Railway monitoring and early warning system and method based on agent and multi-feature fusion

Abstract

A railway monitoring and early warning system and method based on intelligent agents and multi-feature fusion includes a data acquisition module for synchronously capturing historical data sequences from anomaly sensors within a specified time period; a feature processing module for converting historical data sequences into multi-dimensional feature vectors and unstructured information into business semantic vectors; a historical early warning module for persistently storing historical alarm instances and valid handling solutions; a knowledge retrieval module for retrieving related historical cases based on a dual-path parallel retrieval algorithm and a weighted fusion algorithm to output a comprehensive similarity score; an intelligent agent module for autonomously generating structured risk assessment suggestions and handling solutions based on the retrieved historical cases; and an early warning handling module for pushing the handling solutions generated by the intelligent agent module to the processing terminal and receiving verification results and handling feedback. This invention constructs a self-evolving closed loop of "perception-decision-feedback" to achieve dynamic iteration of system capabilities.

Description

Technical Field

[0001] This invention belongs to the field of railway monitoring technology, specifically relating to a railway monitoring and early warning system and method based on intelligent agents and multi-feature fusion. Background Technology

[0002] With the continuous expansion of railway and urban rail transit construction, the number of projects in the railway monitoring field has steadily increased. During the implementation of various monitoring projects, massive amounts of monitoring and testing data have been accumulated, covering core information such as operational status parameters and safety monitoring indicators of key facilities like lines, bridges, tunnels, and stations. This data has become a crucial foundation for ensuring the safe and stable operation of facilities and supporting maintenance decisions. Against this backdrop, accurate, reliable, and intelligent early warning mechanisms are becoming increasingly critical, directly determining whether monitoring data can be efficiently transformed into safety assurance capabilities, and are of great significance for improving the quality of rail transit operation and maintenance and risk prevention and control.

[0003] Current railway monitoring systems generally use fixed thresholds as the core early warning mechanism. While intuitive and simple, its limitations in practical engineering applications are becoming increasingly apparent.

[0004] First, the early warning logic is simplistic and lacks sufficient anti-interference capabilities. Current railway monitoring systems generally use a fixed threshold method as the core early warning mechanism. While this method is intuitive, in the complex railway operating environment, sensors are susceptible to non-fault noise such as lightning interference, random drift, and instantaneous load fluctuations. The reliance on a single threshold lacks the ability to mine the deep characteristics of the monitoring data, resulting in a persistently high false alarm rate and significantly increasing the handling pressure and costs for on-site maintenance personnel.

[0005] Secondly, the data lacks strong correlation and collaborative analysis capabilities. Existing monitoring data exhibits a clear "siloed" state, with systems often monitoring indicators only for a single measuring point independently. Due to the lack of multi-dimensional alignment of monitored objects with changing trends and business scenarios, it is difficult to achieve feature fusion analysis across devices and multiple sensors. This makes it impossible for the system to comprehensively verify complex, hidden faults using historical sequences from related devices, hindering accurate collaborative early warning decisions.

[0006] Finally, the experience reuse rate is low, and there is a lack of knowledge accumulation and reasoning mechanisms. Railway operation and maintenance involves a large number of professional procedures and historical failure cases, but this valuable experience currently exists mainly in the form of unstructured documents or individual expert experience. Existing monitoring systems lack the ability to automatically summarize, structure, archive, and intelligently reason about early warning instances. When new anomalies occur, the system cannot automatically retrieve and generate targeted troubleshooting suggestions from historical early warning modules, resulting in a gap between "data perception" and "intelligent handling," and failing to form a closed-loop learning and evolutionary capability.

[0007] Therefore, this case proposes a railway monitoring and early warning system and method based on intelligent agent and multi-feature fusion to solve the above-mentioned technical problems. Summary of the Invention

[0008] In order to solve the problems existing in the prior art, the present invention provides a railway monitoring and early warning system and method based on intelligent agent and multi-feature fusion, which is used to construct a self-evolving closed loop of "perception-decision-feedback" to realize dynamic iteration of system capabilities.

[0009] The specific technical solution of this invention is as follows:

[0010] A railway monitoring and early warning system based on intelligent agents and multi-feature fusion includes: a data acquisition module, used to acquire monitoring data from field sensors in real time, and perform preliminary screening based on preset fixed threshold logic; when the monitoring data exceeds the threshold, a preliminary alarm signal is triggered, and the historical data sequence of abnormal sensors and surrounding related equipment within a specified time period is captured simultaneously.

[0011] The feature processing module is used to transform the historical data sequence into a multidimensional feature vector representing the evolution trend of the monitored object. And transform the unstructured information related to the anomaly sensors into business semantic vectors. ;

[0012] The historical early warning module is used to persistently store historical alarm instances and their corresponding multi-dimensional feature vectors. Business semantic vector And verify effective treatment plans;

[0013] The knowledge retrieval module is used to calculate the multidimensional feature vector in the historical early warning module based on a dual-path parallel retrieval algorithm. Spatial distance similarity and the business semantic vector The semantic similarity is calculated, and a comprehensive similarity score is output based on a weighted fusion algorithm to retrieve related historical cases;

[0014] The intelligent agent module, with a large language model as its core, is used to perform time-series analysis, causal reasoning, and authenticity verification on the current anomaly based on historical cases retrieved by the knowledge retrieval module and in combination with preset business prompt word templates, and autonomously generate structured risk investigation suggestions and handling plans.

[0015] The early warning and handling module, as a human-computer interaction interface, is used to push the handling plan generated by the intelligent agent module to the processing terminal, and to receive the verification results and handling feedback from the processing terminal.

[0016] The intelligent agent module is also used to automatically summarize and structure the current alarm instance based on the handling feedback, and store it in the historical early warning module to realize the iterative update of knowledge.

[0017] Preferably, the multidimensional feature vector include:

[0018] The trend derivative feature is used to characterize the average slope and maximum abrupt change of the monitored data over a time period.

[0019] Signal morphology features are used to characterize the impact morphology and energy distribution of monitoring data.

[0020] Preferably, the trend derivative feature includes the average slope feature. and the largest mutation feature The signal morphological features include pulse factor. and waveform factor The ;

[0021] in:

[0022]

[0023]

[0024] This represents the monitored value at the current sampling time; This represents the monitoring value at the previous sampling time. This reflects the rate of change of the signal within a unit sampling period;

[0025]

[0026]

[0027] It represents the change in the first-order difference, reflecting the curvature or the degree of abrupt change in the monitored numerical waveform;

[0028]

[0029]

[0030] The maximum absolute value of the signal within the current window; This represents the absolute average value of the signal within the window. The root mean square value reflects the average energy level of the signal;

[0031] .

[0032] Preferably, the business prompt word template preset by the intelligent agent module includes:

[0033] The system role setting field is used to define the identity of the intelligent agent;

[0034] The environmental context field is used to input the device information, installation location, current weather, and historical risk factors for the abnormal sensor.

[0035] The monitoring data description field is used to input time-series data from abnormal sensors and surrounding related sensors;

[0036] The reasoning logic requires fields that are used to make causal inferences based on morphological analysis, environmental correlation analysis, and spatial consistency check analysis.

[0037] The output format specification field is used to limit the agent's output of judgment results, confidence levels, and inference processes in JSON format.

[0038] Preferably, the historical early warning module adopts a hybrid storage mode of relational database and vector database; the relational database is used to store knowledge information summarized by the agent module, including agent judgment conclusions and structured summaries; the vector database is used to store the multidimensional feature vectors. and the business semantic vector .

[0039] Preferably, the knowledge retrieval module employs a dual-path parallel retrieval algorithm, utilizing the multi-dimensional feature vector. Spatial distance similarity and the business semantic vector The weighted fusion of semantic similarity enables a three-dimensional calibration between physical evolution mechanisms and business context; wherein, the multidimensional feature vector The business semantic vector is used to characterize the evolution trend of monitoring indicators. External boundary conditions used to reflect early warning events.

[0040] Another objective of this invention is to provide a railway monitoring and early warning method based on agent-based and multi-feature fusion, and to apply a railway monitoring and early warning system based on agent-based and multi-feature fusion, comprising the following steps:

[0041] Step S1: Acquire monitoring data from field sensors in real time, perform preliminary screening based on preset fixed threshold logic, trigger a preliminary alarm signal when the monitoring data exceeds the threshold, and simultaneously capture historical data sequences of abnormal sensors and surrounding related devices within a specified time period.

[0042] Step S2: Based on the preset prompt word template, the intelligent agent module performs time-series analysis and logical reasoning on the monitoring data of the abnormal sensor and surrounding sensors, combined with the characteristics of the surrounding environment and the types of risks that may occur, to determine whether the alarm signal is a real physical anomaly.

[0043] Step S3: Perform feature processing on sensor data determined to be real physical anomalies, extract change trends and morphological features, and construct a multi-dimensional feature vector of the monitoring data. Simultaneously, it acquires basic information from anomaly sensors, geographical location, weather conditions, etc., constructs semantic text, and vectorizes it using an embedding model to obtain business semantic vectors. ;

[0044] Step S4: The knowledge retrieval module uses a dual-path parallel retrieval algorithm to calculate the multi-dimensional feature vector in the historical early warning module. Spatial distance and the business semantic vector Based on the semantic similarity, a comprehensive similarity score is output using a weighted fusion algorithm to retrieve related historical cases;

[0045] Step S5: Based on the retrieved historical cases, the intelligent agent module generates a structured risk assessment list and handling plan, and pushes it to the processing terminal;

[0046] Step S6: Receive the on-site verification results and handling feedback from the processing terminal, and process the multi-dimensional feature vector of the current alarm instance. Business semantic vector The early warning situation and the results of the handling process are stored in the historical early warning module.

[0047] Preferably, in step S2, the specific steps of the intelligent agent module performing temporal analysis and logical reasoning based on the preset prompt word template include:

[0048] Morphological analysis: Determine whether the data curve is a transient, impulsive change or a smooth, continuous evolution;

[0049] Environmental correlation analysis: Based on the current weather, determine whether the abnormal fluctuations can be attributed to environmental factors;

[0050] Spatial consistency check and analysis: By comparing the monitoring data of surrounding sensors, determine whether there are linkage characteristics that conform to physical laws;

[0051] Causal inference: Based on the above analysis results, it is determined that the signal attribute is a real anomaly or a false alarm caused by environmental interference.

[0052] Preferably, in step S4, the knowledge retrieval module employs a dual-path parallel retrieval algorithm and calculates a comprehensive similarity score using a weighted fusion algorithm.

[0053]

[0054] in, The similarity between the business semantic vector and the semantic vector in the historical early warning module. The similarity between the multidimensional feature vector and the feature vector in the historical early warning module. and The preset weighting coefficients, and .

[0055] Preferably, in step S6, the specific steps for storing the alarm instance in the historical early warning module include:

[0056] The intelligent agent module automatically summarizes alarm instances and handling results to generate structured knowledge.

[0057] Store structured knowledge in a knowledge record table of a relational database;

[0058] Multidimensional feature vectors and business semantic vector Store the warning vector index table in the vector database and establish a connection through the knowledge record ID.

[0059] The beneficial effects of this invention are:

[0060] 1. Significantly reduces false alarm rate and improves the accuracy of early warnings. This system transforms raw data into multi-dimensional feature vectors through a feature processing module and uses an intelligent agent module for time-series analysis and causal reasoning. It can automatically identify and eliminate false anomalies caused by instantaneous load fluctuations or hardware failures, effectively distinguishing between real physical anomalies and environmental interference, greatly reducing the workload of maintenance personnel in performing ineffective checks.

[0061] 2. Breaking down "information silos" to achieve collaborative and interconnected analysis. The system no longer views changes in a single sensor in isolation. Instead, it synchronously captures features from surrounding related devices through the data acquisition module and combines this with multi-dimensional feature extraction technology from the feature processing module. This enables the system to detect complex and trending security risks, improving its early warning capabilities for systemic risks.

[0062] 3. The system facilitates the accumulation and reuse of operational knowledge, enhancing intelligent reasoning capabilities. A historical early warning module is constructed, and the "dual-path parallel retrieval algorithm" of the knowledge retrieval module enables precise location of similar scenarios from massive historical cases. The intelligent agent module transforms fragmented professional experience into structured troubleshooting suggestions, providing expert-level decision support for operations and maintenance personnel, effectively solving the problem of low processing efficiency caused by the loss of expert experience.

[0063] 4. Construct a self-evolving closed loop of "perception-decision-feedback" to achieve dynamic iteration of system capabilities. The system obtains on-site verification results in real time through the early warning and handling module, and drives the intelligent agent module to automatically summarize and structure-archive early warning instances and handling methods. This feedback learning mechanism enables continuous updates to the historical early warning module, allowing the system to evolve in practice, and the accuracy and intelligence level of early warnings continuously improve as the application scale expands. Attached Figure Description

[0064] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that these drawings are designed for illustrative purposes only and are not intended to limit the scope of the present invention. Furthermore, unless specifically indicated, these drawings are intended only to conceptually illustrate the structural construction described herein and are not necessarily drawn to scale.

[0065] Figure 1 This is a schematic diagram of the system composition of the present invention; Detailed Implementation

[0066] First, it should be noted that the specific structure, features, and advantages of the present invention will be described in detail below by way of examples. However, all descriptions are for illustrative purposes only and should not be construed as limiting the present invention in any way. Furthermore, any single technical feature described or implied in the embodiments mentioned herein, or any single technical feature shown or implied in the accompanying drawings, can still be arbitrarily combined or deleted among these technical features to obtain more other embodiments of the present invention that may not be directly mentioned herein. Additionally, for the sake of simplifying the drawings, the same or similar technical features may be indicated only in one place in the same drawing.

[0067] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "screw connection," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0068] The following is in conjunction with the appendix Figure 1 This invention will be described in detail.

[0069] like Figure 1As shown, a railway monitoring and early warning system based on intelligent agents and multi-feature fusion includes: a data acquisition module, used to acquire monitoring data from field sensors in real time, and perform preliminary screening based on preset fixed threshold logic. When the monitoring data exceeds the threshold, a preliminary alarm signal is triggered, and the historical data sequence of abnormal sensors and surrounding related equipment within a specified time period is captured simultaneously.

[0070] The feature processing module is used to transform the historical data sequence into a multidimensional feature vector representing the evolution trend of the monitored object. And transform the unstructured information related to the anomaly sensors into business semantic vectors. ;

[0071] The historical early warning module is used to persistently store historical alarm instances and their corresponding multi-dimensional feature vectors. Business semantic vector And verify effective treatment plans;

[0072] The knowledge retrieval module is used to calculate the multidimensional feature vector in the historical early warning module based on a dual-path parallel retrieval algorithm. Spatial distance similarity and the business semantic vector The semantic similarity is calculated, and a comprehensive similarity score is output based on a weighted fusion algorithm to retrieve related historical cases;

[0073] The intelligent agent module, with a large language model as its core, is used to perform time-series analysis, causal reasoning, and authenticity verification on the current anomaly based on historical cases retrieved by the knowledge retrieval module and in combination with preset business prompt word templates, and autonomously generate structured risk investigation suggestions and handling plans.

[0074] The early warning and handling module, as a human-computer interaction interface, is used to push the handling plan generated by the intelligent agent module to the processing terminal, and to receive the verification results and handling feedback from the processing terminal.

[0075] The intelligent agent module is also used to automatically summarize and structure the current alarm instance based on the handling feedback, and store it in the historical early warning module to realize the iterative update of knowledge.

[0076] Working principle:

[0077] like Figure 1 As shown, this system includes a data acquisition module, a feature processing module, a knowledge retrieval module, an early warning and response module, a historical early warning module, and an intelligent agent module.

[0078] The data acquisition module is deployed at railway monitoring sites and connects to various types of sensors via wired or wireless communication. This module is responsible for acquiring real-time monitoring data from railway sensors and performing preliminary screening of the raw signals using built-in fixed threshold logic. Once an anomaly is detected, it immediately and synchronously captures historical data sequences from the abnormal sensor and surrounding related equipment within a specified time period and sends preliminary alarm signals and relevant data to downstream modules.

[0079] The feature processing module receives data from the data acquisition module and performs multimodal feature processing. This module transforms the monitoring data and unstructured information into multidimensional feature vectors. With business semantic vector This enables multimodal alignment of the monitored objects with their changing trends and monitoring business scenarios.

[0080] The knowledge retrieval module, based on a dual-path parallel retrieval algorithm, calculates the spatial distance and semantic similarity of multi-dimensional feature vectors and semantic vectors in the historical early warning module, respectively. It then outputs a comprehensive similarity score based on a weighted fusion algorithm, retrieving the highest-similar sets of historical early warning cases. This automatically retrieves related historical cases and uses an intelligent agent to generate structured investigation suggestions and handling plans for the current risk. This module utilizes multi-dimensional feature vectors... Characterize the evolution trend of monitoring indicators through business semantic vectors It reflects the external boundary conditions of early warning events, achieving a three-dimensional calibration between physical evolution mechanisms and operational context.

[0081] The early warning and response module, serving as the human-machine interface, is responsible for pushing the early warning and response plans generated by the intelligent agent module to the processing terminal. In this embodiment, the processing terminal is set to the maintenance personnel, who receive in real time the verification results of the accuracy of the early warnings and the corresponding on-site actions taken by the on-site maintenance personnel. After the on-site verification and response are completed manually, the intelligent agent module automatically summarizes and structures the alarm instances and response results, completing the iterative update of the historical early warning module and realizing a fully autonomous closed loop from "data perception - intelligent analysis - plan generation - feedback learning".

[0082] The historical early warning module, as the core experience asset of the system, is responsible for persistently storing historical alarm instances summarized by the intelligent agent, corresponding multi-dimensional features, and verified effective handling solutions, providing data support for the accuracy evaluation of early warning handling and the iterative updating of knowledge.

[0083] The intelligent agent module, centered on a large language model, uses preset business prompt word templates and railway operation and maintenance knowledge to perform time-series analysis, causal reasoning, and authenticity verification on abnormal data, automatically distinguishing between real physical anomalies and interference false alarms. Simultaneously, based on the dual-path retrieval results and the historical early warning module, it autonomously generates structured and scenario-based risk investigation steps, handling procedures, and safety operation suggestions. After manual on-site verification and handling, it automatically summarizes and structures the alarm instances and handling results, completing iterative updates to the historical early warning module. This achieves a fully autonomous closed loop from "data perception—intelligent judgment—solution generation—feedback learning," continuously improving the system's early warning accuracy and handling intelligence level.

[0084] Furthermore, in the embodiments, the multidimensional feature vector can also be considered. include:

[0085] The trend derivative feature is used to characterize the average slope and maximum abrupt change of the monitored data over a time period.

[0086] Signal morphology features are used to characterize the impact morphology and energy distribution of monitoring data.

[0087] Furthermore, in the embodiments, the trend derivative feature may include an average slope feature. and the largest mutation feature The signal morphological features include pulse factor. and waveform factor The ;

[0088] in:

[0089]

[0090]

[0091] This represents the monitored value at the current sampling time; This represents the monitoring value at the previous sampling time. This reflects the rate of change of the signal within a unit sampling period;

[0092]

[0093]

[0094] It represents the change in the first-order difference, reflecting the curvature or the degree of abrupt change in the monitored numerical waveform;

[0095]

[0096]

[0097] The maximum absolute value of the signal within the current window; This represents the absolute average value of the signal within the window. The root mean square value reflects the average energy level of the signal;

[0098] .

[0099] In this embodiment, in order to extract the evolutionary form and trend of the monitoring data and accurately determine the early warning, this case involves monitoring data... The following feature processing is performed:

[0100] (1) Extraction of trend derivative features

[0101] The dynamic evolution characteristics of the monitoring data curve are captured by the difference operator.

[0102] Average slope characteristics It reflects the overall evolution (upward or downward trend) of the monitoring data over a time period. It can effectively filter out subtle fluctuations caused by random noise, extract the global evolution direction of the monitoring data, and provide macro-trend support for early warning.

[0103]

[0104]

[0105] in, This represents the monitored value at the current sampling time; This represents the monitoring value at the previous sampling time. It reflects the rate of change of the signal within a unit sampling period.

[0106] The maximum abrupt change characteristic reflects the most dramatic dynamic response amplitude of the monitored data within a time frame. Capturing the most dramatic response amplitude within a window is a key indicator for identifying instantaneous impacts or structural damage.

[0107]

[0108]

[0109] in, It represents the change in the first-order difference, reflecting the curvature or the degree of abrupt change in the monitored numerical waveform.

[0110] (2) Signal morphological characteristics

[0111] Using time-series monitoring data as the analysis signal, the quantitative extraction of signal waveform distribution characteristics is achieved by calculating dimensionless indices such as waveform factor and impulse factor.

[0112] The impulse factor, or impulse pattern characteristic, is used to distinguish between smooth trends and abnormal fluctuations with impulses. When the monitored object becomes loose, is impacted, or the sensor malfunctions, the impulse factor increases significantly, enabling accurate identification of outlier impulse signals in complex background noise and improving the accuracy of early warnings.

[0113]

[0114]

[0115] in, The maximum absolute value of the signal within the current window; This represents the absolute average value of the signal within the window. The root mean square value reflects the average energy level of the signal.

[0116] Energy distribution characteristics, or waveform factors, are used to characterize the smoothness and stability of the energy distribution of a monitoring signal. Under normal circumstances, the energy distribution of a monitored object usually follows a specific statistical law, but when the monitored object enters a sub-healthy state or suffers fatigue damage, the concentration of the energy distribution will shift. The introduction of this method provides a quantitative basis for assessing the overall stability of the monitored objects.

[0117]

[0118] (3) Multidimensional feature vectors

[0119] To eliminate the order-of-magnitude differences between features of different magnitudes, all calculated features are standardized. Based on these features, a multidimensional feature vector is constructed. To accurately describe the changing trends and evolutionary trajectories of monitoring data, its expression is:

[0120]

[0121] Through multidimensional feature vectors It achieves accurate mapping and comprehensive characterization of the deep features of monitoring data, and its advantages are:

[0122] Comprehensive state characterization. By integrating four types of features, it achieves simultaneous characterization of the dynamic trends, transient changes, distribution patterns, and energy stability of the monitored objects, and can quantitatively describe the evolution trajectory of monitoring data, providing data reference for the generation of response plans.

[0123] Complementarity and completeness. It focuses on the longitudinal evolution along the timeline, while It focuses on the horizontal shape of the statistical distribution. By coupling multiple dimensions, it avoids the problem of single indicators being easily disturbed.

[0124] Precise status characterization. This feature, by constructing a digital status fingerprint, can effectively identify complex operating modes such as slow degradation, sudden failures, and environmental interference, laying a solid data foundation for the leap from reactive alarms to predictive maintenance.

[0125] Furthermore, in the embodiments, the business prompt word template preset by the intelligent agent module may include:

[0126] The system role setting field is used to define the identity of the intelligent agent;

[0127] The environmental context field is used to input the device information, installation location, current weather, and historical risk factors for the abnormal sensor.

[0128] The monitoring data description field is used to input time-series data from abnormal sensors and surrounding related sensors;

[0129] The reasoning logic requires fields that are used to make causal inferences based on morphological analysis, environmental correlation analysis, and spatial consistency check analysis.

[0130] The output format specification field is used to limit the agent's output of judgment results, confidence levels, and inference processes in JSON format.

[0131] In this embodiment, examples of business prompt word templates are shown in Table 1.

[0132] System character settings You are a railway safety monitoring expert, responsible for performing logical verification of the physical authenticity of real-time sensor alarms. Context Abnormal Sensor: [Sensor Number, Type] Burial Location: [Specific Location] Current Weather: [Weather Conditions] Historical Risks: [Static Risks Obtained Based on Sensor Type] Monitoring data description Data sequence for the 30 minutes prior to the alarm: [List of numerical values ​​in JSON format] Data from surrounding related sensors: [List of data in JSON format] Reasoning logic requirements 1. Morphological Analysis: Observe the data curve. If it is a sudden, pulse-like change, it is suspected to be due to non-physical factors such as electromagnetic interference or poor equipment contact. If it is a smooth and continuous evolution, it is consistent with the deformation characteristics of a structure under stress. 2. Environmental Correlation Analysis: Combine the current weather conditions to determine whether the abnormal fluctuations can be directly attributed to environmental factors. 3. Spatial Consistency Check: Compare the monitoring data from surrounding related sensors to determine whether there are any linkage characteristics consistent with physical laws. 4. Causal Inference: Based on the above analysis results, clearly determine the signal attributes. Output format specifications Returns only in JSON format: {"result": [Result: True / False],"confidence": [Confidence level: 0-100%],"agent_reasoning": [A brief description of the reasoning process, not exceeding 200 words]}

[0133] Table 1

[0134] Furthermore, in the embodiments, the historical early warning module may employ a hybrid storage mode of relational database and vector database; the relational database is used to store knowledge-based information summarized by the agent module, including agent judgment conclusions and structured summaries; the vector database is used to store the multidimensional feature vectors. and the business semantic vector This is to support efficient dual-path parallel retrieval.

[0135] This embodiment details the specific data structure of the historical early warning module.

[0136] The historical early warning module stores structured numerical features and high-dimensional semantic vectors, so it adopts a hybrid storage mode of "relational database (MySQL) + vector database (ES, ElasticSearch)".

[0137] A relational database stores the "knowledge-based" information that has been summarized by the agent module, which is used for the assembly of the final solution. Based on the system's early warning table and disposal table, a knowledge record table (knowledge_record) is designed:

[0138] The historical early warning module adopts a hybrid storage mode of relational database and vector database. In the relational database, based on the existing early warning table and handling table of the system, a knowledge record table (knowledge_record) is designed, the structure of which is shown in Table 2.

[0139] Id Long Data table primary key alarm_id Long Core foreign key, associated with the system early warning information table disposition_id Long Core foreign key, associated system processing information table agent_reasoning Text Agent's judgment conclusion: Record the reasoning logic that the agent module determined to be a real anomaly at that time. knowledge_summary Text Structured summary: The "phenomenon-cause-countermeasure" three-part knowledge generated by the agent module after the action is completed serves as direct material for the next generation.

[0140] Table 2

[0141] The vector database is responsible for storing high-dimensional feature data and utilizes its high-performance retrieval capabilities to achieve rapid comparison between physical features and business semantics. A warning vector index table (warning_vector_index) is designed in Elasticsearch, and its structure is shown in Table 3.

[0142] knowledge_record_id keyword Corresponding to the knowledge record ID in a relational database feature_vector dense_vector l2_norm Store the multidimensional feature vectors of the monitoring data generated by the feature processing module. semantic_vector dense_vector cosine Store business semantic vectors generated by the Embedding model

[0143] Table 3

[0144] Through the above data structure design, when performing dual-path parallel retrieval, the system can quickly complete spatial distance and semantic similarity calculations in the vector database, and obtain complete disposal plans and knowledge summaries from the relational database through knowledge_record_id association, thereby achieving efficient and accurate historical case retrieval.

[0145] Furthermore, in the embodiments, the knowledge retrieval module may employ a dual-path parallel retrieval algorithm, utilizing the multi-dimensional feature vectors. Spatial distance similarity and the business semantic vector The weighted fusion of semantic similarity enables a three-dimensional calibration between physical evolution mechanisms and business context; wherein, the multidimensional feature vector The business semantic vector is used to characterize the evolution trend of monitoring indicators. External boundary conditions used to reflect early warning events.

[0146] Example 2:

[0147] A railway monitoring and early warning method based on agent-based and multi-feature fusion, applying a railway monitoring and early warning system based on agent-based and multi-feature fusion, includes the following steps:

[0148] Step S1: Acquire monitoring data from field sensors in real time, perform preliminary screening based on preset fixed threshold logic, trigger a preliminary alarm signal when the monitoring data exceeds the threshold, and simultaneously capture historical data sequences of abnormal sensors and surrounding related devices within a specified time period.

[0149] Step S2: Based on the preset prompt word template, the intelligent agent module performs time-series analysis and logical reasoning on the monitoring data of the abnormal sensor and surrounding sensors, combined with the characteristics of the surrounding environment and the types of risks that may occur, to determine whether the alarm signal is a real physical anomaly.

[0150] Step S3: Perform feature processing on sensor data determined to be real physical anomalies, extract change trends and morphological features, and construct a multi-dimensional feature vector of the monitoring data. Simultaneously, it acquires basic information from anomaly sensors, geographical location, weather conditions, etc., constructs semantic text, and vectorizes it using an embedding model to obtain business semantic vectors. ;

[0151] Step S4: The knowledge retrieval module uses a dual-path parallel retrieval algorithm to calculate the multi-dimensional feature vector in the historical early warning module. Spatial distance and the business semantic vector Based on the semantic similarity, a comprehensive similarity score is output using a weighted fusion algorithm to retrieve related historical cases;

[0152] Step S5: Based on the retrieved historical cases, the intelligent agent module generates a structured risk assessment list and handling plan, and pushes it to the processing terminal;

[0153] Step S6: Receive the on-site verification results and handling feedback from the processing terminal, and process the multi-dimensional feature vector of the current alarm instance. Business semantic vector The early warning situation and the results of the handling process are stored in the historical early warning module.

[0154] Workflow:

[0155] 1. The system acquires real-time sensor data from the field through the data acquisition module, performs preliminary identification through preset fixed threshold logic, and triggers a preliminary alarm signal when the data exceeds the threshold.

[0156] 2. After triggering the initial alarm, the system acquires data from the abnormal sensor and its surrounding sensors for a period of time prior to the alarm time. Simultaneously, based on information such as sensor type, installation location, and weather conditions, the intelligent agent module uses a prompt word template, combined with the characteristics of the surrounding environment and the types of risks that may occur, to generate a business prompt word. The module then performs time-series analysis and logical reasoning on the monitoring data of the abnormal sensor and the surrounding sensors to determine whether the alarm signal is a real physical anomaly.

[0157] 3. Perform feature processing on sensor data that the intelligent agent module identifies as abnormal, extracting features such as trend and morphological logic, and construct a multi-dimensional feature vector of the monitoring data. Acquire basic information from anomaly sensors, geographical location, weather conditions, etc., construct semantic text based on a preset template, and vectorize it using an embedding model to obtain business semantic vector data. .

[0158] 4. Through the early warning and handling module, a dual-path parallel retrieval algorithm is adopted, and a weighted fusion algorithm is used in the historical early warning module to calculate multi-dimensional feature vectors. Euclidean distance and semantic vector data The cosine similarity score is used to extract the n sets of historical early warning data with the highest similarity.

[0159] 5. Through the intelligent agent module, based on the retrieved historical early warning information and on-site handling methods, a risk investigation list and handling suggestions are generated and provided to the early warning and handling module, waiting for the operation and maintenance personnel to confirm the on-site handling.

[0160] 6. The system receives feedback results from manual handling, including the alarm status, handling process results, and multi-dimensional feature vectors of the alarm instance. Store in the historical early warning module.

[0161] Furthermore, in the embodiments, the specific steps of the intelligent agent module performing temporal analysis and logical reasoning based on the preset prompt word template in step S2 include:

[0162] Morphological analysis: Determine whether the data curve is a transient, impulsive change or a smooth, continuous evolution;

[0163] Environmental correlation analysis: Based on the current weather, determine whether the abnormal fluctuations can be attributed to environmental factors;

[0164] Spatial consistency check and analysis: By comparing the monitoring data of surrounding sensors, determine whether there are linkage characteristics that conform to physical laws;

[0165] Causal inference: Based on the above analysis results, it is determined that the signal attribute is a real anomaly or a false alarm caused by environmental interference.

[0166] Furthermore, in the embodiments, it can be considered that in step S4, the knowledge retrieval module adopts a dual-path parallel retrieval algorithm and calculates the comprehensive similarity score through a weighted fusion algorithm:

[0167]

[0168] in, The similarity between the business semantic vector and the semantic vector in the historical early warning module. The similarity between the multidimensional feature vector and the feature vector in the historical early warning module. and The preset weighting coefficients, and .

[0169] In this embodiment, a dual-path parallel retrieval algorithm is used, and the knowledge retrieval module performs dual-path matching in the pre-built "alarm-handling" knowledge base:

[0170] Evolutionary morphology channel: Calculate the current vector The standardized Euclidean distance with the historical feature vectors in the database, by shielding the interference of the sensor's basic values, deeply characterizes the fluctuation slope, energy distribution and impact intensity of the monitored object over time, thereby accurately reflecting the essential similarity of the evolution trend of the monitoring indicators.

[0171] Business context channel: Calculate the current vector By using the cosine similarity of the early warning information vectors in the database, this channel reflects the business background relevance of early warning events under external boundary conditions through semantic association of equipment models, meteorological factors, and geospatial information.

[0172] Cross-modal fusion analysis: An adaptive weighted fusion algorithm is used to calculate the comprehensive similarity score, taking into account both the physical evolution mechanism and the external business context, and extracting the set of historical early warning summaries and successful handling cases with the highest similarity.

[0173]

[0174] in, and These are preset fixed weighting coefficients, and The specific settings depend on the actual scenario. Generally speaking, It is 0.55. It is 0.45.

[0175] The core reason this invention employs dual-path parallel retrieval instead of single-path retrieval is the mutual calibration between the physical characteristics of the monitored data and the business logic of the monitored object. If relying solely on one path, the retrieval system will suffer from significant cognitive bias due to limited cognitive dimensions, where:

[0176] Using only evolutionary morphology channels: Relying solely on the similarity of monitored data waveforms without considering the business context or the constraints of external boundary conditions leads to search results with similar waveforms but completely different actual scenarios. For example, the same fluctuation slope from a stress sensor might represent normal temperature rise in summer but abnormal friction in winter.

[0177] Using only the business context channel: Focusing solely on the textual description of the business scenario without considering real-time monitoring data makes it impossible to capture abnormal evolution and sudden anomalies in sensors. For example, given two sensors in the same business scenario, one is functioning well while the other is about to fail due to internal mechanical fatigue. Contextual retrieval alone cannot detect the ongoing physical evolution risk.

[0178] Dual-path parallel retrieval is adopted: Through cross-validation of dual-path retrieval results, the retrieval algorithm is transformed from "single matching" to "three-dimensional calibration". Only when the data evolution trend and business scenario point to a certain historical case will the case be recalled with high confidence, effectively filtering out interference items.

[0179] Furthermore, in the embodiments, step S6, which involves storing the alarm instance in the historical early warning module, may include the following specific steps:

[0180] The intelligent agent module automatically summarizes alarm instances and handling results to generate structured knowledge.

[0181] Store structured knowledge in a knowledge record table of a relational database;

[0182] Multidimensional feature vectors and business semantic vector Store the warning vector index table in the vector database and establish a connection through the knowledge record ID.

[0183] Furthermore, in the embodiments, step S3, the extraction of business semantic vectors, may include the following steps:

[0184] Data Acquisition: Obtain basic information about the abnormal sensor, including core parameters such as sensor device number, model, installation location, abnormal trigger time, and abnormal type; simultaneously, link the positioning module and the meteorological monitoring interface to synchronously collect the precise geographical location of the abnormal sensor and real-time weather conditions.

[0185] Semantic text construction: Following the preset standardized text structure template, the basic information of the abnormal sensors, geographical location, weather conditions and other data collected above are systematically integrated and filled in to construct a complete, standardized and consistent unified semantic text that fits the data characteristics, ensuring that it can fully reflect the key information related to the abnormal sensors.

[0186] Vectorization: The constructed prompt words are input into the Embedding model for vectorization. Through feature extraction and mapping of text information by the model, vector data of the corresponding prompt words are finally generated. This provides standardized data support for subsequent anomaly analysis, data comparison and other related processing.

[0187] This invention selects Qwen-Embedding-4B as the text vectorization model, which has significant advantages in Chinese semantic understanding and representation. This model can accurately capture synonymous relationships, implicit semantics, and fine-grained semantic differences in text. It has strong modeling capabilities for Chinese sentence structure, professional terminology, and contextual relationships, and can stably adapt to professional semantic information in the railway operation and maintenance field, meeting the practical application needs of text retrieval, knowledge matching, and abnormal semantic recognition in this scenario.

[0188] Application Example 1:

[0189] Based on Examples 1 and 2, the present invention is further illustrated by a real-world project demonstration, as follows:

[0190] Take the monitoring data of a railway slope during a period of heavy rainfall as an example.

[0191] First, the system receives data from the GNSS-03 monitoring point located in the middle of the slope. According to real-time calculations, the cumulative displacement of this point in the horizontal X direction exceeds the preset alarm threshold of 30mm, reaching 34.2mm. The system immediately triggers a preliminary alarm signal.

[0192] Next, the system automatically retrieves monitoring data from GNSS-03 and its adjacent monitoring sensors (GNSS-02 and GNSS-04) over the past 48 hours, while also acquiring data from the on-site rain gauge, and transmits this data to the intelligent agent module for further evaluation. Based on information such as sensor type, installation location, and weather conditions, and in conjunction with the monitoring data from multiple sensors, the intelligent agent module determines that the GNSS displacement curve does not exhibit a sudden jump, but rather a non-linear accelerating trend with increasing rainfall; simultaneously, adjacent monitoring points show coordinated displacement in the same direction, thus confirming the warning is valid.

[0193] Furthermore, the system performs in-depth processing on sensor data to construct multi-dimensional feature vectors; it acquires basic sensor information, geographical location, weather conditions, and other information, and generates semantic text and business semantic vector data based on preset templates. Through the early warning and handling module, a dual-path parallel retrieval algorithm is used to calculate the three sets of historical early warning data with the highest similarity. The retrieval shows that the feature combination has a feature similarity of 0.92 with a certain past landslide precursor feature.

[0194] Finally, based on the retrieved historical handling plans and the current on-site situation, the intelligent agent module generates a risk assessment list and handling suggestions, as follows:

[0195] Risk assessment: The slope is in an unstable creep stage and there is a risk of shallow landslide.

[0196] Risk Assessment Checklist:

[0197] 1. Check whether the drainage ditch at the top of the slope is blocked or cracked.

[0198] 2. Check for seepage and turbidity at the toe of the slope.

[0199] 3. Verify whether the GNSS receiver antenna is being obstructed or interfered with by surrounding vegetation.

[0200] Recommendations for handling:

[0201] Public Works Department:

[0202] 1. Encrypted monitoring: It is recommended to increase the GNSS data sampling frequency from once per hour to once every 5 minutes, and to continuously monitor subsequent deformation.

[0203] 2. Emergency hazard mitigation: Immediately assign maintenance personnel to conduct on-site inspections, investigate and verify potential risks and hazards, focus on cleaning the intercepting ditch, and cover it with tarpaulin to reduce infiltration.

[0204] Dispatch Department:

[0205] 1. Speed ​​Limit Instruction: It is recommended to implement temporary speed limits on both the up and down lines in this section until on-site inspection confirms safety.

[0206] 2. Reserved section: Maintain the train departure interval at the stations ahead of this section to allow for handling of possible emergency shutdowns.

[0207] Reinforcement Recommendation: If the displacement acceleration continues to exceed 2 mm / d², it is recommended to initiate emergency reinforcement work, including adding micropiles or anchor cables for reinforcement.

[0208] On-site inspectors, based on the checklist, discovered that the slope drainage ditch was obstructed by accumulated gravel, leading to rainwater infiltration and localized soil saturation creep. After implementing drainage dredging and temporary reinforcement measures, the displacement rate stabilized. The system then analyzed the multi-dimensional feature vector of this warning. and business semantic vector Data such as the handling results are stored in the historical early warning module to complete the knowledge loop.

[0209] 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 modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An intelligent agent and multi-feature fusion-based railway monitoring and early warning system, characterized in that, include: The data acquisition module is used to acquire monitoring data from field sensors in real time and perform preliminary screening based on preset fixed threshold logic. When the monitoring data exceeds the threshold, a preliminary alarm signal is triggered, and the historical data sequence of abnormal sensors and surrounding related devices within a specified time period is captured simultaneously. The feature processing module is used to transform the historical data sequence into a multidimensional feature vector representing the evolution trend of the monitored object. And transform the unstructured information related to the anomaly sensors into business semantic vectors. ; a history alert module for persistently storing history alert instances and their corresponding multi-dimensional feature vectors , business semantic vectors and validating effective treatment plans; The knowledge retrieval module is used to calculate the multidimensional feature vector in the historical early warning module based on a dual-path parallel retrieval algorithm. Spatial distance similarity and the business semantic vector The semantic similarity is calculated, and a comprehensive similarity score is output based on a weighted fusion algorithm to retrieve related historical cases; The intelligent agent module, with a large language model as its core, is used to perform time-series analysis, causal reasoning, and authenticity verification on the current anomaly based on historical cases retrieved by the knowledge retrieval module and in combination with preset business prompt word templates, and autonomously generate structured risk investigation suggestions and handling plans. The early warning and handling module, as a human-computer interaction interface, is used to push the handling plan generated by the intelligent agent module to the processing terminal, and to receive the verification results and handling feedback from the processing terminal. The intelligent agent module is also used to automatically summarize and structure the current alarm instance based on the handling feedback, and store it in the historical early warning module to realize the iterative update of knowledge.

2. The railway monitoring and early warning system based on intelligent agent and multi-feature fusion according to claim 1, characterized in that, The multi-dimensional feature vector comprises: The trend derivative feature is used to characterize the average slope and maximum abrupt change of the monitored data over a time period. Signal morphology features are used to characterize the impact morphology and energy distribution of monitoring data.

3. A railway monitoring and early warning system based on intelligent agent and multi-feature fusion according to claim 2, characterized in that: The trend derivative feature includes the average slope feature. and the largest mutation feature The signal morphological features include the pulse factor. and waveform factor The ; in: ； ； This represents the monitored value at the current sampling time; This indicates the monitoring value at the previous sampling time. This reflects the rate of change of the signal within a unit sampling period; ； ； It represents the change in the first-order difference, reflecting the curvature or the degree of abrupt change in the monitored numerical waveform; ； ； is the maximum value of the absolute values of the signal within the current window; is the average value of the absolute values of the signal within the window, is the root mean square value, reflecting the average energy level of the signal; 。 4. The railway monitoring and early warning system based on agent and multi-feature fusion according to claim 1, characterized in that: The preset business prompt word templates of the intelligent agent module include: The system role setting field is used to define the identity of the intelligent agent; The environmental context field is used to input the device information, installation location, current weather, and historical risk factors for the abnormal sensor. The monitoring data description field is used to input time-series data from abnormal sensors and surrounding related sensors; The reasoning logic requires fields that are used to make causal inferences based on morphological analysis, environmental correlation analysis, and spatial consistency check analysis. The output format specification field is used to limit the agent's output of judgment results, confidence levels, and inference processes in JSON format.

5. The railway monitoring and early warning system based on agent and multi-feature fusion according to claim 1, characterized in that: The historical early warning module adopts a hybrid storage mode of relational database and vector database; the relational database is used to store knowledge information summarized by the agent module, including agent judgment conclusions and structured summaries; the vector database is used to store the multidimensional feature vectors. and the business semantic vector . 6.The railway monitoring and early warning system based on agent and multi-feature fusion according to claim 1, characterized in that: The knowledge retrieval module employs a dual-path parallel retrieval algorithm, utilizing the multi-dimensional feature vectors. Spatial distance similarity and the business semantic vector The weighted fusion of semantic similarity enables a three-dimensional calibration of physical evolution mechanisms and business context; wherein, the multidimensional feature vector The business semantic vector is used to characterize the evolution trend of monitoring indicators. External boundary conditions used to reflect early warning events.

7. A railway monitoring and early warning method based on agent and multi-feature fusion, applying the railway monitoring and early warning system based on agent and multi-feature fusion according to any one of claims 1-6, characterized in that, Includes the following steps: Step S1: Acquire monitoring data from field sensors in real time, perform preliminary screening based on preset fixed threshold logic, trigger a preliminary alarm signal when the monitoring data exceeds the threshold, and simultaneously capture historical data sequences of abnormal sensors and surrounding related devices within a specified time period. Step S2: Based on the preset prompt word template, the intelligent agent module performs time-series analysis and logical reasoning on the monitoring data of the abnormal sensor and surrounding sensors, combined with the characteristics of the surrounding environment and the types of risks that may occur, to determine whether the alarm signal is a real physical anomaly. Step S3: Feature processing is performed on the sensor data determined as real physical anomaly, change trend and morphological features are extracted, and a multi-dimensional feature vector of monitoring data is constructed ; at the same time, data such as basic information of abnormal sensor, geographical position, weather condition, etc. are acquired, semantic text is constructed, and vectorization is performed through an Embedding model to obtain a business semantic vector ; Step S4: The knowledge retrieval module uses a dual-path parallel retrieval algorithm to calculate the multi-dimensional feature vector in the historical early warning module. Spatial distance and the business semantic vector Based on the semantic similarity, a comprehensive similarity score is output using a weighted fusion algorithm to retrieve related historical cases; Step S5: Based on the retrieved historical cases, the intelligent agent module generates a structured risk assessment list and handling plan, and pushes it to the processing terminal; Step S6: Receive the on-site verification results and handling feedback from the processing terminal, and process the multi-dimensional feature vector of the current alarm instance. Business semantic vector The early warning situation and the results of the handling process are stored in the historical early warning module.

8. The railway monitoring and early warning method based on agent and multi-feature fusion according to claim 7, characterized in that: In step S2, the specific steps of the intelligent agent module performing temporal analysis and logical reasoning based on the preset prompt word template include: Morphological analysis: Determine whether the data curve is a transient, impulsive change or a smooth, continuous evolution; Environmental correlation analysis: Based on the current weather, determine whether the abnormal fluctuations can be attributed to environmental factors; Spatial consistency check and analysis: By comparing the monitoring data of surrounding sensors, determine whether there are linkage characteristics that conform to physical laws; Causal inference: Based on the above analysis results, it is determined that the signal attribute is a real anomaly or a false alarm caused by environmental interference. 9.The railway monitoring and early warning method based on the agent and the multi-feature fusion of claim 7, characterized in that, In step S4, the knowledge retrieval module employs a dual-path parallel retrieval algorithm and calculates the comprehensive similarity score using a weighted fusion algorithm. ； in, The similarity between the business semantic vector and the semantic vector in the historical early warning module. The similarity between the multidimensional feature vector and the feature vector in the historical early warning module. and The preset weighting coefficients, and .

10. The railway monitoring and early warning method based on agent and multi-feature fusion according to claim 7, characterized in that, In step S6, the specific steps for storing the alarm instance in the historical early warning module include: The intelligent agent module automatically summarizes alarm instances and handling results to generate structured knowledge. Store structured knowledge in a knowledge record table of a relational database; The multi-dimensional feature vector and the business semantic vector is stored in the early warning vector index table of the vector database and is associated by the knowledge record ID.

Images