinate field, mileage station field, monitoring time field, patrol time field, and treatment time field, instead of deleting the corresponding record with null values, write the unobserved attribute in the corresponding attribute position and record the data source identifier, field name, and entry time corresponding to the unobserved attribute at the same time, so that subsequent steps can distinguish between the two states of the actual value being zero and the data not being observed.

[0067] Finally, the encoded entities, relationships, and attributes are written into the heterogeneous spatiotemporal knowledge graph of cold and drought slope disasters. The heterogeneous spatiotemporal knowledge graph uses entities as nodes, relationships as edges, and spatiotemporal index keys as retrieval keys. It retains the time attributes, spatial attributes, data source identifiers, and unobserved attributes of each node and each edge, so that disaster records, monitoring data, inspection conclusions, and treatment data from different sources can form a queryable, traceable, and subsequently available data foundation for vector generation and chain identification under the same spatiotemporal index.

[0068] In this specific embodiment, S2 includes:

[0069] Using the slope number and spatiotemporal index key of the target slope as the retrieval entry point, environmental data, monitoring data, and historical disaster chain data connected to the target slope are retrieved from the heterogeneous spatiotemporal knowledge graph. The environmental data includes cold and arid environment zoning, slope aspect, slope angle, altitude, stratigraphic lithology, structural plane occurrence, wind erosion intensity, freeze-thaw cycle number, snowmelt water replenishment, and salinity. The monitoring data includes displacement rate, crack opening, groundwater level, slope temperature, water content, and inspection disease categories. The historical disaster chain data includes inducing entities, process relationships, and disaster morphology.

[0070] The environmental similarity vector is generated by converting the cold and arid environment zoning, stratigraphic lithology and inspection disease categories into unique thermal codes according to the entity type coding table in step S1; converting the slope aspect angle value calculated from due north into a normalized angle code; and normalizing the slope angle, altitude, structural plane attitude, wind erosion intensity, freeze-thaw cycle number, snowmelt water supply and salinity according to the maximum and minimum values ​​of the same type of field in the heterogeneous spatiotemporal knowledge graph. The vector is then spliced ​​together in a fixed order of cold and arid environment zoning, slope aspect, slope angle, altitude, stratigraphic lithology, structural plane attitude, wind erosion intensity, freeze-thaw cycle number, snowmelt water supply and salinity to obtain the environmental similarity vector.

[0071] The monitoring anomaly vector is generated by comparing displacement rate, crack opening, groundwater level, slope temperature, and water content with their corresponding monitoring baseline intervals in a heterogeneous spatiotemporal knowledge graph. The monitoring baseline intervals are determined by historical monitoring records under stable conditions within the same cold-arid environment zone, the same stratum lithology, and the same slope structure. The types of diseases inspected are converted into numerical levels according to a disease level table. The magnitude of exceeding the boundary of the monitoring baseline interval and the duration of continuous exceedances are both written into the monitoring anomaly vector. The abnormal components of each monitoring indicator are calculated according to the formula:

[0072] ;

[0073] Indicates the target slope The Abnormal components of each monitoring indicator Indicates the target slope. Indicates the serial number of the monitoring indicator. Indicates the target slope The The measured values ​​of each monitoring indicator Indicates the first Each monitoring indicator corresponds to the upper boundary of the monitoring benchmark interval. Indicates the first Each monitoring indicator corresponds to the lower boundary of the monitoring benchmark interval. This indicates the operation of finding the maximum value. Indicates the target slope The The duration for which a monitoring indicator continuously exceeds the boundary of the monitoring baseline interval. Indicates the first The length of the statistical time window for each monitoring indicator. To represent multiplication, when lie in and The abnormal component is zero between these intervals;

[0074] The process of generating disaster chain path vectors involves extracting historical disaster chains from a heterogeneous spatiotemporal knowledge graph that are located, adjacent, induced, evolved, or have been affected by measures or effects on the target slope. Each historical disaster chain is arranged in the order of appearance of the inducing entity, process relationship, and disaster form. The entity type code of the inducing entity, the relationship type code of the process relationship, and the entity type code of the disaster form are sequentially written into the path sequence. If the path sequence length is less than the fixed length, it is supplemented with invalid path markers. If the path sequence length exceeds the fixed length, the main chain nodes and relationships from the initial inducing entity to the terminal disaster form are retained. The main chain is determined according to the relationship time sequence and the link confidence from high to low.

[0075] The environmental similarity vector, monitoring anomaly vector, and disaster chain path vector are respectively bound to the spatiotemporal index key, data source identifier, and generation time of the target slope, and used as the model input of the spatiotemporal map transformation coding model in step S3, so that the cold and drought environmental conditions, real-time anomaly status, and historical disaster evolution paths of the target slope can be continuously called in the same map semantic space.

[0076] In this specific embodiment, S3 includes:

[0077] In the heterogeneous spatiotemporal knowledge graph of cold and drought slope disasters, six types of cold and drought inducing events are identified as time window nodes: freeze-thaw period, snowmelt period, strong wind period, salt crystallization period, excavation period, and post-earthquake period. These event time window nodes are graph nodes with start and end times, spatial coverage, inducing event type, and inducing event activation coefficient. The freeze-thaw period is defined as the period within which the daily maximum temperature exceeds a certain value for seven consecutive days. And the daily minimum temperature is less than The number of times the temperature crosses zero is determined, and the snowmelt period is based on the decrease in snow depth over seven consecutive days and the average daily temperature greater than zero. The duration of strong winds is determined by the wind speed not being less than 72 consecutive hours. The cumulative duration and cumulative wind speed are determined; the salt crystallization period is determined according to the salt crystallization saturation during the process of the slope moisture content decrease; the excavation period is determined according to the excavation advance and the spatial distance between the excavation face and the slope unit in the construction log; and the post-earthquake period is determined according to the peak ground acceleration and epicentral distance within 72 hours after the earthquake.

[0078] The activation coefficients of the triggers for each event window node are all normalized to The interval is defined as follows: the number of times the temperature crosses zero, the amount of snow melt, the cumulative wind speed, the salt crystallization saturation, the excavation advance, and the peak ground acceleration are compared with the corresponding upper limit values ​​of the inducing factors. When the upper limit value is exceeded, the activation coefficient of the inducing factor is set to 1; when the upper limit value is not exceeded, it is determined according to the ratio of the measured value to the upper limit value. In this specific embodiment, the upper limit values ​​are 7 times the number of times the temperature crosses zero in seven days, 50 mm of snow melt in seven days, and 72-hour cumulative wind speed. Salt crystallization saturation 1, seven-day excavation advance 10m and seismic peak acceleration 0.20g;

[0079] The activation coefficients of each inducement are written into the time window attributes of entities and relationships that intersect or are adjacent to the target slope space. Spatial intersection means that the spatial coverage of the event time window node includes the coordinates or mileage station of the slope unit. Spatial adjacency means that the linear mileage distance between two slope units is no more than 500m or the planar coordinate distance is no more than 300m.

[0080] A path template is constructed, which is derived from historical disaster chains and expert rules. Each path template includes a starting event window node, a sequence of causal relationships, a sequence of intermediate entity types, a terminal disaster form, and an allowable time interval. The starting event window node corresponds to the freeze-thaw period, snowmelt period, strong wind period, salt crystallization period, excavation period, or post-earthquake period. The sequence of causal relationships is arranged chronologically by the induced relationship and the evolution relationship. The sequence of intermediate entity types consists of entity types in the following categories: wind erosion state, freeze-thaw state, snowmelt water supply, salt expansion state, excavation disturbance, earthquake disturbance, monitoring indicators, and inspection conclusions. The terminal disaster form is crack expansion, rockfall, sliding, collapse, or landslide. The allowable time interval is the maximum interval between the end time of the starting event window node and the occurrence time of the terminal disaster form. The allowable time interval for the freeze-thaw period and snowmelt period is 30 days, the allowable time interval for the strong wind period is 7 days, the allowable time interval for the salt crystallization period is 60 days, the allowable time interval for the excavation period is 90 days after the start of construction, and the allowable time interval for the post-earthquake period is 7 days.

[0081] When the entity type, relation type, event window, and time interval of a candidate propagation edge all fall into any path template, the template mask of the candidate propagation edge is set to 1. When the candidate propagation edge does not fall into any path template, the template mask of the candidate propagation edge is set to 0. Candidate propagation edges with a template mask of 0 do not participate in attention propagation.

[0082] The spatiotemporal graph transformation coding model consists of an input embedding layer, a three-layer path-constrained attention coding layer, and an output layer. The input embedding layer maps environmental similarity vectors, monitored anomaly vectors, disaster chain path vectors, relationship type embeddings, event window node embeddings, spatial adjacency weights, and data reliability into node features of dimension 128. Each of the three path-constrained attention coding layers employs four attention heads, each with a dimension of 32 and an activation function of [function name missing]. The dropout rate is 0.1. The output layer includes a candidate unstable chain classification branch, a chain activation regression branch, a main control factor contribution ranking branch, and a link confidence calibration branch.

[0083] In each path-constrained attention encoding layer, the path-constrained attention weights are calculated according to the formula:

[0084] ;

[0085] in, Represents a node For candidate neighboring nodes Path-constrained attention weights Indicates the current propagation node. Indicates candidate adjacent nodes, Represents a node with candidate neighboring nodes Template mask for candidate propagation edges between them. Represents an exponential function. This represents the transformation matrix of the query vector. Represents the key vector transformation matrix. Represents a node Node characteristics, Indicates candidate neighboring nodes Node characteristics, This indicates the transpose operation. The dimension of a single attention head is represented and its value is... Indicates to Square root The relational window fusion weight vector represents the relationship. This indicates embedding the relation type. Embedded with event window nodes To splice, This represents the vector concatenation operation. Represents a node with candidate neighboring nodes The activation coefficients of the causal factors corresponding to the candidate propagation edges between them. Represents a node with candidate neighboring nodes The spatial adjacency weight between nodes is set to 1 when the spatial distance is no more than 100m, 0.7 when the spatial distance is greater than 100m but no more than 300m, and 0.7 when the spatial distance is greater than 300m but no more than 500m. Represents a node with candidate neighboring nodes The reliability of data for candidate propagation edges is calculated as follows: 1 for manually verified data, 0.9 for automatically monitored data, 0.8 for construction log data, 0.7 for historical disaster records, and 0.5 when unobserved attributes are involved in inference. This represents the summation operation. This represents any candidate neighbor node in the normalization calculation. Represents a node The set of candidate neighboring nodes;

[0086] The model propagates node features along the disaster chain path according to the path constraint attention weights, and performs temporal pooling on the node features on the same candidate unstable chain to obtain the chain activation vector. The chain activation vector has a dimension of 128, and its value is used to characterize the intensity of activation of the candidate unstable chain under the current event window.

[0087] The contribution of the main instability factor is obtained by normalizing the cumulative value of the attention weight of each event window node in the candidate instability chain. The event window node with the highest contribution corresponds to the main instability factor of the target slope.

[0088] The link confidence is determined by the data reliability, template mask, and attention weight of each propagation edge within the candidate unstable chain. Propagation edges with a template mask of 0 correspond to a link confidence of 0, while propagation edges with a template mask of 1 participate in the link confidence calculation based on the product of the attention weight and the data reliability. The spatiotemporal graph transformation coding model is trained using a heterogeneous spatiotemporal knowledge graph from historical periods. During training, a training subgraph with disaster chain annotations, main inducing factor annotations, treatment measure annotations, and treatment effect annotations is extracted from the heterogeneous spatiotemporal knowledge graph from historical periods. The training subgraph is divided into a training set, a validation set, and a test set according to chronological order, with a ratio of [ratio missing]. and follow the steps The same method is used to generate training environment similarity vectors, training monitoring anomaly vectors, and training disaster chain path vectors. The same method is used to generate training causal activation coefficients and path template masks. The training objective is composed of a weighted average of candidate unstable chain classification loss, chain activation regression loss, main causal contribution ranking loss, and link confidence calibration loss. The weights of the four losses are as follows: With a learning rate of 0.001, a batch size of 32, and 100 training epochs, the model parameters with the lowest loss on the validation set are retained when the loss on the validation set does not decrease for ten consecutive epochs.

[0089] After training, the environmental similarity vector, monitoring anomaly vector, disaster chain path vector, and activation coefficient of the inducing factors of the target slope are input into the spatiotemporal graph transformation coding model. The model outputs the chain activation vector, the contribution degree of the main instability factor, and the link confidence of the candidate instability chain. The output results are bound to the spatiotemporal index key of the target slope and written back to the heterogeneous spatiotemporal knowledge graph of cold and drought slope disasters for step S4 to perform similar historical case retrieval and step S5 to calculate the risk level and governance priority.

[0090] In this specific embodiment, S4 includes:

[0091] Using the spatiotemporal index key of the target slope as the retrieval entry point, the target slope sub-map is extracted from the heterogeneous spatiotemporal knowledge graph of cold and drought slope disasters. The target slope sub-map includes the stratigraphy, lithology, slope structure, wind erosion state, freeze-thaw state, snowmelt water supply, salt swelling state, excavation disturbance, seismic disturbance, monitoring indicators, inspection conclusions, disaster morphology, prevention and control measures and treatment effect nodes within the target slope unit and its two-step relationship range. It also retains the relationships of location, adjacent, induced, evolved, adopted measures and formed effect.

[0092] Historical case subgraphs are extracted from the heterogeneous spatiotemporal knowledge graph of historical periods. Each historical case subgraph is centered on a slope unit that has experienced a disaster or has been treated. It includes the main controlling factors, disaster chain, prevention and control measures adopted, post-treatment monitoring indicators, and treatment effect level corresponding to the historical case.

[0093] Graph embedding is performed on the target slope sub-graph and the historical case sub-graph respectively. Graph embedding is completed by the input embedding layer and the three-layer path constraint attention encoding layer in the spatiotemporal graph transformation encoding model in step S3. The candidate unstable chain classification branch, chain activation regression branch, main control cause contribution ranking branch and link confidence calibration branch are removed. The node feature output with dimension 128 is retained. The slope unit nodes, cause entity nodes, disaster morphology nodes and prevention and control measure nodes in the same sub-graph are weighted and pooled according to the link confidence to obtain the sub-graph embedding vector with dimension 128.

[0094] The graph embedding similarity is obtained by normalizing the cosine similarity between the target slope subgraph embedding vector and the historical case subgraph embedding vector. The link overlap is determined by the overlap ratio of the same entity type sequence and the same relationship type sequence in the candidate instability chain of the target slope and the disaster chain of the historical case. The entity type sequence is arranged in the path order of the inducing entity, intermediate entity and terminal disaster form, and the relationship type sequence is arranged in the time order of the induced relationship and the evolution into the relationship. The cold and drought environment zoning consistency is determined by whether the cold and drought environment zoning codes of the target slope and the historical case are consistent. It is 1 when they are consistent and 0 when they are inconsistent.

[0095] The similarity of treatment effects is determined based on the control measures used in historical cases and the changes in monitoring indicators after treatment. The overall decrease in abnormal components of displacement rate, crack opening, groundwater level, slope temperature, and water content after treatment should not be less than [a certain percentage]. When the inspection and review conclusion is adjusted from dangerous or relatively dangerous to stable or basically stable, the treatment effect level is recorded as 2, and the overall decrease in abnormal components after treatment is greater than 0 and less than 1. When the treatment effect level is 1, the treatment effect level is 0 when the abnormal component does not decrease or continues to increase after treatment. The treatment effect similarity is determined by the correspondence between the treatment effect level of historical cases, the control measures adopted and the main controlling factors of the target slope.

[0096] For each historical case, a comprehensive similarity score is calculated using the following formula:

[0097] ;

[0098] in, Indicates the target slope Historical Cases The overall similarity score, Indicates the target slope. Indicates historical cases, This represents the similarity weight of the graph embedding and takes a value of 1. Represents the link overlap weight and takes a value of This indicates the consistency weight of the cold and arid environment zones, with a value of 0.15. This represents the weight of the similarity between treatment effects and takes a value of [value]. Indicates the target slope Historical Cases Graph embedding similarity, This indicates that the graph embeds similarity identifiers. Indicates the target slope Historical Cases Link overlap Indicates the link overlap index. Indicates the target slope Historical Cases Consistency of cold and arid environment zones Indicates the consistency of cold and arid environment zones. Indicates the target slope Historical Cases Similarity of treatment effects Indicates the similarity of treatment effects;

[0099] Sort all historical cases by their overall similarity score from highest to lowest, and select the cases with the highest similarity score. The historical case of the position serves as a similar historical case. This represents the number of similar historical cases and has a value of 5.

[0100] Then, the main controlling factors, disaster chains, adopted prevention and control measures, and treatment effect levels are read from the similar historical cases. When the main controlling factors and disaster chains of the historical cases are the same as those of the target slope and the treatment effect level reaches the preset effective level 2, the basic matching weight of the corresponding prevention and control measures is multiplied by the first correction coefficient, which is 1.25. When the treatment effect level does not reach the preset effective level 2, the basic matching weight of the corresponding prevention and control measures is multiplied by the second correction coefficient, which is 0.60. The basic matching weight is given by the measure mapping table. In the measure mapping table, the freeze-thaw state corresponds to drainage and water interception and slope toe protection, the snow melt water supply corresponds to drainage and water interception, the wind erosion state corresponds to windbreak and sand fixation, the salt expansion state corresponds to salt expansion and water isolation, the excavation disturbance corresponds to excavation sequence adjustment and anchor support, and the earthquake disturbance corresponds to key inspection, slope toe protection, and anchor support.

[0101] The corrective weight for the same prevention and control measure in multiple similar historical cases is obtained by weighting the comprehensive similarity score and calculating the corrective weight for the measure matching.

[0102] Finally, the priority correction coefficient is obtained by looking up the priority coefficient mapping table based on the comprehensive similarity score. When the comprehensive similarity score is not less than 0.80, the priority correction coefficient is 1.20; when the comprehensive similarity score is not less than 0.60 and less than 0.80, the priority correction coefficient is 1.00; and when the comprehensive similarity score is less than 0.60, the priority correction coefficient is 0.85. The similar historical case number, comprehensive similarity score, measure matching correction weight, and priority correction coefficient are bound to the spatiotemporal index key of the target slope and written back to the heterogeneous spatiotemporal knowledge graph of cold and drought slope disasters, so as to calculate the risk level, governance priority, and recommended prevention and control measures in step S5.

[0103] In this specific embodiment, S5 includes:

[0104] Read the target slope The chain activation vector, the contribution of the main controlling factor, the link confidence, the engineering attributes, the protected object attributes, the monitoring integrity attributes, and the correction results of similar historical cases;

[0105] The exposure of the project is determined by a combination of factors including highway traffic volume, project type, distance to surrounding facilities, and construction disturbance status. Highway traffic volume is assigned values ​​based on daily average traffic volume of no more than 500 vehicles, more than 500 but no more than 2000 vehicles, and more than 2000 vehicles, respectively. The project categories are assigned values ​​for general slope engineering, important highway or power transmission channel slopes, and mine or high-grade highway slopes, respectively. The distances to surrounding facilities are assigned values ​​based on three categories: greater than 500m, greater than 100m but not greater than 500m, and not greater than 100m. The construction disturbance status is assigned values ​​according to the following conditions: no construction disturbance, within 90 days after the completion of treatment, and during excavation or blasting construction. and by weight The weighted average is normalized to Engineering exposure within the interval;

[0106] The quantification of disaster consequences is determined jointly by the slope size, potential disaster type, protected object category, and impact range. The slope size is classified into small, medium, and large based on slope height, slope length, and potential unstable body volume, and assigned values ​​accordingly. Potential hazard types are assigned values ​​based on crack expansion or spalling, slippage, collapse, or landslide. The protected objects are categorized into general engineering facilities, roads or power transmission facilities, and residential areas or important public facilities, and are assigned values ​​accordingly. The scope of impact is assigned values ​​based on whether it does not affect traffic, affects a single or partial section of the road, or affects full traffic or external protected objects. 1.0, and weighted The weighted average is normalized to Quantitative values ​​of disaster consequences for the interval;

[0107] Data confidence is determined by the reliability of the data source, the data collection time interval, and the completeness of the monitoring items. The reliability of the data source follows the data reliability value. The data collection time interval is assigned values ​​based on the following criteria: no more than 24 hours from the most recent valid monitoring or inspection, more than 24 hours but no more than 7 days, more than 7 days but no more than 30 days, and more than 30 days from the current calculation time. The completeness rate of monitoring items is calculated by dividing the number of items with valid observation values ​​among displacement rate, crack opening, groundwater level, slope temperature, moisture content, and the types of defects inspected by 6, and then weighted accordingly. The weighted average is normalized to The data confidence level for the interval; the risk score mapping table in this specific implementation is given by the formula:

[0108] ;

[0109] in, Indicates the target slope Risk score, Indicates the target slope. This indicates that the operation takes the smaller value. Indicates the target slope Priority correction factor Indicates the target slope The chain activation strength is determined by the maximum activation component in the chain activation vector. Indicates the target slope Engineering exposure Indicates the target slope The quantification of disaster consequences, Indicates the target slope Data confidence level Indicates the target slope The link confidence score, where 100 represents the proportional coefficient for converting the normalized result into a percentage score.

[0110] A risk score less than 40 is output as Level 1 risk; a risk score of not less than 40 and less than 70 is output as Level 2 risk; and a risk score of not less than 70 is output as Level 3 risk.

[0111] The measure mapping table generates recommended prevention and control measures based on the contribution of the main controlling factors, the confidence of the link, and the measure matching correction weight. The freeze-thaw state and snowmelt water supply correspond to drainage and water interception, the wind erosion state corresponds to windbreak and sand fixation, the salt expansion state corresponds to salt expansion and water isolation, the excavation disturbance corresponds to excavation sequence adjustment and anchor support, the earthquake disturbance corresponds to key inspection, slope toe protection and anchor support, and when the disaster mode is collapse or landslide and the disaster consequence quantification value is not less than 0.70, the slope toe protection and anchor support are simultaneously implemented.

[0112] The recommended value of each candidate prevention and control measure is obtained by multiplying the contribution of the corresponding main inducing factor, the link confidence of the candidate instability chain, and the measure matching correction weight. Prevention and control measures with a recommended value of not less than 0.50 are output as recommended prevention and control measures. When the recommended values ​​of all candidate prevention and control measures are less than 0.50, the prevention and control measure with the highest recommended value is output, and the recommended prevention and control measures are arranged in order of priority treatment from high to low according to the recommended values.

[0113] When prioritizing the treatment of multiple target slopes, the risk level, risk score, engineering exposure, disaster consequence quantification value, data confidence level and recommended prevention and control measures are generated for each target slope. First, the slopes are sorted in order of risk level from level 3 risk, level 2 risk to level 1 risk. Then, within the same risk level, the slopes are sorted in descending order of risk score, engineering exposure and disaster consequence quantification value.

[0114] When the data confidence level of any target slope is less than the preset confidence threshold of 0.65, the key inspection will be added to the recommended prevention and control measures for that target slope, and the priority treatment order of that target slope will be adjusted to after the target slopes with a data confidence level of not less than 0.65 within the same risk level.

[0115] After treatment is completed, the displacement rate, crack opening amount, groundwater level, slope temperature, water content, inspection and verification conclusions, actual control measures adopted, treatment completion time and treatment effect level are written back into the heterogeneous spatiotemporal knowledge graph of cold and drought slope disasters. The treatment effect attributes, formation effect relationships and spatiotemporal index keys of the corresponding control measure nodes are updated so that the latest treatment results can be called for subsequent similar historical case retrieval, measure matching correction weight calculation and priority correction coefficient calculation.

[0116] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0117] This invention uses a heterogeneous spatiotemporal knowledge graph to uniformly represent disaster records, monitoring data, inspection conclusions, and treatment information. It also establishes a computable association between the target slope and historical disaster chains through environmental similarity vectors, monitoring anomaly vectors, and disaster chain path vectors. This enables the originally scattered data to serve the identification of candidate instability chains, the calculation of risk levels, and the output of recommended measures. As a result, it forms a continuous technical processing chain to address the problems of multiple causes of slope instability in cold and arid regions, complex data sources, and difficulty in determining treatment priorities.

[0118] This invention further sets up a time window node for cold and drought inducing events and a disaster chain path constraint attention mechanism in the knowledge graph reasoning layer. It transforms historical disaster chains and expert rules into path templates, and combines the treatment effect correction measures of similar historical cases with matching weights and priority correction coefficients. This allows the spatiotemporal graph transformation coding model to consider the inducing time period, spatial adjacency, relationship type and data reliability at the same time when attention propagation and result output, making it more suitable for the risk prevention and control scenario of cold and drought slopes in Xinjiang.

Claims

1. A knowledge graph-based method for risk prevention and control of cold and arid slopes in Xinjiang, characterized in that, include: S1. Convert slope disaster records, monitoring data, inspection conclusions, and treatment data into entities, relationships, and attributes with temporal and spatial attributes to construct a heterogeneous spatiotemporal knowledge graph for cold and drought slope disasters; S2. Map the environmental data, monitoring data, and historical disaster chain data of the target slope to the heterogeneous spatiotemporal knowledge graph to generate environmental similarity vectors, monitoring anomaly vectors, and disaster chain path vectors; S3. Construct cold and drought causative event time window nodes, calculate the causative activation coefficients of each event time window node, and input the environmental similarity vectors, monitoring anomaly vectors, disaster chain path vectors, and causative activation coefficients into the spatiotemporal graph transformation coding model. Generate path constraint attention weights based on relationship type, time window activation, spatial adjacency, and data reliability, and output the chain activation vector, main controlling causative contribution, and link confidence of candidate instability chains; S4. Retrieve similar historical cases based on graph embedding similarity, link overlap, cold and drought environmental zoning consistency, and treatment effect similarity to obtain measure matching correction weights and priority correction coefficients; S5. Calculate the risk level and governance priority based on the chain activation vector, engineering exposure, disaster consequence quantification value, data confidence level and the priority correction coefficient, and generate recommended prevention and control measures and priority treatment order based on the contribution of the main controlling factors, the link confidence level and the measure matching correction weight.

2. The method for risk prevention and control of cold and arid slopes in Xinjiang based on knowledge graphs according to claim 1, characterized in that, S1 includes: The entity types are slope units, stratigraphic lithology, slope structure, wind erosion status, freeze-thaw status, snowmelt water supply, salt swelling status, excavation disturbance, seismic disturbance, monitoring indicators, inspection conclusions, disaster morphology, prevention and control measures, and treatment effects; the relation types are location, adjacent, induced, evolved, adopted measures, and formed effects; and the coordinates, mileage markers, monitoring time, inspection time, and treatment time in data from different sources are uniformly coded to generate a spatiotemporal index key for each entity and relation. Missing fields are marked as unobserved attributes and data source identifiers are recorded to obtain the heterogeneous spatiotemporal knowledge graph.

3. The knowledge graph-based risk prevention and control method for cold and arid slopes in Xinjiang according to claim 1, characterized in that, S2 include: The environmental similarity vector is generated from the cold and arid environment zoning, slope aspect, slope angle, altitude, stratum lithology, structural plane occurrence, wind erosion intensity, freeze-thaw cycle number, snowmelt water supply, and salinity of the target slope. The displacement rate, crack opening, groundwater level, slope temperature, water content, and inspection disease type are compared with the corresponding monitoring benchmark interval, and the monitoring anomaly vector is generated according to the magnitude and duration of exceeding the interval boundary. The causal entities, process relationships, and disaster forms in the historical disaster chain are encoded according to the order of path occurrence to generate the disaster chain path vector. The environmental similarity vector, the monitoring anomaly vector, and the disaster chain path vector are used together as the model input for step S3.

4. The method for risk prevention and control of cold and arid slopes in Xinjiang based on knowledge graphs according to claim 1, characterized in that, S3 includes: constructing the freeze-thaw period, snowmelt period, strong wind period, salt crystallization period, excavation period, and post-earthquake period as event time window nodes; calculating the corresponding causal activation coefficients based on the number of zero-crossing temperatures, snowmelt amount, cumulative wind speed, salt crystallization saturation, excavation progress, and peak ground acceleration; writing the causal activation coefficients into the time window attributes of adjacent entities and relationships; in the spatiotemporal graph transformation coding model, multiplying and normalizing the relationship type embedding, event time window node embedding, spatial adjacency weight, and data reliability to generate the path constraint attention weight; and obtaining the chain activation vector, the main controlling causal contribution, and the link confidence based on the path constraint attention weight propagating along the disaster chain path node characteristics.

5. The method for risk prevention and control of cold and arid slopes in Xinjiang based on knowledge graphs according to claim 1, characterized in that, S4 includes: performing graph embedding on the target slope sub-map and the historical case sub-map respectively, and calculating the graph embedding similarity; calculating the proportion of the same entity type sequence and the same relationship type sequence in the candidate instability chain of the target slope and the disaster chain of the historical case to obtain the link overlap; setting the value of consistent cold and drought environment zoning coding to one and the value of inconsistent coding to zero to obtain the cold and drought environment zoning consistency; calculating the treatment effect similarity based on the prevention and control measures adopted in the historical cases and the changes in monitoring indicators after treatment; weighting and summing the graph embedding similarity, the link overlap, the cold and drought environment zoning consistency, and the treatment effect similarity, selecting the top m historical cases as similar historical cases, where m is a preset positive integer, and generating the measure matching correction weight and the priority correction coefficient from the similar historical cases.

6. The method for risk prevention and control of cold and arid slopes in Xinjiang based on knowledge graphs according to claim 1, characterized in that, S5 includes: The exposure of the project is calculated based on highway traffic volume, project type, distance to surrounding facilities, and construction disturbance status; the quantification of disaster consequences is calculated based on slope scale, potential disaster type, protected object category, and impact range; the confidence level of the data is calculated based on data source reliability, collection time interval, and monitoring item completeness rate; the chain activation vector, project exposure, quantification of disaster consequences, data confidence level, and priority correction coefficient are input into a risk scoring mapping table to obtain a risk score, and the risk is output as Level 1, Level 2, or Level 3 risk according to a preset risk score range; The contribution of the main controlling factor, the confidence of the link, and the matching and correction weight of the measure are input into the measure mapping table, and at least one of the following is output as a recommended prevention and control measure: drainage and water interception, windbreak and sand fixation, anti-salt swelling and water isolation, slope toe protection, anchoring and support, excavation sequence adjustment, and key inspection.

7. The knowledge graph-based risk prevention and control method for cold and arid slopes in Xinjiang according to claim 4, characterized in that, The path constraint attention weights are also generated based on path templates: historical disaster chains and expert rules are converted into path templates, each path template including a starting event window node, a sequence of causal relationships, a sequence of intermediate entity types, a terminal disaster form, and an allowed time interval; when the entity type, relationship type, and event window of a candidate propagation edge fall into any path template, the attention propagation coefficient of that candidate propagation edge is retained; when a candidate propagation edge does not fall into any path template, the attention propagation coefficient of that candidate propagation edge is set to zero. The path-constrained attention weights are obtained by normalizing the retained attention propagation coefficients.

8. The knowledge graph-based risk prevention and control method for cold and arid slopes in Xinjiang according to claim 5, characterized in that, The matching correction weight and the priority correction coefficient are generated as follows: the main controlling factors, disaster chains, adopted prevention and control measures, and treatment effect levels are read from the similar historical cases; when the main controlling factors and disaster chains of the historical cases are the same as those of the target slope and the treatment effect level reaches the preset effective level, the basic matching weight of the corresponding prevention and control measures is multiplied by the first correction coefficient, where the first correction coefficient is a preset value greater than one. When the treatment effect level does not reach the preset effective level, the basic matching weight of the corresponding prevention and control measures is multiplied by a second correction coefficient, where the second correction coefficient is a preset value that is greater than zero and less than one. The priority correction coefficient is obtained by searching the priority coefficient mapping table based on the weighted results of the map embedding similarity, the link overlap, the consistency of the cold and arid environment zoning, and the treatment effect similarity.

9. The knowledge graph-based risk prevention and control method for cold and arid slopes in Xinjiang according to claim 4, characterized in that, The spatiotemporal graph transformation coding model is trained in the following way: a training subgraph with disaster chain annotation, main inducing factor annotation, treatment measure annotation and treatment effect annotation is extracted from the heterogeneous spatiotemporal knowledge graph of historical periods; The training environment similarity vector, training monitoring anomaly vector, and training disaster chain path vector are generated in the same manner as in step S2, and the training induced activation coefficients are generated in the same manner as in step S3. The weighted sum of the candidate unstable chain classification loss, chain activation regression loss, main induced factor contribution ranking loss, and link confidence calibration loss is used as the training objective to obtain the spatiotemporal graph transformation coding model.

10. The knowledge graph-based risk prevention and control method for cold and arid slopes in Xinjiang according to claim 6, characterized in that, The priority treatment order is generated as follows: risk level, risk score, engineering exposure, disaster consequence quantification value, data confidence level, and recommended prevention and control measures are generated for multiple target slopes respectively; first, they are sorted in order of risk level from level 3 risk, level 2 risk to level 1 risk, and then sorted in descending order of risk score, engineering exposure, and disaster consequence quantification value; when the data confidence level of any target slope is less than the preset confidence threshold, the recommended prevention and control measures for that target slope will be added to the key inspection, and the priority treatment order of that target slope will be adjusted to after the target slopes within the same risk level that have met the preset confidence threshold; After treatment is completed, the post-treatment monitoring indicators and inspection and verification conclusions are written back to the heterogeneous spatiotemporal knowledge graph to update the treatment effect attributes of the corresponding prevention and control measures.

Images