Turnout state abnormity early warning method and system based on coding and large language model

Abstract

The invention discloses a coding and large language model-based turnout state abnormity early warning method and system, and belongs to the technical field of intelligent operation and maintenance and fault prediction of rail transit, and the method comprises the steps: building a time sequence curve of multi-modal physical parameters in a turnout conversion process through collecting the multi-modal physical parameters in the turnout conversion process; comprising a conversion resistance curve of the switch machine and a stroke displacement curve of the switch rail; performing feature extraction of a switch machine conversion resistance curve and a switch rail stroke displacement curve on the turnout conversion process according to the change features of the time sequence curve, and inputting the extracted features into a natural language template together to generate a structured language description fusing conversion resistance and stroke displacement for querying a turnout conversion knowledge base; fault maintenance knowledge fragments related to turnout state abnormity are found; and through structured language description and related knowledge fragments, cue words are constructed, and turnout states are evaluated based on a large language model to diagnose abnormalities and generate early warnings.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance and fault prediction technology of rail transit, specifically involving a method and system for early warning of abnormal turnout status based on coding and large language model. Background Technology

[0002] Turnouts are one of the most critical and vulnerable links in railway lines, and their reliability is directly related to traffic safety. Currently, the maintenance mode of turnouts is shifting from the traditional periodic "planned maintenance" to a more proactive "condition-based maintenance".

[0003] The following are the main problems and defects in the maintenance of turnouts in the existing technology:

[0004] 1. Limited data utilization: Most turnout monitoring systems rely solely on single parameter curves such as switch machine current and power for analysis, lacking collaborative perception and correlation analysis of multi-dimensional forces (such as switching resistance) and mechanical actions (such as switch rail stroke) throughout the entire switching process;

[0005] 2. Limited level of diagnostic intelligence: Existing diagnostic methods mostly rely on preset fixed thresholds or simple rule bases, making it difficult to identify complex and compound abnormal patterns. They are not sensitive to early and minor performance degradation (sub-health state), have a high false alarm rate, and cannot provide in-depth fault cause analysis and maintenance guidance.

[0006] 3. Knowledge dependence and poor interpretability: Traditional algorithm models are like "black boxes". Their decision-making process lacks transparency, relies heavily on expert experience for later interpretation, and cannot effectively utilize and integrate existing and constantly updated massive amounts of professional knowledge (such as maintenance manuals and historical case libraries).

[0007] Therefore, there is an urgent need in this field for an intelligent early warning solution that can deeply integrate multi-source monitoring data, possess advanced cognitive reasoning capabilities, and provide interpretable diagnostic results. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies and achieve early, accurate, and interpretable early warning of turnout status anomalies, this invention adopts the following technical solution:

[0009] The turnout status anomaly early warning method based on encoding and large language model includes the following steps:

[0010] Multimodal physical parameters of the turnout switching process are collected to construct time-series curves of the changes in multimodal physical parameters during the turnout switching process. The time-series curves include the switching resistance curve of the switch machine and the dynamic displacement curve of the switch rail.

[0011] Based on the changing characteristics of the time-series curves, features of the switch machine switching resistance curve and the switch rail travel displacement curve are extracted during the turnout switching process. The extracted features are then input into a predefined natural language template to generate a structured language description that integrates the switching resistance and travel displacement. Through multimodal fusion analysis of the resistance curve and travel curve, latent faults and early performance degradation trends that cannot be reflected by a single parameter can be captured, thereby achieving true "predictive maintenance".

[0012] A turnout switching knowledge base is constructed. Based on retrieval enhancement generation, the knowledge base is queried through structured language description to find a set of fault maintenance knowledge fragments related to turnout status anomalies, thereby constructing the association between natural language descriptions of multimodal physical parameters and fault targeting and maintenance knowledge.

[0013] By using structured language descriptions and retrieved relevant knowledge fragments, prompt words are constructed. Based on a large language model, the state of the turnout is evaluated to diagnose anomalies and generate early warnings. This enables the large language model to not only provide early warnings and analyze the causes of failures for early and minor abnormal data, but also to further provide key monitoring points, potential risk evolution directions, and possible intervention strategies.

[0014] Furthermore, a complete turnout switching process is divided into multiple stages, each corresponding to a specific timing curve. Turnout information and reference curves for each stage are preset, and the timing curves for each stage are compared with the reference curves. The turnout information is combined with the generation of structured language descriptions to assist in the generation of the descriptions. This transforms complex timing curve data into structured language descriptions that can be understood by both machines and humans, improving the interpretability of complex or compound anomalies.

[0015] Furthermore, the extracted features of each stage of the switching resistance curve include the slope of the switching resistance curve and the shaking amplitude of the switch machine; the slope is obtained by obtaining a time series based on the resistance sequence of the switching resistance curve, calculating the resistance slope through the resistance values ​​of adjacent times in the time series, and describing the shaking amplitude of the switch machine through the standard deviation of the resistance slope sequence.

[0016] Furthermore, the extracted features of each stage of the conversion resistance curve also include the peak value, mean, standard deviation, rising / falling slope, and specific shape of the conversion resistance curve. The specific shape includes, but is not limited to, the amplitude of the jitter and the duration.

[0017] Furthermore, the features extracted from each stage of the dynamic displacement curve include the synchronicity of the switch rail being pulled by multiple points. The synchronicity is achieved by downsampling the dynamic curves of two different traction positions to obtain two sets of sequence values ​​for different positions. The sequence value of one position in the current group is subtracted from the sequence values ​​of all positions in the other group, and the second norm is calculated. The smallest second norm value is taken as the second norm value corresponding to the sequence value of that position in the current group. Similarly, the sequence value of one position in the other group is subtracted from the sequence values ​​of all positions in the current group, and the second norm is calculated. The smallest second norm value is taken as the second norm value corresponding to the sequence value of that position in the other group. The largest second norm value in the current group and the other group is taken to describe the synchronicity of the two different traction positions.

[0018] Furthermore, the features extracted from each stage of the dynamic displacement curve also include displacement amount and conversion speed.

[0019] Furthermore, the stages include before unlocking, unlocking, switching, locking, and after locking, and the turnout information includes, but is not limited to, the rated force of the switch machine.

[0020] Furthermore, the multimodal physical parameters also include auxiliary parameters for feature extraction, including but not limited to current, voltage, vibration acceleration, and rail temperature.

[0021] Furthermore, the collected multimodal physical parameters are preprocessed, including but not limited to timestamp alignment, filtering and noise reduction, and outlier removal.

[0022] The turnout status anomaly early warning system based on coding and a large language model includes a switch machine switching resistance sensor, a switch rail travel displacement sensor, a time-series curve generation module, a structured language description generation module, a turnout switching knowledge database, and a large prediction model module. Employing the aforementioned turnout status anomaly early warning method based on coding and a large language model, the system generates a time-series curve containing switching resistance and travel displacement curves through sensor data acquisition. Curve features are extracted to generate a structured language description. This description is then used to query the database for turnout switching knowledge. The large language model module, based on the structured language description and turnout switching knowledge, constructs prompts to assess the turnout status, diagnose anomalies, and generate early warnings.

[0023] The advantages and beneficial effects of this invention are as follows:

[0024] The present invention relates to a turnout condition anomaly early warning method and system based on coding and a large language model. Through multimodal fusion analysis of resistance and stroke curves, it can capture latent faults and early performance degradation trends that cannot be reflected by a single parameter. Its perception is more comprehensive and its early warning is earlier, achieving true predictive maintenance. Utilizing the cognitive capabilities of the large language model, it breaks through the limitations of traditional rule bases, can handle complex anomalies, and its diagnosis is more intelligent and the results are interpretable. Through retrieval-enhanced generation, it can easily integrate and update the latest professional knowledge and cases, enabling the diagnostic capabilities to continuously evolve with the enrichment of the knowledge base, achieving self-learning and avoiding the high cost of model retraining. Through accurate early warning, it avoids the expansion of faults and operational interruptions. Condition-based maintenance optimizes the allocation of maintenance resources, thereby effectively reducing operation and maintenance costs and ensuring traffic safety. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the system structure in an embodiment of the present invention.

[0026] Figure 2 This is a flowchart of the method in an embodiment of the present invention.

[0027] Figure 3 This is a five-stage resistance curve diagram of the turnout switching process in an embodiment of the present invention.

[0028] Figure 4 This is a graph showing the motion curve of the turnout switching process in an embodiment of the present invention. Detailed Implementation

[0029] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0030] like Figure 1 As shown, this invention proposes a turnout status anomaly early warning system based on coding and a large language model, including a switch machine switching resistance sensor, a switch rail dynamic displacement sensor, a data preprocessing module, a time-series curve generation module, a structured language description generation module, a turnout switching knowledge database, and a large prediction model module.

[0031] The switch machine switching resistance sensor is a patch-type strain gauge located on the action connecting rod, used to collect the switching resistance of the switch machine in real time in order to construct the switching resistance curve of the switch machine.

[0032] The switch rail dynamic displacement sensor is a magnetostrictive sensor used to collect the dynamic displacement of the switch rail in real time in order to construct the dynamic displacement curve of the switch rail.

[0033] The data preprocessing module is used to perform preprocessing on the collected raw data, such as timestamp alignment, filtering and noise reduction, and outlier removal.

[0034] The timing curve generation module collects multimodal physical parameters during the turnout switching process to construct timing curves of the changes in multimodal physical parameters during the turnout switching process. The timing curves include the switching resistance curve of the switch machine and the dynamic displacement curve of the switch rail.

[0035] The structured language description generation module extracts features from the switch machine switching resistance curve and the switch rail displacement curve during the turnout switching process based on the changing characteristics of the timing curves. The extracted features are then input into a predefined natural language template to generate a structured language description that integrates the switching resistance and displacement. A complete turnout switching process is divided into multiple stages, each corresponding to a specific timing curve. Turnout information and reference curves for each stage are preset, and the timing curves of each stage are compared with the reference curves. The turnout information is then used to assist in the generation of the structured language description.

[0036] The turnout switching knowledge database employs enhanced retrieval generation to enable structured language descriptions of database queries, finding a set of fault maintenance knowledge fragments related to turnout status anomalies.

[0037] The Big Prophecy Model module constructs prompt words by describing them in structured language and retrieving relevant knowledge fragments. Based on the Big Language Model, it evaluates the switch status to diagnose anomalies and generate early warnings.

[0038] like Figure 2 As shown, this invention also proposes a turnout status anomaly early warning method based on encoding and large language models. Combining signal processing, natural language processing, and large language models, it performs turnout equipment status monitoring and early anomaly warning, specifically including the following steps:

[0039] Step S1: Real-time acquisition and preprocessing of multimodal physical parameters during turnout switching process to construct time-series curves of multimodal physical parameter changes during turnout switching process.

[0040] like Figure 3 , Figure 4 As shown, a high-precision sensor network is deployed at key parts of the turnout to collect the switching resistance curve of the switch machine (the patch strain gauge located on the action connecting rod) and the dynamic displacement curve of the switch rail (based on a magnetostrictive sensor) in real time. At the same time, auxiliary parameters such as current, voltage, vibration acceleration, and rail temperature can be collected, thereby realizing the fine perception and feature extraction of multi-dimensional physical parameters (resistance, dynamic displacement, etc.). Furthermore, the collected raw data is preprocessed by timestamp alignment, filtering and noise reduction, and outlier removal. Figure 3 The test was conducted on September 18, 2025 at 17:08:12, at a temperature of 30.0℃, with a pull-in force (i.e., the maximum resistance value of 5070.0N).

[0041] Step S2: Based on the changing characteristics of the time-series curve, the turnout switching process is divided into stages and the characteristic parameters of each stage are extracted. The characteristic parameters are input into a predefined natural language template to generate a structured language description. Turnout information and reference curves for each stage are preset, and the time-series curves of each stage are compared with the reference curves. The turnout information is combined to assist in the generation of the structured language description, thereby transforming the complex time-series curve data into a structured language description that can be understood by both machines and humans.

[0042] Specifically, based on the characteristics of the preset information, the language encoding is quantified as follows:

[0043] 1. Stage division: Based on the changing characteristics of the resistance and stroke curves, a complete turnout switching process is divided into five typical stages: before unlocking, unlocking, switching, locking, and after locking.

[0044] 2. Preset turnout information and reference curve: The turnout information (such as the rated force of the switch machine) and reference curve are used to generate auxiliary language descriptions.

[0045] For example, exceeding the rated force by 20% results in a deviation of 500N from the reference curve.

[0046] 3. Feature Extraction and Quantization: Extract key feature parameters from the two curves at each stage. For example:

[0047] Resistance curve: peak value, mean, standard deviation, rising / falling slope, and specific shape (such as oscillation amplitude and duration) for each stage; among which, peak value, mean, and standard deviation are easy to obtain, while for the slope, this invention uses a quantized differential method to obtain the slope sequence, specifically:

[0048] The original resistance sequence was downsampled to obtain the time series:

[0049]

[0050] Calculate the slope of each element in the time series:

[0051]

[0052] in, Indicates a point in time The sampled resistance value, while the jitter amplitude is obtained through the resistance slope sequence. It is expressed as the standard deviation.

[0053] Stroke curve: displacement, conversion speed, and synchronicity at each stage (for multi-point traction); among these, displacement and conversion speed are readily available, while synchronicity is calculated using the following method:

[0054] camera position With camera position Synchronization through The description and formula are as follows:

[0055]

[0056]

[0057]

[0058] Where x represents the curve based on position n. The downsampling obtained camera position sequence value, y represents the curve based on camera position m. The downsampling results in a sequence of machine positions. ||·|| represents the norm operation, min represents the minimum value operation, and max represents the maximum value operation.

[0059] 4. Language Description Generation: The extracted quantitative features are filled into a predefined natural language template to generate a structured language description that integrates resistance and stroke information. For example: During the locking phase, the peak resistance is 5800N, which is 15% higher than the benchmark value; at the same time, the stroke curve shows a 0.5-second delay in locking.

[0060] By using multimodal fusion analysis of resistance and stroke curves, latent faults and early performance degradation trends that cannot be reflected by a single parameter can be captured, thus achieving true "predictive maintenance".

[0061] Step S3: Construct a turnout conversion knowledge base, and use structured language to describe and retrieve the knowledge base. Based on Retrieval-Augmented Generation (RAG), obtain a set of relevant knowledge fragments.

[0062] 1. Knowledge Base Construction: Unstructured text data such as the structural principles of turnouts, maintenance manuals, historical failure cases, and expert experience rules are digitized and then stored in a vector database after text segmentation and vectorization to construct a professional knowledge base for turnouts.

[0063] 2. Enhanced Generation Retrieval: After generating the linguistic description of the current curve in step S2, this description is used as a query request to perform a similarity search in the vector database, identifying the most relevant knowledge fragments (such as fault descriptions and maintenance records corresponding to similar curve shapes). Through the RAG architecture, the system can easily integrate and update the latest professional knowledge and cases, enabling diagnostic capabilities to continuously evolve as the knowledge base grows, avoiding the high cost of model retraining.

[0064] Step S4: Construct prompt words using structured language descriptions and retrieved relevant knowledge fragments, and evaluate the turnout status based on a large language model to diagnose anomalies and generate early warnings.

[0065] Specifically, the language description generated in step S2 and the relevant knowledge fragments retrieved in step S3 are combined into an enhanced prompt word using a carefully designed prompt word template, and then input into the large language model.

[0066] Leveraging the powerful knowledge processing and reasoning capabilities of large language models, it breaks through the limitations of traditional rule bases, enabling it to handle complex anomalies. Combined with professional domain knowledge, it can achieve intelligent diagnosis of abnormal states, root cause analysis, and provide maintenance suggestions. Its diagnostic process is similar to expert reasoning, and the generated reports are easy to understand, greatly improving the credibility and usability of the results.

[0067] Specifically, based on its powerful natural language understanding and logical reasoning capabilities, the large language model accomplishes the following tasks:

[0068] Status assessment: Determine whether the turnout is currently in a "healthy", "sub-healthy (abnormal)" or "faulty" state.

[0069] Abnormal diagnosis: Analyze the possible causes of the abnormality (e.g., "the slide plate may cause a resistance spike in the locking section").

[0070] Warnings and Recommendations: Generate a comprehensive diagnostic report that includes warning level, description of specific phenomena, possible cause analysis, and maintenance guidance recommendations.

[0071] Early and accurate warnings can prevent the escalation of faults and operational interruptions, while condition-based maintenance optimizes the allocation of maintenance resources, thereby effectively reducing the total lifecycle maintenance cost and improving the safety and reliability of rail transit systems.

[0072] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for abnormal warning of turnout state based on coding and large language model, characterized in that The method comprises the following steps: Collecting multi-modal physical parameters of the switch conversion process to construct time sequence curves of multi-modal physical parameters in the switch conversion process, the time sequence curves including a conversion resistance curve of the switch machine and a travel displacement curve of the frog; According to the change characteristics of the time sequence curves, extracting features of the conversion resistance curve and the travel displacement curve of the switch conversion process, and inputting the extracted features into a natural language template to generate a structured language description fusing the conversion resistance and the travel displacement; Constructing a switch conversion knowledge base, and querying the knowledge base through the structured language description based on retrieval enhancement generation to find a group of fault maintenance knowledge fragments related to the abnormal switch state; Based on the structured language description and the retrieved related knowledge fragments, constructing prompt words, and evaluating the switch state based on a large language model to diagnose the abnormality and generate a warning.

2. The method for warning of turnout state abnormality based on coding and large language model according to claim 1, characterized in that: A complete switch conversion process is divided into multiple stages, each stage corresponding to a corresponding time sequence curve; presetting switch information of each stage and a reference curve, and comparing the divided time sequence curves of each stage with the reference curve to assist the generation of the structured language description.

3. The method of Claim 2, wherein the method is characterized in that: The extracted features of each stage of the conversion resistance curve include the slope of the conversion resistance curve and the jitter amplitude of the switch machine; The slope is obtained based on a resistance sequence of the conversion resistance curve, the resistance slope is calculated through the resistance values of adjacent times in the time sequence, and the jitter amplitude of the switch machine is described through the standard deviation of the resistance slope sequence.

4. The method of Claim 2, wherein the method is characterized by: The extracted features of each stage of the conversion resistance curve also include the peak value, the mean value, the standard deviation, the rising / falling slope and the specific form of the conversion resistance curve, and the specific form includes but is not limited to the jitter amplitude and the duration.

5. The method of Claim 2, wherein the method is characterized by: The extracted features of each stage of the travel displacement curve include the synchronization of the frog being pulled by multiple points; the synchronization is obtained by performing downsampling operation on the travel curves of two different traction positions to obtain two groups of sequence values of different positions, calculating the second norm by respectively subtracting the sequence value of one position in the current group from all sequence values of positions in the other group, taking the minimum second norm value as the second norm value corresponding to the sequence value of the position in the current group, and similarly, calculating the second norm by respectively subtracting the sequence value of one position in the other group from all sequence values of positions in the current group, taking the minimum second norm value as the second norm value corresponding to the sequence value of the position in the other group, and taking the maximum second norm value in the current group and the other group to describe the synchronization of the two different traction positions.

6. The method of Claim 2, wherein the method is characterized by: The extracted features of each stage of the travel displacement curve also include the displacement and the conversion speed.

7. The method of Claim 2, wherein the method is characterized by: The stages include before unlocking, unlocking, conversion, locking, and after locking.

8. The method of Claim 1, wherein the method is based on encoding and large language model for turnout state anomaly early warning. The multi-modal physical parameters also include auxiliary parameters for feature extraction, and the auxiliary parameters include but are not limited to current, voltage, vibration acceleration and track temperature.

9. The method of Claim 1, wherein the method is characterized by: The collected multi-modal physical parameters are preprocessed, and the preprocessing includes but is not limited to timestamp alignment, filtering and denoising, and outlier rejection.

10. A turnout state anomaly early warning system based on coding and large language model, comprising a switch machine conversion resistance sensor and a point travel displacement sensor, characterized in that: The system further comprises a time curve generation module, a structured language description generation module, a turnout conversion knowledge database, a big prophecy model module, adopts the turnout state abnormal early warning method based on coding and a large language model in any one of claims 1 to 9, generates a time curve containing a conversion resistance curve and a moving distance curve through the collection of a sensor, extracts curve features to generate a structured language description, queries the database through the structured language description to obtain turnout conversion knowledge, and a large language model module based on the prompt word constructed by the structured language description and the turnout conversion knowledge evaluates the turnout state to diagnose the abnormality and generates the early warning.

Images