Track detection data analysis method and system based on knowledge graph

Abstract

The invention discloses a track detection data analysis method and system based on a knowledge graph, and relates to the technical field of knowledge graphs, and the method comprises the steps: collecting and aligning multi-source heterogeneous data such as a track vibration sensor, a surface image, a work order text and the like, carrying out section division in combination with consistency and relevance, and distributing credibility weights for various evidences; and establishing a knowledge graph, generating a disturbance factor by using time sequence convolution, and outputting probability distribution and an index prediction value of an orbit risk type in combination with spatial topology and rule matching degree propagation. Attention weight is generated through anti-fact evaluation, risk cost is defined through Bayesian coincidence probability, and finally attention redistribution is optimized through genetic iteration, so that the result is converged to the minimum risk cost. The problems of difficulty in multi-source data fusion, insufficient prediction accuracy and lack of interpretability of results can be effectively solved, the accuracy, stability and interpretability of track defect analysis are improved, and reliable support is provided for operation and maintenance decision making.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to a method and system for analyzing orbit detection data based on knowledge graphs. Background Technology

[0002] During the long-term operation of rail transit, rails, sleepers, and connecting components are susceptible to multiple effects from vibration, environmental erosion, and train loads, leading to potential hazards such as cracks, spalling, and structural loosening. To ensure operational safety, current track inspection methods primarily rely on single sensors or manual inspections, but these methods have the following shortcomings:

[0003] First, multi-source data such as vibration signals, image sampling, and maintenance work orders are difficult to process uniformly. The spatiotemporal scales and recording methods of different sources vary significantly, making fusion analysis difficult. Second, most existing methods are based on independent statistical or deep models, which cannot fully utilize the correlation between multi-source data and make it difficult to build a holistic understanding across time periods and modalities. Third, the formation and evolution of defects are influenced by both time series and spatial topology. Traditional methods often ignore the spatiotemporal propagation laws, resulting in insufficient prediction stability and accuracy. Fourth, in risk management and operation and maintenance decisions, users need to understand "which type of evidence is most critical and which adjustment can reduce risk," but existing black-box models lack interpretability and are difficult to support practical applications. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a knowledge graph-based method for analyzing track detection data to address the problems of difficulty in unifying and integrating multi-source heterogeneous data, insufficient accuracy in predicting defect evolution processes, and lack of interpretability of results in track detection.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a knowledge graph-based method for analyzing orbit detection data, which includes: acquiring multi-source heterogeneous data from an orbit detection system; aligning the multi-source heterogeneous data and dividing it into segments; and assigning credibility weights to each data piece of evidence; and simultaneously performing binary analysis on the divided segments.

[0008] A knowledge graph is generated by constructing a knowledge graph with the state distribution of the evidence-corresponding identification results as nodes. Temporal and topological data are integrated on the knowledge graph to perform spatiotemporal prediction, and the prediction results are output: the probability distribution of risk types and the index prediction value of each risk type.

[0009] The attention weight of each evidence or perturbation factor is generated by counterfactual evaluation; and the risk cost of the prediction result of each site is generated by the rule detection result;

[0010] The quantification result is fed back to the prediction result, and genetic iteration optimization is performed to generate a final analysis result;

[0011] The two tuples are centered on a probability distribution; and the two tuples include a section recognition result and a rule detection result, and represent a probability distribution of an abnormal entity recognition result and a probability distribution of a theoretical influence result of an abnormal entity, respectively.

[0012] As a preferred scheme of the track detection data analysis method based on a knowledge graph, the multi-source heterogeneous data includes vibration data collected by a vibration sensor arranged on a track, surface images collected by a camera, and work order texts generated by maintenance;

[0013] The vibration data and the surface images are positioned on the track through positioning information of the vibration sensor and positioning information during image acquisition; and the work order texts are positioned on the track through historical records of flaw detection data and maintenance data under a section division result, and time stamps of the flaw detection data and the maintenance data are marked;

[0014] Based on the positioning result, the data evidence of each site on the track is generated; wherein, based on different characteristics of section division in each record, the section in the record is updated using the latest historical record, so that the original data of the area outside the work order is retained;

[0015] Real-time features on each site i: ;

[0016] wherein, represents real-time vibration data on site i, represents real-time surface images on site i, represents the latest flaw detection data on site i, represents the latest maintenance data on site i, represents the time stamp of the real-time vibration data on site i, represents the time stamp of the real-time surface images on site i, represents the time stamp of the latest flaw detection data on site i, represents the time stamp of the latest maintenance data on site i.

[0017] As a preferred scheme of the track detection data analysis method based on a knowledge graph, the segment division includes dividing each type of data as an analysis dimension through analysis of data consistency and correlation between a previous sampling period and a current sampling period based on multi-source heterogeneous data.

[0018] Step one: in a single dimension, analyze the consistency of data of each site within a period, if the volatility of the data is less than the corresponding threshold, the site is determined not to be processed, if the volatility of the data is not less than the corresponding threshold, the site is determined as a candidate site;

[0019] Step two: on the track of each dimension, the candidate site is taken as a seed, and growth is performed along the track, if the adjacent site is also a candidate site or the correlation value of the adjacent site is higher than the evidence value of the site itself, then one growth is realized, otherwise, the growth is stopped, and a segment division result in each dimension is generated;

[0020] The correlation value is a value analysis realized by an LSTM neural network, which is used to represent the support degree of data for the detection result within the interval of the current dimension in the analysis dimension, and the support degree of each detection result is weighted and summed by taking the probability of each detection result as a weight to obtain the correlation value.

[0021] The evidence value itself is represented by two states, and the maximum value is obtained: state 1: fixed threshold; state 2: calculating the correlation value of the site and the first seed in the opposite direction;

[0022] The credibility weight includes taking the segment division result in each dimension as an evidence, and taking the difference between the maximum and minimum probabilities of the detection result in each evidence as the credibility weight of the evidence.

[0023] As a preferred scheme of the track detection data analysis method based on a knowledge graph, the analysis of the two-tuple includes identifying the evidence in each dimension by using an intelligent agent, generating a state distribution of the segment identification result according to the probability distribution of the segment identification result of each evidence based on the probability distribution of each evidence;

[0024] And generating a state distribution of the two-tuple according to the rule detection result corresponding to each segment identification result;

[0025] Each rule detection result corresponds to a different segment identification result.

[0026] According to historical records, the state distribution of the corresponding recognition result of the evidence of the same site in different dimensions is taken as a node, the relationship between the nodes is learned to obtain the knowledge graph; and when all edges perform information propagation, the credibility weight held by the starting point evidence is multiplied by the information transmittable amount.

[0027] As a preferred scheme of the track detection data analysis method based on the knowledge graph, the intelligent agent is trained by taking a time sequence convolutional neural network as a model; when outputting the recognition result, a perturbation factor is generated for each dimension of the evidence according to time sequence analysis, which is used to represent the influence of the feature of the evidence on other sites in the time sequence;

[0028] The perturbation factor includes a probability distribution of the recognition result when the evidence section is used to identify other sites;

[0029] The space-time prediction includes, using the spatial topological correlation, the two-tuple of the perturbation factor or the evidence is attenuated in the space dimension of the edge and then propagated;

[0030] After each evidence completes the propagation, the state distribution of the corresponding recognition result of the two-tuple at each site is multiplied by the credibility weight, and the result of the perturbation factor after weighting is multiplied by the credibility weight of the corresponding source; after inputting the knowledge graph, the prediction result is obtained;

[0031] The spatial topological correlation includes, when analyzing the correlation of any site r in interval j, if r is in j, the perturbation factor is completely lost, and the two-tuple corresponding to the section j is taken as the evidence element of r in the dimension of the section; if r is outside j, the two-tuple corresponding to the section j is completely lost, and the perturbation factor is propagated:

[0032] When propagating, the matching degree of the rule detection result and the perturbation factor in the two-tuple of each dimension at each site on the propagation path is calculated, the matching degree is multiplied by a standard attenuation rate to obtain the attenuation rate of each site in each dimension on the propagation path; the actual attenuation rate of each site on the propagation path is obtained by superimposing the attenuation rate of each dimension;

[0033] The actual attenuation rate is integrated on the path to obtain the propagation result of the perturbation factor of interval j on any site r; the actual attenuation rate of the perturbation factor at r is taken as a weight coefficient of the influence, and the perturbation factor after propagation and attenuation is weighted.

[0034] As a preferred scheme of the track detection data analysis method based on the knowledge graph, the counterfactual evaluation includes removing any evidence or disturbance factor in the target site to obtain a counterfactual prediction result; the counterfactual evaluation is compared with an actual prediction result, and differences between the two are used as importance basis; after calculating all importance bases on the site, the attention weight of each evidence or disturbance factor is obtained.

[0035] The risk cost includes the consistency of the rule detection result in the two-tuple of the prediction result on the analysis site and the evidence where the site is located, which is defined as:

[0036]

[0037] Where k represents the two-tuple index on the site r; K represents the number of two-tuples on the site r; represents the prediction result on the site r; represents the rule detection result of the kth two-tuple on the site r; represents the analysis function of the consistency, which is realized by a pre-trained Bayesian model, and the consistency is analyzed by using the Bayesian model to generate a consistency probability after encoding two input contents by an encoder.

[0038] As a preferred scheme of the track detection data analysis method based on the knowledge graph, the optimization of the genetic iteration includes attention re-distribution of the input of the knowledge graph.

[0039] The original attention of the input of the knowledge graph is: the weight corresponding to each evidence or disturbance factor, which is the result of normalization; the distribution of the original attention is used as an initial individual.

[0040] For all evidence or disturbance factors on the site, the attention weight is used to trigger the generation of new individuals in the genetic process, and the exchange or disturbance of the local weight distribution between individuals is realized by selection / crossover / variation in the new individuals to explore new schemes and generate a set of results of attention re-distribution of the evidence or disturbance factors.

[0041] In the result of the attention re-distribution, the increase proportion of the attention is any value from 0 to L;

[0042] The minimum risk cost is used as an objective function for optimization; the final attention distribution result is obtained by iterating to the maximum number of times or the risk cost no longer decreases, and the final analysis result is generated by using the knowledge graph according to the result of the attention re-distribution.

[0043] Secondly, the present invention provides a knowledge graph-based track detection data analysis system, including a data acquisition unit that acquires multi-source heterogeneous data from a track detection system, aligns the multi-source heterogeneous data and divides it into segments, and assigns credibility weights to each data piece of evidence; at the same time, it performs binary analysis on the divided segments.

[0044] The prediction unit constructs a knowledge graph with the state distribution of the evidence-corresponding identification results as nodes. It integrates temporal and topological data on the knowledge graph to perform spatiotemporal prediction, corrects for physical constraints, and outputs the prediction results: the probability distribution of risk types and the predicted index value of each risk type.

[0045] The computational unit generates the attention weight for each piece of evidence or perturbation factor through counterfactual evaluation; and generates the risk cost of the prediction result for each site based on the rule detection results.

[0046] The output unit feeds back the quantification results to the prediction results, performs genetic iteration optimization, and generates the final analysis results.

[0047] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the knowledge graph-based orbit detection data analysis method as described in the first aspect of the present invention.

[0048] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the knowledge graph-based orbit detection data analysis method as described in the first aspect of the present invention.

[0049] The beneficial effects of this invention are as follows: By unifying and aligning vibration data, image information, and work order text during track detection, this invention establishes a binary knowledge graph centered on segment identification results and rule detection results, effectively solving the problem of difficult fusion of multi-source heterogeneous data. Temporal convolution is used to generate perturbation factors, and combined with track spatial topology propagation and rule matching attenuation mechanisms, making the prediction results more consistent with the true laws of defect evolution. Attention weights for each piece of evidence or perturbation factor are obtained through counterfactual evaluation, quantifying the importance of key factors and improving the interpretability and verifiability of the results. Furthermore, the risk cost is defined using Bayesian coincidence probability, and attention is redistributed through genetic iteration to gradually optimize the prediction structure, minimizing the risk cost globally and ensuring the stability and robustness of the prediction. This invention not only improves the accuracy of track defect detection and risk prediction but also provides clear decision-making basis for maintenance personnel, demonstrating promising engineering application prospects. Attached Figure Description

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0051] Figure 1 A flowchart of a track detection data analysis method based on a knowledge graph. DETAILED DESCRIPTION

[0052] In order to make the above objectives, features and advantages of the present application more apparent, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0053] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced without the specific details, other than those described herein, and it is understood that the present application will encompass these and other variations, modifications and alternatives.

[0054] Secondly, the "one embodiment" or "embodiment" referred to herein can include specific features, structures or characteristics contained in at least one implementation of the present application. The "in one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments.

[0055] Reference Figure 1 For one embodiment of the present application, the embodiment provides a track detection data analysis method based on a knowledge graph, comprising the following steps:

[0056] S1: Obtain multi-source heterogeneous data from a track detection system, align the multi-source heterogeneous data, and then perform section division, and assign a credibility weight to each data evidence; at the same time, analyze the divided sections for two-tuple.

[0057] The two tuple takes a probability distribution as the core; the two tuple includes a section identification result and a rule detection result; and the two tuple respectively represents a probability distribution of an identification result of an abnormal entity and a probability distribution of a theoretical influence result of the abnormal entity. Specifically, the section identification result is a result of intelligent identification of a certain section of a track through multi-source heterogeneous data (vibration signals, track surface images, flaw detection or maintenance records). The section identification result indicates whether the section has an abnormal entity, such as a crack, a falling block, a foreign object or a poor joint. The rule detection result is a result of deduction or prediction of a theoretical influence of the abnormal entity in track operation according to a preset detection rule or model (such as a safety specification, a structural mechanics calculation or an experience threshold). The rule detection result describes a possible consequence caused by the abnormal entity, such as increased vibration, accumulated structural fatigue or increased risk of derailment. Therefore, the two tuple not only describes a “detected abnormality”, but also expresses a “theoretical influence possibly caused by the abnormality”, thereby realizing mapping from a phenomenon to deduction.

[0058] Suppose that multi-source detection of a certain section i of a track obtains the following results: a section identification result: through vibration sensors and image analysis, it is found that the section has a track crack (an abnormal entity = a crack). A rule detection result: according to the crack width and position, combined with mechanical rules, it is deduced that the crack may cause an increase in the longitudinal vibration amplitude during train operation, thereby aggravating track fatigue. Then, the two tuple of the section can be expressed as: (crack identification, increase in longitudinal vibration amplitude).

[0059] For example, a section identification result: a fastener is found to be loose; a rule detection result: it may cause uneven settlement of the track and deterioration of the track geometry. The two tuple is expressed as: (loose fastener, increased risk of uneven settlement).

[0060] Further, the multi-source heterogeneous data includes vibration data collected by vibration sensors arranged on the track, surface images collected by cameras, and work order texts generated by maintenance. The vibration data and the surface images are positioned on the track through positioning information of the vibration sensors and positioning information during image collection. Flaw detection data and maintenance data under a section division result (in the historical division result, the division result in each record may be different) in the historical record of the work order text are positioned on the track and marked with a timestamp of the flaw detection data and the maintenance data.

[0061] It is to be known that first, multi-source data is aligned under a unified space-time coordinate, enabling different types of information to be fused and analyzed within the same reference framework, avoiding misjudgment caused by inconsistent sampling caliber. Second, by retaining the latest records and combining historical information, the integrity and continuity of the data are ensured, making the analysis of the track state more consistent with the real evolution process. Third, the introduction of positioning and timestamps enables subsequent knowledge graph modeling to have higher precision and traceability, which helps to quickly lock in key abnormal points. Finally, this design improves the robustness and scalability of the system, ensuring stable operation of the overall analysis even in the case of data loss or partial abnormalities, thereby better supporting the intelligent and refined management of track detection.

[0062] Based on the positioning results, the data evidence of each point on the track is generated; wherein, the defect detection data and the maintenance data are updated based on the different characteristics of the section division in each record, using the latest historical records to update the section in the record, and keeping the original data outside the work order. It is to be known that the analysis process of each cycle will be re-divided into sections. In the maintenance stage, only the sections with demand are processed. Therefore, only the work order data under this segmentation is obtained, and then the work order data of each position before this cycle is updated.

[0063] It is to be said that since the section division of the track state is regenerated in each cycle, if the historical data is directly replaced, it is easy to cause information gaps or overall deviation. By only processing the sections with demand in the maintenance stage, the system can avoid redundant modification of areas not involved, thereby maintaining the stability of the global data. At the same time, superimposing and updating the work order data of this cycle with historical records can supplement the latest maintenance results without losing past information, so that the evidence of each point not only reflects the latest state, but also inherits the historical evolution process. This updating mechanism ensures the traceability and dynamic consistency of the track data, provides more reliable input conditions for subsequent knowledge graph-based space-time prediction, and improves the accuracy of anomaly analysis and the efficiency of data management.

[0064] Real-time features of each point i: .

[0065] wherein, represents the real-time vibration data of point i, represents the real-time surface image of point i, represents the latest defect detection data of point i, represents the latest maintenance data of point i, represents the timestamp of the real-time vibration data of point i, represents the timestamp of the real-time surface image of point i, a timestamp representing the latest inspection data at the site i, a timestamp representing the latest maintenance data at the site i.

[0066] Further, the segment division comprises dividing each kind of data as an analysis dimension by analyzing the consistency and correlation of the data of the multi-source heterogeneous data between the last sampling period and the current sampling period:

[0067] Step one: in a single dimension, analyze the consistency of the data of each site within a period, if the volatility of the data is less than the corresponding threshold, the site is determined not to be processed; if the volatility of the data is not less than the corresponding threshold, the site is determined as a candidate site. Wherein, the volatility is measured by the numerical value for numerical data, and for images, the similarity is measured (the higher the similarity, the smaller the volatility).

[0068] Step two: on the track of each dimension, take the candidate site as a seed and grow along the track, if the adjacent site is also a candidate site or the correlation value of the adjacent site is higher than the evidence value of the site itself, then grow once; otherwise, stop growing; generate the segment division result under each dimension.

[0069] Wherein, the correlation value is realized by the value analysis of the LSTM neural network, which is used to represent the support degree of the data to the detection result within the interval of the current dimension under the analysis dimension (the judgment probability of each result under the data environment within the interval); by taking the probability of each detection result as a weight, the support degree of each detection result is weighted and summed to obtain the correlation value;

[0070] The evidence value of the site itself is represented by two states, and the maximum value is obtained: state 1: fixed threshold; state 2: calculate the correlation value of the site and the first seed in the opposite direction.

[0071] The credibility weight comprises taking the segment division result under each dimension as an evidence, and taking the difference between the maximum and minimum probabilities of the detection result in each evidence as the credibility weight of the evidence.

[0072] By introducing "volatility" as a consistency judgment index, different types of data (numeric, image) can be compared in a unified framework, solving the problem of multi-source heterogeneous data difficult to directly align. Further, using the growth mechanism of candidate sites, continuous segments can be dynamically formed, avoiding the fragmentation of results caused by single-point anomalies. The introduction of relevance value and self-evidence value makes the segment division not only rely on the data itself, but also consider the context environment and neighborhood support, ensuring that the results are more in line with the actual evolution law of track defects. Finally, by measuring the stability of the detection results through the credibility weight, a quantifiable confidence basis can be provided for the construction of the subsequent knowledge graph, making the evidence have a distinction and priority when propagating and reasoning. Overall, this design improves the scientificity, robustness and explainability of segment division, providing a solid foundation for subsequent prediction and optimization.

[0073] The analysis of the two-tuple includes (not only the analysis of the two-tuple for the interval, but also the analysis of each site in the interval), the identification of evidence in each dimension is performed by an agent, and the state distribution of the segment identification result is generated for each evidence according to the probability distribution of the segment identification result of each evidence (the identification of each evidence segment is performed by a pre-trained agent. The agent is trained based on a time series convolutional neural network). By introducing a pre-trained agent (based on a time series convolutional neural network), dynamic features can be captured in the time series of multi-source data, so that the segment identification result is no longer limited to the judgment of static features. The state distribution is generated in the form of probability distribution, aiming to avoid the bias caused by "single result", but to describe uncertainty with the probability of different risk types, so that the identification result has flexibility and credibility. This design can not only reflect the diversity of abnormal occurrence, but also provide quantitative input basis for subsequent knowledge graph construction and spatio-temporal propagation. By analyzing at the segment and site levels, the invention realizes double-level expression from macro to micro, maintains the integrity of the global structure, and takes into account the difference in fine granularity, ensuring that the track anomaly analysis result is more scientific, stable and traceable.

[0074] According to the rule detection result corresponding to each section identification result, the state distribution of the two-tuple is generated. Each rule detection result corresponds to different section identification results. When the system obtains the identification result of a certain section, the result is often not single, but a distribution with probability, such as "the possibility of crack is 60%, the possibility of fastener loosening is 30%, and the possibility of no abnormality is 10%". Next, the consequences that may be brought about according to these identification results, that is, the rule detection results, need to be deduced. For example, if the identification is "crack", the rule library will correspond to "increased risk of track fatigue" or "increased longitudinal vibration"; if the identification is "fastener loosening", it will correspond to "track geometric irregularity risk" or "settlement risk". When implemented, the system will first find the corresponding relationship between each identification result and its possible consequences in the rule library, and then combine the identification probability to weight and summarize the possibility of these consequences. In this way, each section will finally obtain a set of "probability distribution of rule detection results", that is, the so-called two-tuple state distribution.

[0075] According to the state distribution of the identification result corresponding to the same site in different dimensions in the historical record, the knowledge graph is obtained by learning the relationship between nodes; and all edges are multiplied by the credibility weight held by the starting evidence when information is propagated, as the transmittable amount of information.

[0076] In this embodiment, the construction of the knowledge graph is not directly inputted with multi-modal original data such as vibration signals, surface images or maintenance texts, but uses the probability distribution of the identification result corresponding to the same site in different dimensions in the historical record as the state representation of the node. Heterogeneous data is uniformly mapped to the semantic space of section state, avoiding the interference of dimension inconsistency and noise level difference between different modalities on relationship learning; through probability distribution representation, only the statistical characteristics directly related to section risk identification are retained, effectively reducing the redundant factors and noise interference in the original signal; at the same time, the probability distribution directly reflects the support degree of the evidence for a specific identification result, so that the connection between nodes can be measured by probability difference and similarity, thereby enhancing the transparency of the knowledge graph in reasoning and explanation. In addition, it is convenient to introduce credibility weight in the information propagation process, so that the propagation amount of the edge reflects not only the support degree of the probability distribution, but also the reliability of the evidence source, improving the robustness of the overall reasoning. The input based on probability distribution also makes the knowledge graph have good dynamic updating ability. When new detection data comes, only the corresponding probability distribution needs to be updated to refresh the node state, without the need to reprocess the original multi-modal data, thereby significantly reducing the computational complexity of the system and better meeting the online analysis and real-time risk assessment needs of track detection.

[0077] S2: Construct a knowledge graph with the state distribution of the evidence corresponding to the recognition result as the node, fuse the time sequence and topology on the knowledge graph for spatio-temporal prediction, and output the prediction result: the probability distribution of the risk type and the index prediction value of each risk type.

[0078] It is to be known that each risk type will be fixed with several corresponding indicators. For visualization, it is provided to maintenance personnel.

[0079] The agent is trained based on a time sequence convolutional neural network; when outputting the recognition result, a perturbation factor is generated for each dimension of the evidence according to the analysis of the time sequence, which is used to represent the influence of the characteristics of the evidence in the time sequence on other sites; that is, the agent has two outputs: 1 is the recognition result for the current section; and 2 is the influence probability distribution on other sections or sites, which is propagated through the edge.

[0080] The perturbation factor includes the probability distribution of the recognition result when identifying other sites using the evidence section. By adding a second output, i.e., the perturbation factor, in the time sequence convolutional neural network, not only the recognition result of the current section is obtained, but also the influence probability distribution of the evidence on other sections or sites is quantified. In this way, the track anomaly is no longer regarded as a single-point event in isolation, but as a dynamic factor that may spread and evolve to the surrounding area, thereby more truly depicting the diffusion law of the track defect. This design ensures that the subsequent edge propagation process in the knowledge graph has a clear quantitative basis, can reflect the coupling relationship between anomalies across sections, and improves the completeness and robustness of the prediction. At the same time, through the modeling of the perturbation factor, more detailed influence path information can be provided for counterfactual analysis and attention optimization, making the final risk assessment and optimization result more credible and interpretable.

[0081] The spatio-temporal prediction includes utilizing the spatial topology correlation to propagate the two-tuple of the perturbation factor or the evidence after the attenuation in the spatial dimension of the edge.

[0082] After each evidence completes the propagation, the state distribution of the corresponding recognition result of the two-tuple at each site is obtained, multiplied by the result of the credibility weight, and the result of the perturbation factor weighted and multiplied by the credibility weight of the corresponding source (multiplied by two weights); after inputting the knowledge graph, the prediction result is obtained.

[0083] The spatial topology correlation includes, when analyzing the correlation of interval j to any site r, if r is within j, the perturbation factor is completely lost, and the two-tuple corresponding to section j is taken as the evidence element of r in the section dimension; if r is outside j, the two-tuple corresponding to section j is completely lost, and the perturbation factor is propagated:

[0084] In the propagation, for each site on the propagation path, the matching degree of the rule detection result and the perturbation factor in each dimension of the two-tuple is calculated, the matching degree is multiplied by the standard attenuation rate to obtain the attenuation rate of each site on the propagation path in each dimension, and the attenuation rates in each dimension are superimposed to obtain the actual attenuation rate of each site on the propagation path. Wherein, the matching degree is actually the comparison of the rule detection result corresponding to the perturbation factor (obtained according to the corresponding relationship), and the similarity of the two is compared to obtain the matching degree. If the site has multiple two-tuple states, the average value is calculated respectively.

[0085] The actual attenuation rate is integrated on the path to obtain the propagation result of the perturbation factor for any site r in the interval j; and the actual attenuation rate of the perturbation factor at r is used as the weight coefficient of the influence to weight the perturbation factor after propagation attenuation.

[0086] By introducing spatial topological correlation before propagation, the different performances of anomalies inside and outside the section can be distinguished: for the inside of the section, the two-tuple is directly used as evidence to avoid repeated superposition of the perturbation factor; for the outside of the section, the influence diffusion across the section is embodied through the propagation and attenuation of the perturbation factor. The matching degree between the rule detection result and the perturbation factor is introduced, so that the propagation intensity depends not only on the spatial distance, but also on the semantic consistency, thereby ensuring that the propagation path is more targeted and reliable. Further, by accumulating the attenuation rate during the propagation process and performing integral calculation, the infinite amplification of abnormal influence can be avoided, and the propagation intensity can be gradually weakened with the distance, which conforms to the physical characteristics of track defect diffusion. Finally, the design makes the prediction result contain not only the direct evidence contribution inside the section, but also the dynamic conduction effect across the section, thereby improving the integrity, rationality and explainability of track risk prediction.

[0087] In addition, it should be noted that in track detection, different sites are often in different physical environments and operating states, and even if they belong to the same section, they may also exhibit differentiated abnormal characteristics. If only the whole section is analyzed, local characteristics may be hidden, leading to the neglect of slight abnormalities or the misjudgment of single-point abnormalities as large-scale problems. By analyzing each site separately, data fluctuations and abnormal signals can be captured at a fine-grained level, and the precise positioning of track hazards can be achieved. In this way, the sensitivity of abnormal identification is improved, and the local detection results can be further aggregated into section-level evidence, which preserves the details while ensuring overall stability. In addition, the site-level analysis provides more rich node information for subsequent knowledge graph modeling, so that the propagation and reasoning process can be based on more detailed spatio-temporal data, thereby enhancing the accuracy and explainability of the prediction result.

[0088] S3: generating attention weights of each evidence or perturbation factor through counterfactual evaluation; generating risk cost of the prediction result of each site through the rule detection result.

[0089] The counterfactual evaluation includes removing any evidence or perturbation factor in the target site to obtain a counterfactual prediction result; comparing the counterfactual evaluation with the actual prediction result, and taking the difference between the two (both are same-dimensional comparisons, and the Euclidean distance between the two can be directly compared to obtain) as the importance basis; after calculating the importance basis of all sites, the attention weight of each evidence or perturbation factor is obtained. By removing the evidence or perturbation factor of the target site, the influence of the factor on the prediction result in the absence of the factor can be simulated, thereby truly reflecting the contribution of the factor to the final result. Using the difference between the actual prediction result and the counterfactual prediction result as a measure, and calculating in the same dimension using the Euclidean distance, not only can ensure the uniformity and intuitiveness of the measurement method, but also can avoid interference caused by mixed multi-dimensional information. The importance basis obtained in this way can be converted into attention weight after normalization, which clearly identifies the influence of each evidence or perturbation factor in the prediction. This mechanism ensures that the model can still provide transparent explanation results after multi-source heterogeneous data fusion and spatiotemporal propagation, helping operation and maintenance personnel understand the key risk sources, and providing reliable quantitative support for subsequent optimization and decision-making.

[0090] The risk cost includes defining the rule detection result consistency in the two-tuple of the prediction result on the analysis site and the evidence where the site is located.

[0091]

[0092] Wherein, k represents the two-tuple index on site r; K represents the number of two-tuples on site r; represents the prediction result on site r; represents the rule detection result of the kth two-tuple on site r; represents the consistency analysis function, which is realized by a pre-trained Bayesian model. After encoding the two input contents by an encoder, the Bayesian model is used for consistency analysis to generate a consistency probability.

[0093] By comparing the prediction results with the rule detection results in the evidence tuple, the consistency between the prediction and the theoretical influence can be effectively verified, and the situation of only having prediction probability but lacking actual credibility can be avoided. By introducing a pre-trained Bayesian model as the consistency analysis function, robust probability estimation can be provided under conditions of uncertainty and noise, making the risk cost calculation statistically reasonable and fault-tolerant. At the same time, the use of an encoder for feature extraction and semantic alignment of input content can ensure that data of different modalities are compared in the same space. The risk cost defined in this way can provide a clear objective function for the subsequent optimization process, so that genetic iteration and attention redistribution converge towards reducing the risk cost, thereby realizing the practicality and explainability of track anomaly prediction.

[0094] In the track detection scenario, simply relying on the recognition result can only indicate whether an anomaly exists in a certain section, but it is difficult to determine the actual impact range and evolution trend of the anomaly. The purpose of introducing rule detection results is to provide directional reference for recognition, so that the recognition result not only stays at the level of "whether it exists", but also can be mapped to "what consequences it may cause". For example, when a track crack is recognized, the rule detection result will give a directional guide that the crack may cause an increase in longitudinal vibration or accumulation of track fatigue, thereby providing a more explicit causal chain for subsequent prediction. In this way, the rule detection result is equivalent to building a bridge between the recognition layer and the reasoning layer, so that the recognition result is no longer isolated, but can be embedded in the overall logic of the knowledge graph. This design ensures that the recognition result is extensible and explainable, and lays a directional foundation for anomaly propagation analysis and risk cost calculation.

[0095] S4: feedback the quantization result to the prediction result, perform genetic iteration optimization, and generate a final analysis result.

[0096] The genetic iteration optimization includes redistributing attention to the input of the knowledge graph.

[0097] Let the original attention of the input of the knowledge graph be: the weight corresponding to each evidence or disturbance factor (the weight of the evidence is the credibility weight, and the weight of the disturbance factor is the credibility weight multiplied by "the actual attenuation rate of the disturbance factor at r as the weight coefficient of the influence"), the result of normalization; the allocation of the original attention is taken as the initial individual. Individual code: individual = attention allocation scheme of a group of evidence or disturbance factors. Each gene site = attention value corresponding to the evidence or disturbance factor (after normalization). Initial population = composed of original attention allocation results and their disturbance samples.

[0098] All the evidence or disturbance factors on the site are triggered to generate new individuals in the genetic process according to the attention weight, and the exchange or disturbance of the local weight distribution between individuals is realized through selection / crossover / mutation to explore new solutions and generate a set of results of the attention re-allocation of the evidence or disturbance factors.

[0099] Specifically, the generation of new individuals in the genetic process according to the attention weight is actually the generation process of the first generation of individuals, and N first generation individuals are generated. Each individual is different, and the number of all evidence or disturbance factors is n, so the number of individuals in the first generation that are weighted for a single evidence or a single disturbance factor is n. (It is ensured that there is an individual that is weighted for only one evidence or disturbance factor)

[0100] The other N-n individuals are randomly triggered according to the attention weight as a probability when generated.

[0101] Further, selection (Selection): probabilistic selection based on fitness, common methods include roulette selection or tournament selection. Individuals with low risk cost are retained to enter the next generation, and the genetic probability of high-quality solutions is improved.

[0102] Crossover (Crossover): Randomly select two individuals (attention allocation schemes) and exchange the attention values of part of the gene sites according to the set crossover rate. For example: the part of the offspring evidence weight is taken from parent A, and the other part is taken from parent B. The crossover result is normalized to keep the attention sum to 1.

[0103] Mutation (Mutation): Randomly disturb some gene sites of an individual to change its attention value. The disturbance amplitude is in the range of [0, L], which simulates the proportion of attention increase or decrease. After mutation, normalization is also performed to ensure the legality of the probability distribution.

[0104] The selection-crossover-mutation process is repeated to generate a new generation population.

[0105] Among them, the increase proportion of the attention re-allocation result is any value in 0 to L. This allows the optimization process to explore different amplitude adjustment schemes, rather than being limited to fixed proportion changes. By limiting the increase proportion to the range of 0 to L, it can be ensured that the change amplitude of the attention will neither be too large to cause model oscillation instability, nor be too small to make the optimization stagnate. In this way, the genetic process can generate diversified new individuals within a controllable range when performing crossover and mutation, which not only guarantees the global search ability, but also maintains the interpretability and convergence of the results. At the same time, the proportion limited in the interval [0, L] also facilitates the combination with the normalization mechanism to ensure that the final attention distribution meets the probability constraint, thereby forming an optimal allocation result that meets the actual needs.

[0106] The optimization is performed with the minimum risk cost as the objective function, and the final attention allocation result is obtained by iterating to the maximum number of iterations or when the risk cost no longer decreases. The final analysis result is generated using the knowledge graph based on the attention re-allocation result, the result of multiplying the state distribution of the evidence corresponding to the recognition result by the credibility weight, and the result of multiplying the weighted disturbance factor by the credibility weight of the corresponding source (multiplied by two weights).

[0107] Through the genetic optimization mechanism, the attention weight allocation can achieve global optimization between multi-source evidence and disturbance factors, rather than being limited to single-point greedy adjustment. Through selection operation, high-quality individuals (low-risk cost allocation schemes) can be preferentially retained. Through cross operation, information combination between different schemes is realized to explore new potential optimal solutions. Through mutation operation, random disturbance is introduced to avoid the algorithm falling into local optimum. Overall, this evolutionary optimization mechanism can continuously iterate in a complex search space, so that the final attention allocation tends to be a scheme with the minimum risk cost. This design ensures that the model is not only effective in a single prediction, but also continuously improves robustness and stability in iterative optimization, thereby outputting analysis results that are more consistent with the actual operation rules of the track.

[0108] In an optional embodiment, to ensure the spatial continuity of the track anomaly prediction results during the iterative optimization process, there is a constraint between adjacent two sites during the iteration process: the probability distribution similarity of anomalies between the two nodes cannot be lower than a preset value. If the anomaly probability distribution of a certain site is too different from that of the adjacent site during the iteration process, it is determined that the result does not conform to the physical characteristics of the track and needs to be corrected by a penalty factor. Specifically, after generating a new individual in each iteration, the system calculates the similarity of the prediction probability distribution of adjacent sites, for example, using cosine similarity or Euclidean distance inverse index. If the result is lower than the preset threshold, a penalty term is added to the fitness function to reduce the fitness of the individual. In this way, the iterative optimization can automatically converge to a solution that not only reduces the risk index, but also ensures the spatial distribution of the track prediction results to be continuous and reasonable. It can avoid the abrupt situation of "predicting high risk at a certain site and low risk at the adjacent site", which is more consistent with the actual rules of the gradual diffusion of track defects along the space. At the same time, the introduction of the adjacent constraint can also improve the stability and interpretability of the system prediction results, making the generated maintenance arrangement more practical.

[0109] The final analysis result includes the probability distribution of each risk type and the corresponding index prediction value, and the maintenance arrangement is generated accordingly. In this embodiment, specifically: the system first performs threshold determination on the probability distribution of different risk types, and marks the risk types with probability exceeding the set threshold as the focus of attention; then, in combination with the index prediction value corresponding to the risk type, the influence degree is further evaluated. When the index prediction value of a certain risk type exceeds the upper limit of safety, the system divides the section into a first-level risk, and immediately generates an emergency maintenance task; when the probability is high and the index prediction value is in the critical interval, it is divided into a second-level risk, and is included in the near-term maintenance plan; when the probability is low but the index prediction value is close to the threshold, it is divided into a third-level risk, and is included in the periodic inspection range; when both the probability and the index are at a low level, it is marked as a fourth-level risk, and only routine monitoring is performed. Through the above-mentioned manner, the system can dynamically generate a maintenance task list, and automatically match the maintenance timeliness and resource allocation according to the different risk levels, so as to realize the fine and scientific management of track maintenance.

[0110] The embodiment also provides a track detection data analysis system based on a knowledge graph, comprising:

[0111] A collection unit acquires multi-source heterogeneous data from a track detection system, performs alignment on the multi-source heterogeneous data, performs section division after alignment, and assigns a credibility weight to each data evidence; meanwhile, the divided sections are analyzed in two-tuples.

[0112] A prediction unit constructs a knowledge graph taking state distribution of evidence corresponding to recognition result as a node, performs spatio-temporal prediction on the knowledge graph by fusing time sequence and topology, corrects the prediction result by physical constraint, and outputs the prediction result: probability distribution of risk type and index prediction value of each risk type.

[0113] A calculation unit generates an attention weight of each evidence or disturbance factor through counterfactual evaluation; and generates a risk cost of the prediction result of each site through the rule detection result.

[0114] An output unit feeds back the quantization result to the prediction result, performs genetic iteration optimization, and generates a final analysis result.

[0115] The embodiment also provides a computer device suitable for the track detection data analysis method based on a knowledge graph, comprising: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the track detection data analysis method based on a knowledge graph proposed in the above embodiment.

[0116] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved by WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0117] The embodiment also provides a storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the track detection data analysis method based on a knowledge graph. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0118] To sum up, the present application realizes the fine identification of track anomaly, the modeling of propagation path and the risk grading by the following steps: acquiring the multi-source heterogeneous data of the track detection system and performing alignment and section division, combining the credibility weight to construct the two-tuple evidence, further realizing the fusion prediction of time sequence and spatial topology with the knowledge graph as the carrier, outputting the probability distribution of the risk type and the corresponding index prediction value, generating the attention weight by using the counterfactual evaluation, and forming the risk level and directional guidance based on the rule detection result, introducing the genetic iteration mechanism to optimize the attention distribution, and combining the similarity constraint of adjacent sites to ensure the rationality and continuity of the result. Through the above steps, the present application can realize the fine identification of track anomaly, the modeling of propagation path and the risk grading, and provide quantifiable and interpretable decision basis for the maintenance arrangement, so as to improve the scientificity and safety of track maintenance.

[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A knowledge graph-based method for analyzing orbit detection data, characterized in that: This includes acquiring multi-source heterogeneous data from the track detection system, aligning the multi-source heterogeneous data and dividing it into segments, and assigning credibility weights to each data piece of evidence; at the same time, performing binary analysis on the divided segments. A knowledge graph is generated by constructing a knowledge graph with the state distribution of the evidence-corresponding identification results as nodes. Temporal and topological data are integrated on the knowledge graph to perform spatiotemporal prediction, and the prediction results are output: the probability distribution of risk types and the index prediction value of each risk type. Attention weights for each piece of evidence or perturbation factor are generated through counterfactual evaluation; The risk cost of generating the prediction result for each site based on the detection results of the aforementioned rules; The quantification results are fed back to the prediction results for genetic iteration optimization, generating the final analysis results. The two-tuples are based on probability distributions; the two-tuples include segment identification results and rule detection results; representing the probability distribution of the identification results of abnormal entities and the probability distribution of the theoretical influence results of abnormal entities, respectively.

2. The knowledge graph-based orbit detection data analysis method as described in claim 1, characterized in that: The multi-source heterogeneous data includes vibration data collected by vibration sensors arranged on the track, surface images collected by cameras, and work order text generated during maintenance. The vibration data and surface image are located on the track using the positioning information of the vibration sensor and the positioning information during image acquisition; the flaw detection data and maintenance data under the section division results are located on the track using the historical records of the work order text, and the timestamps of the flaw detection data and maintenance data are marked. Based on the positioning results, the data evidence for each point on the track is generated; wherein, the flaw detection data and the maintenance data are updated based on the different characteristics of segment division in each record, using the latest historical record, so that the original data is retained in the area outside the work order; Real-time features at each site i: ; in, This represents the real-time vibration data at point i. This represents the real-time surface image at site i. This represents the latest flaw detection data at point i. This represents the latest maintenance data at site i. The timestamp represents the real-time vibration data of site i. This represents the timestamp of the real-time surface image at site i. This represents the timestamp of the latest flaw detection data at point i. The timestamp represents the latest maintenance data at point i.

3. The knowledge graph-based orbit detection data analysis method as described in claim 2, characterized in that: The segmentation includes dividing the data into analytical dimensions by analyzing the consistency and correlation of multi-source heterogeneous data between the previous sampling period and the current sampling period, using each type of data as an analytical dimension. Step 1: Analyze the consistency of data for each site within a single dimension. If the data volatility is less than the corresponding threshold, the site is deemed to be left unprocessed; if the data volatility is not less than the corresponding threshold, it is deemed to be a candidate site. Step 2: On each dimension's track, using the candidate sites as seeds, grow along the track. If adjacent sites are also candidate sites or the correlation value of adjacent sites is higher than the site's own evidentiary value, then one growth is achieved; otherwise, growth stops; generate the segmentation results for each dimension. The correlation value is a value analysis implemented through an LSTM neural network, used to represent the degree of support of the data for the detection results within the current dimension under the analysis dimension; the correlation value is obtained by weighting and summing the support degree of each detection result with the probability of each detection result as the weight. The intrinsic evidence value is represented by two states, with the maximum value being obtained: State 1: fixed threshold; State 2: calculate the correlation value between the site and the first seed in the opposite direction. The credibility weights include taking the segmentation result under each dimension as a piece of evidence, and taking the difference between the maximum and minimum probabilities of the detection results in each piece of evidence as the credibility weight of the evidence.

4. The knowledge graph-based orbit detection data analysis method as described in claim 3, characterized in that: The analysis of the binary tuple includes using an agent to identify evidence in various dimensions, and generating a state distribution of the segment identification results for each piece of evidence based on the probability distribution of the segment identification results. Furthermore, based on the rule detection results corresponding to the identification results of each segment, a state distribution of two tuples is generated; Each rule's detection result corresponds to a different segment identification result; Based on the state distribution of the identification results corresponding to the same site in different dimensions in the historical record, the knowledge graph is obtained by learning the relationship between the nodes; and all edges are multiplied by the credibility weight of the starting evidence when information is propagated, which is used as the amount of information that can be propagated.

5. The knowledge graph-based orbit detection data analysis method as described in claim 4, characterized in that: The intelligent agent is trained using a temporal convolutional neural network as a model. When outputting the recognition result, based on the temporal analysis, a perturbation factor is generated for the evidence in each dimension to represent the influence of the evidence's features on other sites in the temporal sequence. The perturbation factor includes the probability distribution of the identification results when other sites are identified using evidence segments; The spatiotemporal prediction includes using spatial topological correlation to propagate the perturbation factor or evidence tuple after attenuating it in the spatial dimension of the edge. After each piece of evidence has been propagated, the result of multiplying the state distribution of the identification result corresponding to the two tuples at each site with the credibility weight, and the result of multiplying the perturbation factor by the credibility weight of the corresponding source, is obtained; after inputting into the knowledge graph, the prediction result is obtained. The aforementioned spatial topological correlation includes the following: when analyzing the correlation between interval j and any point r, if r is within j, then all the perturbation factors are lost, and the binary tuple corresponding to segment j is used as evidence elements of r in the dimension of the segment; if r is outside j, then all the binary tuple corresponding to segment j is lost, and the perturbation factors are used for propagation. During propagation, the degree of matching between the rule detection result and the perturbation factor in the tuple of each dimension at each point in the propagation path is calculated, and the degree of matching is multiplied by the standard decay rate to obtain the decay rate at each point in each dimension of the propagation path. The actual attenuation rate at each point along the propagation path is obtained by superimposing the attenuation rates of each dimension. Integrating the actual attenuation rate along the path yields the propagation result of the perturbation factor for any point r in interval j. The actual attenuation rate of the disturbance factor at point r is used as the weighting coefficient of the influence to weight the disturbance factor after propagation attenuation.

6. The knowledge graph-based orbit detection data analysis method as described in claim 5, characterized in that: The counterfactual assessment includes removing any evidence or perturbation factor from the target site to obtain a counterfactual prediction result; The counterfactual assessment is compared with the actual prediction results, and the difference between the two is used as the basis for importance. After calculating the importance of all evidence at a site, the attention weight of each piece of evidence or perturbation factor is obtained; The risk cost is defined by the consistency between the prediction result at the analysis site and the rule detection result in the binary tuple of evidence at the site: Where k represents the index of the binary tuple at site r; K represents the number of binary tuples at site r; This represents the prediction result at site r; This represents the rule detection result for the k-th binary tuple at site r; The analysis function representing the consistency is implemented through a pre-trained Bayesian model. After the encoder encodes the two input contents respectively, the Bayesian model is used to perform consistency analysis and generate the consistency probability.

7. The knowledge graph-based orbit detection data analysis method as described in claim 6, characterized in that: The optimization of the genetic iteration includes reallocating attention to the input of the knowledge graph; Let the original attention score of the input to the knowledge graph be the result of normalization of the weight corresponding to each piece of evidence or perturbation factor; and let the allocation of the original attention score be used as the initial individual. For all evidence or perturbation factors at the locus, the generation of new individuals in the genetic process is probabilistically triggered according to the attention weights. In the new individuals, through selection / crossover / mutation, the local weight allocation of individuals is exchanged or perturbed, new schemes are explored, and a set of evidence or perturbation factor attention redistribution results are generated. In the result of the redistribution of attention, the increase in attention ratio is any value from 0 to L. The optimization is performed with the minimum risk cost as the objective function; the final attention allocation result is obtained by iterating until the maximum number of iterations or until the risk cost no longer decreases, and the final analysis result is generated using the knowledge graph based on the attention redistribution result.

8. A knowledge graph-based orbit detection data analysis system, based on the knowledge graph-based orbit detection data analysis method according to any one of claims 1 to 7, characterized in that: This includes a data acquisition unit that acquires multi-source heterogeneous data from the track detection system, aligns the multi-source heterogeneous data, divides it into segments, and assigns credibility weights to each data piece of evidence; simultaneously, it performs binary analysis on the divided segments. The prediction unit constructs a knowledge graph with the state distribution of the evidence-corresponding identification results as nodes. It integrates temporal and topological data on the knowledge graph to perform spatiotemporal prediction, corrects for physical constraints, and outputs the prediction results: the probability distribution of risk types and the predicted index value of each risk type. The computational unit generates attention weights for each piece of evidence or perturbation factor through counterfactual evaluation; The risk cost of generating the prediction result for each site based on the detection results of the aforementioned rules; The output unit feeds back the quantification results to the prediction results, performs genetic iteration optimization, and generates the final analysis results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the knowledge graph-based orbit detection data analysis method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the knowledge graph-based orbit detection data analysis method according to any one of claims 1 to 7.

Images