An intention-driven railway slope intelligent monitoring and instability early warning method and system

Abstract

The present application relates to the technical field of railway slope safety monitoring, and proposes an intention-driven railway slope intelligent monitoring and instability early warning method and system. The method uses an intention analysis model to extract the intention entity of the natural language disaster prevention intention and encode it, and performs feature matching with the knowledge node vector in the pre-constructed knowledge graph in the field of railway geological disasters to generate a structured intention strategy. A multi-objective optimization model is constructed based on the structured intention strategy to solve and generate an optimal device scheduling scheme. Each multi-source heterogeneous monitoring device performs a cooperative perception task according to the device scheduling scheme, performs spatio-temporal registration and feature fusion, and outputs the instability evaluation result of the current railway slope. The intention achievement degree is verified based on the instability evaluation result and the structured intention strategy. The present application realizes the automatic mapping of disaster prevention business requirements and monitoring devices, breaks through the energy consumption bottleneck of field devices, and significantly improves the timeliness and robustness of slope disaster early warning.

Description

Technical Field

[0001] This invention relates to the fields of data processing and geological disaster monitoring technology, and in particular to an intent-driven intelligent monitoring and instability early warning method and system for railway slopes. Background Technology

[0002] In the construction and operation of railway projects under complex terrain conditions, slope stability is directly related to the safety of train operation and the lives and property of people along the line. Especially under the influence of factors such as extreme rainfall, geological tectonic activity, and changes in groundwater, slope instability (such as landslides, collapses, and debris flows) is often characterized by its suddenness, concealment, and high destructiveness. To ensure railway operation safety, a large number of multi-source heterogeneous IoT monitoring devices are currently deployed along railway lines. These devices mainly include rain gauges, surface displacement monitoring equipment (such as GNSS or Beidou monitoring stations), deep inclinometers, crack gauges, video surveillance equipment, and inspection drones, used to continuously acquire information on the physical state of slopes.

[0003] Currently, the publicly available technologies related to railway slope monitoring mainly include the following categories: The first is the online fixed sampling and monitoring system, which mainly relies on the equipment's preset sampling frequency to continuously collect slope monitoring data; Second, a threshold-triggered early warning system that triggers an alarm when the data monitored by a single sensor exceeds a preset safety threshold. Third, risk prediction systems based on data-driven or machine learning methods use historical monitoring data to train models and predict and analyze slope stability.

[0004] However, under complex and ever-changing disaster prevention scenarios and high-concurrency scheduling requirements, the aforementioned existing technologies still have certain limitations, mainly in the following aspects: First, it does not support intent-driven operations, leading to difficulties in semantic mapping between disaster prevention business needs and underlying monitoring equipment, resulting in low scheduling efficiency. Existing systems typically cannot automatically parse disaster prevention business needs in natural language form. When encountering sudden severe weather or emergencies, railway disaster prevention dispatchers often need to propose monitoring requirements based on experience, such as conducting high-frequency inspections of a high-risk slope. However, existing systems still require technicians to manually configure sensor parameters and modify the sampling frequency or monitoring strategy of each device, a process that is time-consuming and prone to configuration omissions.

[0005] Second, it lacks multimodal collaboration capabilities and has insufficient coordination among heterogeneous monitoring devices from multiple sources. Currently, in railway slope monitoring networks, surface sensors, deep monitoring equipment, and inspection drones are often deployed on different system platforms, resulting in a certain degree of information silos. When data quality deteriorates due to communication anomalies or hardware failures in a certain area, traditional systems can usually only passively trigger alarms and cannot automatically schedule other devices to conduct collaborative observations based on the monitoring target, lacking multimodal linkage and dynamic blind spot filling capabilities.

[0006] Third, the lack of a dynamic resource orchestration mechanism and the static monitoring strategy result in low system resource utilization efficiency. Traditional slope monitoring systems typically employ fixed sampling or low-power periodic wake-up modes. When the equipment maintains high-frequency acquisition for extended periods, it is prone to generating a large amount of redundant data and consuming communication bandwidth, while also accelerating the energy consumption of battery-powered equipment in the field. Conversely, when the sampling frequency is low, it may be difficult to capture early, minute deformation characteristics of the slope in a timely manner, affecting the early warning effect.

[0007] Therefore, there is an urgent need for a railway slope monitoring method that can automatically understand the disaster prevention business intent and accordingly conduct intelligent collaborative scheduling and dynamic resource arrangement of multi-source heterogeneous monitoring equipment, so as to improve the railway slope disaster monitoring and early warning capabilities. Summary of the Invention

[0008] To address the technical problems in existing railway slope monitoring systems, such as the disconnect between disaster prevention business needs and the semantics of underlying physical equipment, the lack of intelligent collaborative perception among heterogeneous equipment, and the inefficient resource allocation and excessive energy consumption of nodes caused by static monitoring strategies, this invention provides an intent-driven intelligent monitoring and instability early warning method and system for railway slopes.

[0009] This invention aims to achieve accurate analysis of disaster prevention intentions through vectorized alignment of large language models and knowledge graphs, and to construct a multi-objective optimized orchestration algorithm to drive multi-source heterogeneous monitoring devices to perform collaborative sensing. Under the premise of strictly meeting disaster prevention constraints, it maximizes the system monitoring efficiency and minimizes node energy consumption, thereby realizing dynamic decision-making and closed-loop early warning for slope instability.

[0010] To achieve the above objectives, this invention provides an intent-driven intelligent monitoring and instability early warning method for railway slopes, the method comprising the following steps: Step S1: Receive natural language disaster prevention intentions input by railway disaster prevention personnel, extract intention entities including slope location, risk type and monitoring needs using an intention parsing model based on pre-trained large language model and named entity recognition, encode the intention entities into intention feature vectors, perform feature matching with knowledge node vectors in a pre-constructed railway geological disaster domain knowledge graph, and generate structured intention strategies. Step S2: The physical topology and equipment resource attributes of the current multi-source heterogeneous monitoring equipment on the slope are obtained in real time through the edge computing gateway deployed along the railway line. A multi-objective optimization model is constructed based on the structured intent strategy, and the optimal equipment scheduling scheme is solved and generated through the resource orchestration algorithm. Step S3: Each multi-source heterogeneous monitoring device in the underlying physical network dynamically updates its local sampling parameters according to the device scheduling scheme and performs collaborative sensing tasks. It performs spatiotemporal registration and feature fusion on the collected multimodal monitoring data and outputs the current instability assessment result of the railway slope. Step S4: Based on the instability assessment results and the structured intent strategy, verify the intent achievement degree to determine whether the disaster prevention intent constraint is met. If the constraint is met, the process ends. If the constraint is not met, trigger the closed-loop reconstruction mechanism, dynamically update the equipment scheduling scheme, and distribute it to the underlying physical network for collaborative blind spot filling.

[0011] Specifically, in step S1, the specific steps of encoding the intent entity into an intent feature vector and performing feature matching with the knowledge graph include: The extracted intent entities are mapped to intent feature vectors using a word embedding model. The domain entities in the railway geological disaster domain knowledge graph are mapped to knowledge node vectors. ; The matching degree between the intent feature vector and the knowledge node vector is calculated using cosine similarity. The calculation formula is as follows: ; Set similarity threshold Only when The mapping relationship between the intent entity and the domain entity is preserved; otherwise, the association rules in the knowledge graph of railway geological disaster domain are used for forced correction to generate a structured intent strategy that includes target spatial coordinates, target sensor type and performance constraint parameters; wherein, the performance constraint parameters include maximum tolerable delay and minimum sampling frequency.

[0012] Specifically, in step S2, the step of constructing a multi-objective optimization model based on the structured intent strategy includes: The generation of equipment scheduling schemes is transformed into a constrained multi-objective optimization problem, and decision variables are set. ,in Indicates the task Distributed to multi-source heterogeneous monitoring equipment , This indicates no allocation; Define the optimization objective function To maximize the overall monitoring service quality and minimize the total energy consumption of field equipment, the calculation formula is as follows: ; in, This is a set of sub-tasks derived from the current disaster prevention objectives. A collection of multi-source heterogeneous monitoring devices available within the target area; Multi-source heterogeneous monitoring equipment Execute the task Expected data quality gain at that time; Multi-source heterogeneous monitoring equipment Energy consumption for waking up from hibernation to perform high-frequency sensing and transmission tasks; and These are the weighting coefficients for service quality and energy consumption, respectively.

[0013] Specifically, the optimization objective function must satisfy the physical constraints of the underlying device: ; ; in, Multi-source heterogeneous monitoring equipment Execute the task Required bandwidth This represents the maximum available bandwidth for the node. Multi-source heterogeneous monitoring equipment Current remaining battery power, Minimum battery level for safe sleep mode; After solving the optimization objective function using a heuristic simulated annealing algorithm or a multi-agent reinforcement learning algorithm, a method is generated that... The largest equipment scheduling scheme is implemented, and a control command matrix is ​​issued.

[0014] Specifically, in step S3, the feature fusion method of the multimodal monitoring data is as follows: The weighted feature fusion algorithm is used to calculate the overall probability of slope instability. : ; in, To sense the total number of modes; For the first Observational data of various modes after spatiotemporal registration; This represents the local instability probability based on single-mode data output. For dynamic confidence weights; When the edge computing gateway detects a surge in environmental rainfall that obstructs the view from the visual sensors, it dynamically reduces the weight of the visual modality and correspondingly increases the weight of the microwave radar or the underground inclinometer.

[0015] Specifically, in step S4, the method for verifying intent achievement is as follows: Real-time calculation of information entropy and data continuity indicators of instability assessment results; Compare the current monitoring service quality Quality requirements in the structured intent strategy ; When the condition is met If this is deemed an intent violation, an abnormal interruption command is sent to the resource orchestration module. Upon receiving an abnormal interruption command, the multi-objective optimization solution is re-executed to find a suboptimal alternative path. Commands are issued to actively wake up nearby dormant PTZ cameras, or to dispatch inspection drones to approach the target slope section. The image change detection algorithm is used to extract the crack width features on the slope surface. The supplementary visual features are then sent to the feature fusion stage in step S3 for multimodal secondary fusion until the final output comprehensive instability probability meets the disaster prevention intention constraint, thus forming an adaptive closed loop.

[0016] This invention also provides an intent-driven intelligent monitoring and instability early warning system for railway slopes, comprising: The intent parsing module, deployed in the application service layer, includes a pre-trained large language model and graph database. It is used to extract and encode the intent entities of natural language disaster prevention intents, and perform feature matching with knowledge node vectors to generate structured intent strategies. The resource orchestration module is used to receive the physical topology and equipment resource attributes of the multi-source heterogeneous monitoring equipment for slopes uploaded by the structured intent policy and the edge computing gateway, and run a multi-objective optimization model to solve the optimal equipment scheduling scheme. The collaborative sensing module includes multi-source heterogeneous monitoring devices and edge computing gateways distributed on site. It is used to perform collaborative sensing tasks and perform spatiotemporal registration and feature fusion on the collected multimodal monitoring data, and output the current instability assessment results of the railway slope. The dynamic evaluation module is used to calculate the deviation between the instability evaluation result and the structured intent strategy, verify the degree of intent achievement, and trigger the closed-loop reconstruction mechanism.

[0017] Furthermore, the internal structure of the intent parsing module includes: The semantic encoding module is used to extract intention entities, including slope location, risk type and monitoring needs, from natural language disaster prevention intentions using a pre-trained large language model and named entity recognition technology, and encode the intention entities into intention feature vectors. The knowledge alignment module is used to extract knowledge node vectors from the pre-constructed railway geological disaster knowledge graph, calculate the cosine similarity between the intent feature vector and the knowledge node vector, filter logical illusions in the large language model and perform association correction based on the set similarity threshold. The constraint generation module is used to extract corresponding attribute parameters based on the deterministic mapping relationship output by the knowledge alignment module, and generate a structured intent strategy containing target spatial coordinates, target sensor type and performance constraint parameters.

[0018] The present invention also provides an electronic device, characterized in that it includes: a processor, a memory, and a communication interface; the memory is used to store computer instructions; the processor, the memory, and the communication interface complete the mutual electrical signal transmission between them through a communication bus, and when executing the computer instructions stored in the memory, it realizes the intention-driven intelligent monitoring and instability early warning method for railway slopes.

[0019] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the intent-driven intelligent monitoring and instability early warning method for railway slopes.

[0020] Compared with existing technologies, the core innovations and beneficial effects of this invention are mainly reflected in the following three aspects: (1) An intent-driven and strongly constrained mapping mechanism to prevent logical illusions in large-scale models was constructed, solving the problem of secure conversion of disaster prevention requirements into device instructions in complex contexts. Existing IoT scheduling methods based on large language models (such as conventional semantic-spatial scheduling) have the risk of logical illusions in safety-critical areas. This invention utilizes natural language models to extract key entities and construct intent feature vectors. Subsequently, instead of directly implementing scheduling, a pre-constructed railway geological disaster knowledge graph node vector is introduced. Through similarity calculation A rigorous domain entity matching process is implemented. When the matching degree falls below a threshold, a correction mechanism based on pre-defined association rules in the graph is forcibly triggered. This mechanism effectively filters out non-professional instructions from large models and generates structured intent strategies with strict performance constraints such as maximum tolerable latency and minimum sampling frequency. This improves the level of scheduling automation while safeguarding the absolute safety bottom line of railway disaster prevention and scheduling.

[0021] (2) A multi-objective physical resource orchestration method based on edge network topology awareness is proposed, breaking through the bottleneck of computing power and energy consumption of heterogeneous devices in the field. Unlike traditional centralized scheduling in the cloud or single-dimensional network optimization, this invention distributes structured intent strategies to edge computing gateways along the railway line. The gateways perceive the physical topology status of underlying multi-source heterogeneous monitoring devices (such as inclinometers, rain gauges, and cameras) in real time. This is achieved by constructing a system that combines quality gain parameters... High-frequency wake-up of devices and energy consumption costs The multi-objective optimization function, while strictly satisfying the available bandwidth of the underlying nodes. Safety sleep minimum power The solution is obtained under rigid physical constraints. This method can achieve an optimal balance between high-frequency monitoring of key high-risk areas and the endurance of global nodes in harsh railway slope environments with no mains power and weak communication.

[0022] (3) A cross-modal hardware-level collaborative blind-filling mechanism based on information entropy and intent achievement verification was designed, which significantly improved the robustness of early warning under extreme weather conditions. Existing technologies mostly rely on static multimodal data fusion algorithms. Once a single modality (such as visual failure caused by heavy rain) is blocked, the system reliability will drop sharply. This invention establishes a dynamic closed-loop self-healing mechanism. The system calculates the information entropy of the current instability assessment result in real time and evaluates the quality of monitoring services. When the monitoring quality is lower than the quality requirements of the intent strategy, it is judged as an intent violation and triggers hardware-level reconstruction. The system can not only dynamically adjust the confidence weight of feature fusion at the algorithm level, but also actively issue control commands to physically wake up nearby dormant PTZ cameras or cross-regional dispatch inspection drones for close reconnaissance. This dual closed loop, combining "dynamic adjustment of algorithm weights" and "active blind-filling of physical equipment", ensures the continuity and high confidence of railway slope disaster early warning under complex and harsh environments. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0024] Figure 1 A flowchart illustrating the steps of an intent-driven intelligent monitoring and instability early warning method for railway slopes, as provided in an embodiment of the present invention; Figure 2 This is a block diagram of the module structure of an intent-driven intelligent monitoring and instability early warning system for railway slopes, provided in an embodiment of the present invention. Figure 3This is a schematic diagram illustrating the deployment and workflow of the intelligent monitoring and instability early warning system for railway slopes driven by the intent of this invention. Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be clearly and completely described below with reference to embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Examples of embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention provides an intent-driven intelligent monitoring and instability early warning method for railway slopes. It possesses good adaptive scheduling capabilities for multi-source heterogeneous monitoring equipment and cross-modal data fusion evaluation capabilities, offering a new approach to disaster prevention scheduling in complex railway environments: "business intent directly reaches physical equipment." The flowchart of the method is shown below. Figure 1 As shown, the specific steps include: Step S1: Encode intent entities and perform feature matching to generate a structured intent strategy: Receive natural language disaster prevention intent input by railway disaster prevention personnel, extract intent entities including slope location, risk type, and monitoring requirements using an intent parsing model based on a pre-trained large language model and named entity recognition, encode the intent entities into intent feature vectors, and perform feature matching with knowledge node vectors in a pre-constructed railway geological disaster domain knowledge graph to generate a structured intent strategy.

[0027] For example, based on a specific emergency disaster prevention scenario for railway slopes, when a sudden increase in rainfall is detected, the dispatch center staff input a voice or text command: "Immediately conduct a high-frequency investigation of the deep sliding and surface seepage conditions of the steep mudstone slope in section K105." The system uses a pre-trained Large Language Model (LLM) to perform preliminary Named Entity Recognition (NER) to extract entities to be identified.

[0028] To eliminate the "logical illusion" and common-sense errors in large language models, the extracted intent entities are mapped to intent feature vectors. Furthermore, it maps the professional entities in the pre-constructed railway geological disaster knowledge graph into knowledge node vectors. The matching degree between the intent feature vector and the knowledge node vector is calculated using cosine similarity. The calculation formula is as follows: ; in, for Dimensional intention feature vector; for 3D knowledge graph node vectors; and | represent the L2 norm of the vectors. A similarity threshold is set. (like ), only when If the alignment is successful, the entity mapping relationship is preserved; otherwise, forced correction is performed based on the strong association rules in the knowledge graph. After alignment, a machine-readable structured intent strategy is output, including: target spatial coordinates (3D bounding box of section K105), target sensor type (deep inclinometer, groundwater level gauge), and performance constraint parameters (maximum tolerable latency). minimum sampling frequency ).

[0029] Step S2: Construct a multi-objective optimization model to generate the optimal equipment scheduling scheme: The physical topology status and equipment resource attributes of the current slope multi-source heterogeneous monitoring equipment are obtained in real time by the edge computing gateway deployed along the railway line. Based on the structured intent strategy, a multi-objective optimization model is constructed, and the optimal equipment scheduling scheme is solved and generated by the resource orchestration algorithm.

[0030] The generation of the equipment scheduling scheme is transformed into a constrained multi-objective optimization problem. Decision variables are defined. ,in Indicates the task Assigned to physical monitoring equipment ; This indicates no allocation. Define the system's optimization objective function. To maximize overall Quality of Service (QoS) and minimize total energy consumption of field equipment, the calculation formula is as follows: ; in, This is a set of sub-tasks derived from the current disaster prevention objectives. A collection of multi-source heterogeneous monitoring devices available within the target area (such as the K105 section) (including rain gauges, surface displacement gauges, deep inclinometers, drones, and cameras). For equipment Execute the task The expected data quality gain at that time, which is determined by the device's historical state and sensor characteristics; For equipment Energy consumption for waking up from hibernation to perform high-frequency sensing and transmission tasks; and These are the weighting coefficients for service quality and energy consumption, respectively. Under the intent of "emergency disaster prevention," the system adaptively adjusts these coefficients. To ensure high-precision monitoring.

[0031] At the same time, the objective function must satisfy the physical constraints of the underlying device: ; ; in, Bandwidth required for the task. This represents the maximum available bandwidth for the node. This represents the device's current remaining battery power. To ensure safe sleep, a minimum power level is maintained. As a preferred implementation, when using the Multi-Agent Reinforcement Learning (MARL) algorithm, the edge computing gateway in the system is considered as an agent, and the available status and task queues of all current monitoring devices are used as the environment state space; the agent's action space is defined as the generation decision variable matrix. The allocation combination; the reward function is defined as the optimization objective function. The value of the penalty term is combined with the constraints. The agent, through continuous interaction with the environment, is trained to generate... The optimal equipment scheduling scheme is determined, and a control command matrix is ​​issued.

[0032] Step S3: Multimodal monitoring data acquisition and instability assessment: Each multi-source heterogeneous monitoring device in the underlying physical network dynamically updates its local sampling parameters according to the device scheduling scheme and performs collaborative sensing tasks. The acquired multimodal monitoring data is spatiotemporally registered and feature fused to output the current instability assessment result of the railway slope.

[0033] After receiving the scheduling command, the bottom-level deep inclinometer node modifies its internal timer register, increasing the wake-up frequency from a low-power "1 time / day" to "1 time / 10 minutes". The edge computing gateway receives one-dimensional time-series data from rain gauges and displacement gauges, as well as two-dimensional image streams from cameras.

[0034] The weighted feature fusion algorithm is used to calculate the overall probability of slope instability. : ; in, To sense the total number of modes; For the first Observational data of various modes after spatiotemporal registration; This refers to the local instability probability output by a pre-set mechanical or machine learning model based on single-modal data (such as radar displacement data alone); For dynamic confidence weights ( If a sudden increase in ambient rainfall is detected, causing the visual camera image to become blurry or obstructed, the system will automatically reduce the visual modality weight. And accordingly increase the weight of microwave radar or underground inclinometers. This ensures high confidence in the fusion results under complex environments.

[0035] Step S4, Intent Achievement Verification: Based on the instability assessment results and the structured intent strategy, the intent achievement is verified to determine whether the disaster prevention intent constraint is met. If the constraint is met, the process ends. If the constraint is not met, the closed-loop reconstruction mechanism is triggered to dynamically update the equipment scheduling scheme and distribute it to the underlying physical network for collaborative blind spot filling.

[0036] The system calculates the information entropy and data continuity index of the current evaluation output in real time. The information entropy... The calculation is used to assess the uncertainty of current monitoring results, and its formula is: ; in This represents the probability distribution of the data returned by each sensor in each state. When the information entropy... A sharp increase exceeding a set threshold indicates that the current sensing data is chaotic, and the system uses this as one of the bases to determine the current monitoring service quality. A sharp drop. The system compares the current monitoring service quality. With the structural intent of stage S1 .

[0037] When the condition is met When this occurs, the system determines it as an "Intent Violation." At this point, the dynamic evaluation module sends an exception interrupt command to the resource orchestration module, such as... Figure 3 As shown, the system automatically re-executes step S2 to find a suboptimal alternative path: it issues a command to actively wake up nearby dormant high-definition PTZ cameras, or dispatches the nearest inspection drone to approach section K105, and uses AI image change detection algorithms (such as frame difference method or deep learning segmentation network) to extract the width features of cracks on the slope surface. The supplemented visual features are then fed back into step S3 for multimodal secondary fusion until the final output is obtained. It can accurately respond to the initial "high-frequency investigation" intentions of disaster prevention personnel, forming a self-healing closed loop without human intervention.

[0038] Furthermore, this invention also provides an intent-driven intelligent monitoring and instability early warning system for railway slopes, such as... Figure 2As shown, the system consists of four core modules that communicate with each other: an intent parsing module, a resource orchestration module, a collaborative perception module, and a dynamic evaluation module. The intent parsing module, deployed at the application service layer, includes a pre-trained large language model and a graph database. It is used to extract and encode the intent entities of natural language disaster prevention intents and perform feature matching with knowledge node vectors to generate structured intent strategies. The resource orchestration module receives the structured intent strategies and the physical topology and equipment resource attributes of the multi-source heterogeneous monitoring equipment on the slope uploaded by the edge computing gateway, and runs a multi-objective optimization model to solve for the optimal equipment scheduling scheme. The collaborative perception module includes multi-source heterogeneous monitoring equipment distributed on-site and an edge computing gateway. It performs collaborative perception tasks and performs spatiotemporal registration and feature fusion on the collected multimodal monitoring data, outputting the current instability assessment result of the railway slope. The dynamic evaluation module calculates the deviation between the instability assessment result and the structured intent strategy, verifies the degree of intent achievement, and triggers a closed-loop reconstruction mechanism.

[0039] In a preferred embodiment, to further clarify the engineering application scenarios of the present invention and improve the practical deployment feasibility of the system in complex railway slope environments, the system of the present invention preferably adopts a three-layer physical and logical collaborative deployment architecture of "cloud, edge, and field," such as... Figure 3 As shown, its specific structural components and workflow are as follows: Application Service Layer: Used to perform global intent parsing and resource optimization calculations. This layer mainly deploys the intent parsing module and resource orchestration module, which have high computational resource requirements. Specifically, the intent parsing module includes an NLP large model inference engine and a Neo4j graph database, which is further divided into three core sub-modules: (1) Semantic encoding module: responsible for receiving natural language instructions and converting them into intent feature vectors; (2) Knowledge alignment module: responsible for performing high-dimensional spatial similarity matching between the intent feature vectors and domain entities (node ​​vectors) in the knowledge graph, and forcibly correcting instructions that do not conform to geological common sense; (3) Constraint generation module: converts the aligned semantic logic into a configuration language that can be read by the underlying machine and outputs a structured intent strategy. At the same time, the resource orchestration module, as the "nerve center" of the system, receives the above structured strategy, combines the physical topology status of the devices uploaded by the underlying layer, runs a multi-objective optimization function to solve the optimal device combination and link resource configuration, and sends the generated scheduling scheme to the lower-level edge computing gateway in the form of a control instruction matrix.

[0040] Edge computing layer: Used for the fusion processing of multimodal monitoring data and dynamic evaluation of intent achievement. The edge computing layer mainly consists of edge computing gateways deployed along the railway line, carrying the fusion computing end of some collaborative sensing modules and deploying a dynamic evaluation module. The edge computing gateway receives instructions from the cloud and coordinates with field devices. Its core responsibility is to receive multimodal monitoring data collected by various heterogeneous monitoring devices at the lower level locally, and efficiently complete the spatiotemporal registration and feature fusion calculation of the multimodal monitoring data locally, outputting the current railway slope instability assessment result. Simultaneously, the dynamic evaluation module on the edge side executes the intent monitor function, calculating the deviation between the actual output and the business intent; if it finds that the disaster prevention intent constraints are not met, it proactively sends an abnormal interruption command to the cloud to trigger a closed-loop reconstruction mechanism. By offloading data fusion and evaluation tasks to the edge side, the network bandwidth pressure caused by uploading massive amounts of raw data to the cloud is significantly reduced, lowering communication latency.

[0041] The underlying physical network is used to perform on-site environmental perception and multimodal data acquisition tasks. It consists of multi-source heterogeneous monitoring devices (such as rain gauges, GNSS surface displacement gauges, deep inclinometers, PTZ cameras, and inspection drones) distributed across the target area of ​​the railway slope, collectively forming the underlying collaborative perception module. These devices maintain communication with the edge computing gateway, receiving control command matrices, dynamically adjusting sampling frequency, state switching, and perception actions, and continuously transmitting the collected underlying environmental features to the edge computing layer for fusion processing, thereby forming an adaptive closed-loop on-site perception network.

[0042] The present invention also provides an electronic device for performing the above-described intention-driven intelligent monitoring and instability early warning method for railway slopes, the hardware structure of which is as follows: Figure 4 As shown, it includes: a processor 100, a memory 200, and a communication interface 300. The processor 100 can be a central processing unit (CPU) or a graphics processing unit (GPU) for high-speed execution of vector similarity calculation and constrained multi-objective optimization algorithms; the memory 200 is used to store operating system programs, large language model weights, knowledge graph data, and computer instructions for executing the method of this invention; the processor 100, the memory 200, and the communication interface 300 complete the electrical signal transmission between them through a communication bus 400.

[0043] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the intent-driven intelligent monitoring and instability early warning method for railway slopes.

[0044] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept by means of the above teachings or the technology or knowledge in related fields. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An intent-driven railway slope intelligent monitoring and instability early warning method, characterized in that, Includes the following steps: Step S1: Receive natural language disaster prevention intentions input by railway disaster prevention personnel, extract intention entities including slope location, risk type and monitoring needs using an intention parsing model based on pre-trained large language model and named entity recognition, encode the intention entities into intention feature vectors, perform feature matching with knowledge node vectors in a pre-constructed railway geological disaster domain knowledge graph, and generate structured intention strategies. Step S2: The physical topology and equipment resource attributes of the current multi-source heterogeneous monitoring equipment on the slope are obtained in real time through the edge computing gateway deployed along the railway line. A multi-objective optimization model is constructed based on the structured intent strategy, and the optimal equipment scheduling scheme is solved and generated through the resource orchestration algorithm. Step S3: Each multi-source heterogeneous monitoring device in the underlying physical network dynamically updates its local sampling parameters according to the device scheduling scheme and performs collaborative sensing tasks. It performs spatiotemporal registration and feature fusion on the collected multimodal monitoring data and outputs the current instability assessment result of the railway slope. Step S4: Based on the instability assessment results and the structured intent strategy, verify the intent achievement degree and determine whether the disaster prevention intent constraint is met. If the constraint is met, the process ends. If the constraint is not met, trigger the closed-loop reconstruction mechanism, dynamically update the equipment scheduling scheme and distribute it to the underlying physical network for collaborative blind spot filling. In step S1, the specific steps of encoding the intent entity into an intent feature vector and performing feature matching with the knowledge graph include: mapping the extracted intent entities to intent feature vectors using a word embedding model mapping the domain entities in the railway geological disaster domain knowledge graph to knowledge node vectors ; The matching degree between the intent feature vector and the knowledge node vector is calculated using cosine similarity. The calculation formula is as follows: ； Set similarity threshold Only when The mapping relationship between the intent entity and the domain entity is preserved; otherwise, the association rules in the knowledge graph of the railway geological disaster domain are used to perform forced correction in order to generate a structured intent strategy that includes target spatial coordinates, target sensor type and performance constraint parameters; wherein, the performance constraint parameters include maximum tolerable delay and minimum sampling frequency. In step S2, the step of constructing a multi-objective optimization model based on the structured intent strategy includes: The generation of equipment scheduling schemes is transformed into a constrained multi-objective optimization problem, and decision variables are set. ,in Indicates the task Distributed to multi-source heterogeneous monitoring equipment , This indicates no allocation; Define the optimization objective function To maximize the overall monitoring service quality and minimize the total energy consumption of field equipment, the calculation formula is as follows: ； in, This is a set of sub-tasks derived from the current disaster prevention objectives. A collection of multi-source heterogeneous monitoring devices available within the target area; Multi-source heterogeneous monitoring equipment Execute the task Expected data quality gain at that time; Multi-source heterogeneous monitoring equipment Energy consumption for waking up from hibernation to perform high-frequency sensing and transmission tasks; and These are the weighting coefficients for service quality and energy consumption, respectively. The optimization objective function must satisfy the physical constraints of the underlying device: ； ； in, Multi-source heterogeneous monitoring equipment Execute the task Required bandwidth This represents the maximum available bandwidth for the node. Multi-source heterogeneous monitoring equipment Current remaining battery power, Minimum battery level for safe sleep mode; After solving the optimization objective function using a heuristic simulated annealing algorithm or a multi-agent reinforcement learning algorithm, a function is generated that... The largest equipment scheduling scheme is implemented, and a control command matrix is ​​issued.

2. The intention-driven intelligent monitoring and instability early warning method for railway slopes according to claim 1, characterized in that, In step S3, the specific method for feature fusion of multimodal monitoring data is as follows: The weighted feature fusion algorithm is used to calculate the overall probability of slope instability. : ； in, To sense the total number of modes; For the first Observational data of various modes after spatiotemporal registration; This represents the local instability probability based on single-mode data output. For dynamic confidence weights; When the edge computing gateway detects a surge in environmental rainfall that obstructs the view from the visual sensors, it dynamically reduces the weight of the visual modality and correspondingly increases the weight of the microwave radar or the underground inclinometer.

3. The intention-driven intelligent monitoring and instability early warning method for railway slopes according to claim 1, characterized in that, In step S4, the specific method for verifying intent achievement is as follows: Real-time calculation of information entropy and data continuity indicators of instability assessment results; Compare the current monitoring service quality Quality requirements in the structured intent strategy ; When the condition is met If this is deemed an intent violation, an abnormal interruption command is sent to the resource orchestration module. Upon receiving an abnormal interruption command, the multi-objective optimization solution is re-executed to find a suboptimal alternative path. Commands are issued to actively wake up nearby dormant PTZ cameras, or to dispatch inspection drones to approach the target slope section. The image change detection algorithm is used to extract the crack width features on the slope surface. The supplementary visual features are then sent to the feature fusion stage in step S3 for multimodal secondary fusion until the final output comprehensive instability probability meets the disaster prevention intention constraint, thus forming an adaptive closed loop.

4. An intent-driven intelligent monitoring and instability early warning system for railway slopes, used to implement the intent-driven intelligent monitoring and instability early warning method for railway slopes as described in any one of claims 1 to 3, characterized in that, include: The intent parsing module, deployed in the application service layer, includes a pre-trained large language model and graph database. It is used to extract and encode the intent entities of natural language disaster prevention intents, and perform feature matching with knowledge node vectors to generate structured intent strategies. The resource orchestration module is used to receive the physical topology and equipment resource attributes of the multi-source heterogeneous monitoring equipment for slopes uploaded by the structured intent policy and the edge computing gateway, and run a multi-objective optimization model to solve the optimal equipment scheduling scheme. The collaborative sensing module includes multi-source heterogeneous monitoring devices and edge computing gateways distributed on site. It is used to perform collaborative sensing tasks and perform spatiotemporal registration and feature fusion on the collected multimodal monitoring data, and output the current instability assessment results of the railway slope. The dynamic evaluation module is used to calculate the deviation between the instability evaluation result and the structured intent strategy, verify the degree of intent achievement, and trigger the closed-loop reconstruction mechanism.

5. The system according to claim 4, characterized in that, The internal structure of the intent parsing module includes: The semantic encoding module is used to extract intention entities, including slope location, risk type and monitoring needs, from natural language disaster prevention intentions using a pre-trained large language model and named entity recognition technology, and encode the intention entities into intention feature vectors. The knowledge alignment module is used to extract knowledge node vectors from the pre-constructed railway geological disaster knowledge graph, calculate the cosine similarity between the intent feature vector and the knowledge node vector, filter logical illusions in the large language model and perform association correction based on the set similarity threshold; The constraint generation module is used to extract corresponding attribute parameters based on the deterministic mapping relationship output by the knowledge alignment module, and generate a structured intent strategy containing target spatial coordinates, target sensor type and performance constraint parameters.

6. An electronic device, characterized in that, include: Processor, memory, and communication interface; The memory is used to store computer instructions; the processor, the memory, and the communication interface communicate with each other via a communication bus to transmit electrical signals. When executing the computer instructions stored in the memory, the processor implements the intention-driven intelligent monitoring and instability early warning method for railway slopes as described in any one of claims 1 to 3.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the intention-driven intelligent monitoring and instability early warning method for railway slopes as described in any one of claims 1 to 3.

Images