Track circuit fault cause analysis method based on text record data mining

Abstract

The invention discloses a track circuit fault cause analysis method based on text record data mining, and belongs to the technical field of track circuit fault diagnosis and analysis in a railway signal system, and the method comprises the following steps: S1, collecting a track circuit fault text, and processing the track circuit fault text into a standardized fault text R; s2, semantic features and word order features of the R are extracted and fused, and a feature set S is output after optimization of an SMOTE algorithm; s3, inputting the S into an FEML model, and outputting a large-class label corresponding to the fault text through parameter optimization, parallel training of a base learner and integration of a meta learner; s4, in combination with the weight value and a Dirichlet multi-term hybrid model, extracting and outputting a fine-grained fault type and a cause thereof under the large-class label; and S5, constructing a visual map by using Neo4j, and realizing association analysis through a Cyber query language. According to the method, the analysis model and the visual knowledge graph are constructed, fault causes are accurately mined, and the operation and maintenance efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of track circuit fault diagnosis and analysis technology in railway signaling systems, and in particular to a method for analyzing the causes of track circuit faults based on text record data mining. Background Technology

[0002] Track circuits, as a crucial component of railway train operation control systems, are widely used in my country's rail transit lines, undertaking key functions such as section occupancy detection, rail breakage monitoring, and information transmission. The reliability and stability of track circuits are vital guarantees for railway traffic safety and transportation efficiency; their performance directly determines the operational stability and accident risk of the railway system.

[0003] Currently, the academic research focus on track circuit faults mainly involves fault diagnosis and cause analysis of equipment failures through quantitative analysis of monitoring data. There is still a significant gap in the research on in-depth mining and knowledge discovery of fault text records in historical operation and maintenance data. In terms of cause analysis of track circuit faults, there is a lack of diagnostic and analysis methods that can accurately extract fault features and efficiently analyze faults, especially in the face of complex conditions such as large quantities of unstructured fault texts. Summary of the Invention

[0004] The purpose of this invention is to provide a method for analyzing the causes of track circuit faults based on text record data mining, thereby solving the aforementioned technical problems.

[0005] To achieve the above objectives, this invention provides a method for analyzing the causes of track circuit faults based on text record data mining, comprising the following steps: S1. Text Collection and Preprocessing: Collect track circuit fault texts and process them into standardized fault texts. ; S2, Dual-view text feature extraction and SMOTE algorithm: Extracting standardized fault text from the output of S1. The semantic and word order features are analyzed, and then fused together. After optimization using the SMOTE algorithm, a fault text feature set is output. ; S3. Fault Text Classification Based on FEML Model: Construct a three-layer FEML model and classify the fault text feature set output from S2. Input the FEML model, and after parameter optimization, parallel training of base learners and integration with random forest meta-learners, output the major category labels corresponding to the fault texts; S4. Fault text clustering based on Dirichlet multinomial mixture model: For the major category labels output by S3, combine TF-IDF weights and Dirichlet multinomial mixture model to extract and output the fine-grained fault types and causes under the major category labels. S5. Construct a visual knowledge graph of fault causes: Use Neo4j to store the fault types and causes output by S4 and construct a visual knowledge graph. Then, use Cypher query language to perform multi-dimensional correlation analysis and output the analysis results of fault types and causes.

[0006] Preferably, S1 specifically includes the following steps: The collected track circuit fault texts were cleaned and standardized by referring to a railway professional thesaurus, transforming non-standard texts into standard fault texts. .

[0007] Preferably, S2 specifically includes the following steps: S21. Standard fault text output by S1 An improved mutual information algorithm is used to extract semantic dimension features, and a distributed memory paragraph vector model is used to extract word order dimension features, thus constructing a dual-view feature extraction framework. Among them, through word frequency To calculate semantic features in the semantic dimension of the text The frequency of occurrence is calculated using the following formula: (1); in, Semantic features; semantic features Number of times it appears; Total number of semantic features; word frequency semantic features Distribution frequency and semantic features in faulty text datasets With word frequency Proportional; In addition, through word frequency and average word frequency Calculate semantic features In the major categories of faults The frequency of occurrence in [the text] is calculated using the following formula: (2); (3); in, , semantic features In the major categories of faults The number of times it appears in; Total number of fault categories; semantic features In the major categories of faults The frequency of occurrence in [the context of troubleshooting] and in major categories of troubleshooting The frequency of occurrence of faults outside the major categories is inversely proportional to the frequency of category words. and average word frequency Numerical values ​​and characteristics Representative categories of faults Proportional; S22, Based on word frequency , word frequency Average word frequency The indicators are used to filter out fault texts whose indicator values ​​are higher than other indicator values, thus identifying high-frequency fault texts. Then, the selected high-frequency fault texts The data structure is optimized using the SMOTE algorithm, and a fault text feature set is output. .

[0008] Preferably, S3 specifically includes the following steps: S31. Based on the FEML model, the fault text feature set output by S22... The data were mapped to numerical labels and then divided into training and test sets in a 7:3 ratio. S32. Based on the training set, use the first layer of the FEML model, employ K-fold cross-validation to train the model performance, and combine grid search cross-validation to optimize the base learner parameters. S33. Based on the optimized parameters of the first layer of the FEML model, the second layer of the FEML model is used to train three base learners in parallel: gradient boosting decision tree, linear support vector machine and ridge regression classifier, to capture the features of fault text from different perspectives and output intermediate results. S34. Using the third layer of the FEML model, the intermediate results output from the second layer are input into the random forest meta-learner for integration, and the final fault text is output. The corresponding category tags and their corresponding fault types are: Track line cause A, Power supply cause B, Equipment cause C, Environmental cause D, Construction and human-related cause E, and System and software cause F.

[0009] Preferably, S4 specifically includes the following steps: S41, Final fault text based on the output of S34 For the corresponding category labels, a dynamic feature weight adjustment mechanism is introduced, where the feature weights... The specific calculation formula is as follows: (4); The final fault text after S3 processing The characteristics are , It is a feature The weight, It is a feature In the final fault text The corresponding category tags Weight value; S42. Adjust the dynamically weighted features. Input the Dirichlet multinomial mixture model to perform fine-grained clustering mining on the fault category data, and extract and output the fine-grained fault type and cause information under the corresponding category label.

[0010] Preferably, S5 specifically includes the following steps: S51. Based on the Neo4j graph database, the entities and their relationships of the fine-grained fault types and causal information output by S4 are integrated and stored as a graph structure knowledge base, and a visualization interface is built to present the entities and their relationships in the graph structure knowledge base in an intuitive form. S52. Retrieve the graph structure knowledge base generated in S51 using the Cypher query language, perform multi-dimensional correlation analysis based on fault type and cause information, and output the analysis results.

[0011] Preferably, in step S21, an improved mutual information algorithm is introduced for the semantic dimension. This algorithm uses average class frequency... Together with the inverted category frequency, they constitute a composite weighting factor, which dynamically adjusts the feature weights; the word order dimension establishes the connection between the grammatical structure of the faulty text sequence and the context through a distributed memory paragraph vector model, and constructs a temporal feature vector with high discriminative power.

[0012] Preferably, the distributed memory paragraph vector model is a neural network model for learning the vector representation of text paragraphs. This model maps faulty text paragraphs to fixed-length feature vectors, captures faulty text information, and extracts word order features. To mitigate the impact of random fluctuations during feature vector training, the distributed memory paragraph vector model defines stable feature vectors for faulty text by independently repeating the training process multiple times and calculating the arithmetic mean of the feature vectors. Stable eigenvectors The specific calculation formula is as follows: (5); in, For the first i The fault text is in the first Feature vectors generated in the next iteration of training. This represents the total number of iterations.

[0013] Preferably, the fine-grained fault types and cause information output by S42 are written into a CSV file using the Neo4j LOAD CSV statement, and then the data in the CSV file is imported into the Neo4j graph database to provide a structured data foundation for the construction of a knowledge graph of track circuit fault causes.

[0014] Therefore, the present invention employs the above-mentioned method for analyzing the causes of track circuit faults based on text record data mining, which has the following beneficial effects: 1. By integrating dual-view feature extraction and ensemble learning, a semantic-word order dual-modal feature space is constructed, overcoming the problems of data dimensional redundancy and distribution skewness. By improving the mutual information (IMI) algorithm to optimize the feature selection mechanism, and combining the enhanced distributed memory paragraph vector (PV-DM) model to achieve collaborative representation of syntactic structure and contextual semantics, the discriminativeness of the feature space is effectively improved.

[0015] 2. Innovatively construct the FEML (Feature Enhanced Meta-Learning) model to classify fault texts, and then use the Dirichlet multinomial mixture model to cluster the classified fault texts, ultimately achieving multi-granularity analysis of fault texts and overcoming the problem of inter-class confusion in complex fault modes.

[0016] 3. Establish a fault knowledge graph with causal reasoning capabilities to achieve visualized mapping of fault associations, further enhancing the ability to analyze fault causes. In addition, use Neo4j's visualization tools to display the knowledge graph and analyze fault associations through Cypher query language to provide decision support for track circuit maintenance and fault analysis.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method steps of the present invention; Figure 2 This is a diagram summarizing the steps of the method of the present invention; Figure 3 This is a flowchart of the fault text representation and classification model of the present invention; Figure 4 This is a framework diagram for dual-view fault feature extraction in this invention; Figure 5 This is a structural diagram of the FEML model of the present invention; Figure 6 This is a flowchart of the fine-grained clustering and knowledge graph construction of track circuit fault text in this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0020] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0021] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] This invention provides a method for analyzing the causes of track circuit faults based on text record data mining, such as... Figure 1 As shown, it includes 5 steps: S1. Text Collection and Preprocessing: Collect track circuit fault texts and process them into standardized fault texts R to solve the problems of chaotic fault text format and inconsistent terminology, providing high-quality, structured input data for subsequent feature extraction and model training. S2, Dual-view text feature extraction and SMOTE algorithm: Extract the semantic features and word order features of the standardized fault text R output by S1, and fuse the semantic features and word order features. After optimization by the SMOTE algorithm, output the fault text feature set S, which solves the problems of data dimensional redundancy, distribution skewness and insufficient feature extraction, and provides highly discriminative feature input for the classification model. S3. Fault text classification based on FEML (Feature Enhanced Meta-Learning) model: The fault text feature set S output by S2 is input into the FEML (Feature Enhanced Meta-Learning) model. After parameter optimization, parallel training of base learners and integration by random forest meta-learners, the major category labels corresponding to the fault texts are output. This process improves the accuracy and stability of fault text classification through ensemble learning, realizes accurate classification of fault texts, and lays the foundation for subsequent fine-grained analysis. S4. Fault Text Clustering Based on Dirichlet Multinomial Hybrid Model: For the major category labels output by S3, combine TF-IDF weights and the Dirichlet multinomial hybrid model to extract fine-grained fault types and causes under the major fault categories. This breaks through the problem of inter-class confusion under complex fault modes, realizes in-depth decomposition of fault categories, accurately identifies specific fault types and their core causes, and improves the fine-grainedness of fault analysis. S5. Construct a visual knowledge graph of fault causes: Use Neo4j to store the fault types and their causes output by S4 and construct a visual knowledge graph. Then, use Cypher query language to perform multi-dimensional correlation analysis and output the analysis results of fault types and causes. This process provides visual decision support for operation and maintenance personnel by intuitively presenting the correlation path between fault types and causes, core influencing factors and evolution patterns, thereby improving the efficiency of fault handling.

[0023] like Figure 2 As shown, S1 specifically includes the following steps: The collected track circuit fault text is cleaned and standardized using a railway professional thesaurus, including special character removal, vocabulary normalization, and word segmentation, to convert the unstructured text into standard fault text. This lays the foundation for subsequent feature extraction.

[0024] S2 specifically includes the following steps: S21, processing the standard fault text output by S1. An improved mutual information algorithm is used to extract semantic dimension features, and a distributed memory paragraph vector model is used to extract word order dimension features, thus constructing a dual-view feature extraction framework. Among them, through word frequency To calculate semantic features in the semantic dimension of the text The frequency of occurrence is calculated using the following formula: (1); in, Semantic features; semantic features Number of times it appears; Total number of semantic features; word frequency semantic features Distribution frequency and semantic features in faulty text datasets With word frequency Proportional, word frequency Larger means greater semantic features The stronger the ability to distinguish; In addition, through word frequency and average word frequency Calculation features In the major categories of faults The frequency of occurrence in [the text] is calculated using the following formula: (2); (3); in, , semantic features In the major categories of faults The number of times it appears in; Total number of fault categories; semantic features In the major categories of faults The frequency of occurrence in [the context of troubleshooting] and in major categories of troubleshooting The frequency of occurrence of faults outside the major categories is inversely proportional to the frequency of category words. and average word frequency Numerical and semantic features Representative categories of faults Proportional, semantic features In the major categories of faults The more times a term appears in a fault category, and the fewer times it appears in other fault categories, the higher the frequency of the term. and average word frequency The larger the semantic features It better represents this type of fault. S22, based on word frequency , word frequency Average word frequency The indicators are used to filter out fault texts whose indicator values ​​are higher than other indicator values, thus identifying high-frequency fault texts. This type of high-frequency fault text High-frequency and highly representative texts were then selected as high-frequency fault texts. The data structure is optimized using the SMOTE algorithm, and a fault text feature set is output. .

[0025] S3 specifically includes the following steps: S31, FEML (Feature Enhanced) The Meta-Learning model first labels the fault text feature set output by S22, defining the data labels of fault categories A to F as 0 to 5 respectively. That is, track line cause A corresponds to 0, power supply cause B corresponds to 1, equipment cause C corresponds to 2, environmental cause D corresponds to 3, construction and human cause E corresponds to 4, and system and software cause F corresponds to 5. The above 6 categories of faults are common track circuit faults, which are further subdivided as follows: The specific causes of track line cause A are: rail wear, deformation, and abnormal rail gaps; the specific causes of power supply cause B are: unstable power supply voltage and harmonic interference of power supply lines; the specific causes of equipment cause C are: aging and damage of track relays or insulation components, loose or broken connecting parts; the specific causes of environmental cause D are: rain, low temperature, and icing; the specific causes of construction and human cause E are: construction damage to rail insulation or connecting wires; the specific causes of system and software cause F are: misjudgment of control system faults and abnormal signal transmission.The fault categories described in the text are then converted into numerical labels that can be recognized by the machine learning model. The training and test sets are divided in a 7:3 ratio. The training set is used for parameter learning of the FEML (Feature Enhanced Meta-Learning) model, such as adjusting the weights of the base learner and fitting the rules of the meta learner, enabling the model to learn the association between features and labels in the samples to form classification logic. The test set, as data not used in training, is used to evaluate the model's classification accuracy for new fault texts, avoiding overfitting due to the accidental appearance of specific terms. These training and test sets provide a scientific dataset foundation for subsequent parameter optimization and model training. S32. Based on the divided dataset, the first layer of the FEML (Feature Enhanced Meta-Learning) model uses K-fold cross-validation to train model performance and combines it with grid search cross-validation (Grid Search CV) to optimize the base learner parameters, laying the parameter foundation for model training. S33. Using the optimized parameters of the first layer of the FEML (Feature Enhanced Meta-Learning) model, the FEML (Feature Enhanced Meta-Learning) model... The second layer of the Meta-Learning model uses an improved Stacking algorithm to train three base learners in parallel: Hist GBM (Gradient Boosting Decision Tree), Linear Support Vector Machine (Linear SVC), and Ridge Regression Classifier. This captures the features of the fault text from different perspectives and outputs intermediate results. The third layer of the model then inputs the intermediate results from the second layer into a Random Forest Meta-Learner for integration, and outputs the final fault text. The corresponding major category tags and their corresponding fault types are: track line cause A, power supply cause B, equipment cause C, environmental cause D, construction and human cause E, and system and software cause F, for a total of 6 categories.

[0026] S4 specifically includes the following steps: S41, S42, the final fault text based on the output of S34. For the corresponding class label AF, a dynamic feature weight adjustment mechanism is introduced, where the feature weights... The specific calculation formula is as follows: (4); Among them, the fault text after S3 processing The characteristics are , It is a feature The weight, It is a feature In the final fault text The corresponding category tag AF Weight value, The purpose of the weight value is to calculate word frequency. and inverse document frequency To measure the importance of a word in a document. S42, dynamically adjust the feature weights. The input is a Dirichlet multinomial mixture model, which considers the relationships between texts and can perform cluster analysis more accurately, meticulously uncovering fault types and causes that are more consistent with reality. This model extracts and outputs fine-grained fault type and cause information under major category labels, realizing fine-grained clustering of fault types and accurate extraction of high-frequency causes. A fault knowledge graph with a priority evaluation model is constructed to solve the problems of fine-grained classification difficulties and fuzzy identification of high-frequency causes caused by the complexity of track circuit fault text data.

[0027] S5 specifically includes the following steps: S51, based on the Neo4j graph database, integrate and store the entities and interrelationships of the fine-grained fault types and causal information output by S4 into a graph structure knowledge base, and build a visualization interface to present the entities and interrelationships in the graph structure knowledge base in an intuitive form; S52, use the Cypher query language to retrieve the graph structure knowledge base generated in S51, perform multi-dimensional correlation analysis based on fault types and causal information, and output the analysis results to provide decision support for track circuit maintenance and fault analysis.

[0028] like Figure 3 , 4 As shown, the semantic dimension is achieved through an improved mutual information (IMI) algorithm, which uses average class word frequency. Together with the inverted class frequency, it constitutes a composite weighting factor that dynamically adjusts feature weights. The inverted class frequency, by inversely measuring the frequency of a feature's occurrence in non-target categories, strengthens the feature's representativeness for the target fault category. This improves the Mutual Information (IMI) algorithm by amplifying the weight of features that appear frequently in the target category but infrequently in other categories, such as "rainwater" appearing frequently in environmental cause D, thus enhancing its representativeness for the target category. For features that appear frequently across categories, such as broad terms like "fault," the weighting factor reduces their weight to prevent them from interfering with the distinction between fault categories. This addresses the problem of feature misjudgment caused by uneven distribution of fault text data, accurately capturing feature F. IMIThe core representativeness within a specific fault category; the word order dimension is achieved through the Distributed Memory Phrase Vector (PV-DM) model, a neural network model for learning text paragraph vector representations. This model maps faulty text paragraphs to fixed-length feature vectors, capturing faulty text information and extracting word order features. To reduce the impact of random fluctuations during feature vector training, the PV-DM model is independently and repeatedly trained and the arithmetic mean of the feature vectors is calculated to define stable feature vectors for faulty text. Stable eigenvectors The specific calculation formula is as follows: (5); in, For the first i The fault text is in the first Feature vectors generated in the next iteration of training. The distributed memory segment vector (PV-DM) model, representing the total number of iterations, serves to provide each feature F... PV-DM By assigning a unique vector and training the vector representation of features in different contexts, a temporal feature vector containing grammatical structure and contextual associations is finally generated, overcoming the limitation of the traditional bag-of-words model that ignores word order and supplementing the structural information not covered by the semantic dimension. Figure 4 The "feature fusion" step in the process combines the category association features extracted from the semantic dimension with the structural association features extracted from the word order dimension into a unified feature set, which is... Figure 3 The algorithm extracts features at the term and character levels, and then optimizes them using the SMOTE algorithm to address data skewness. The output is a high-quality feature set that can be finely classified in the FEML classifier. The above uses semantic and word order features for collaborative representation, and constructs a dual-view feature extraction framework.

[0029] like Figure 5As shown, the FEML (Feature Enhanced Meta-Learning) model achieves high-precision classification of track circuit fault text through a three-layer structure design of "parameter optimization layer - parallel base learner layer - meta learner integration layer". The first layer is the parameter optimization mechanism, which adopts K-fold cross-validation. That is, the dataset is divided into K folds, and K training and validation are performed. Finally, the average value of the K results is taken, where K is an adjustable parameter. In this embodiment, K=5, that is, five-fold cross-validation combined with grid search cross-validation (Grid Search CV) method is used. The preset parameter space such as the learning rate, tree depth, and regularization coefficient of the gradient boosting decision tree (HistGBM) and linear support vector machine (linear SVC) is traversed to select the parameter combination that optimizes the performance of the base learner. This solves the problem of blind setting of base learner parameters and avoids underfitting or overfitting caused by unreasonable parameters. This lays a stable parameter foundation for the parallel training of the second layer of base learners. The second layer consists of three base learners: Hist GBM (Gradient Boosting Decision Tree) handles nonlinear relationships and high-dimensional sparse features. It iteratively constructs multiple decision trees, each focusing on correcting the errors of the preceding trees, accurately capturing complex interactions between features to solve nonlinear association problems that are difficult to cover with a single rule. Linear SVC (Linear Support Vector Machine) is used for linear classification of high-dimensional text features. It achieves efficient linear classification by finding the optimal classification hyperplane. For high-frequency text features, it quickly defines linear boundaries, clearly distinguishing them from other fault categories and improving classification speed. Ridge Regression Classifier limits the model parameter size through regularization, effectively reducing the risk of overfitting caused by high-dimensional text features such as niche terms and low-frequency fault descriptions. For text features with limited sample sizes, Ridge Regression Classifier suppresses noisy features through shrinkage coefficients, ensuring classification stability and improving the recognition accuracy of niche categories. These three base learners complement each other in three dimensions: "capturing nonlinear relationships," "high-dimensional linear partitioning," and "anti-overfitting stable output," avoiding the inherent defects of a single learner. The third layer is the random forest meta-learner integration layer. The third layer takes the intermediate classification results of the three base learners in the second layer as new feature inputs, learns the prediction patterns of each base learner, and solves the inherent limitations of a single base learner. By integrating complementary information, the accuracy and stability of classification are improved.

[0030] like Figure 6As shown, the fine-grained fault types, causal factors, and other entity relationships and interrelationships output by S42 are first written into a CSV (Comma-Separated Values) file. This file contains three core fields: entity information fields, relationship type fields, and attribute information fields. It is the key data format connecting fine-grained clustering results with knowledge graph storage, solving the compatibility problem of importing unstructured text analysis results into structured graph databases. Then, when importing using Neo4j's LOAD CSV (Comma-Separated Values) statement, the statement maps the CSV (Comma-Separated Values) data to the "node-relationship" structure of the graph database according to the field mapping, ensuring lossless data storage and guaranteeing the integrity of the graph structure. Compared with traditional table storage, this structured storage has the advantage that the graph structure can directly retain multi-level relationships between entities, avoiding the efficiency loss of multi-table join queries and improving query response speed. The visualization interface can use colored nodes to distinguish fault types, causal factors, etc., and relationships are connected by labeled edges. It also supports quick location of association paths and output of results through Cypher queries. This process, based on the frequency, weight, and other attributes stored in the graph structure, can automatically identify high-frequency causes and provide accurate suggestions for operation and maintenance.

[0031] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for analyzing the causes of track circuit faults based on text record data mining, characterized in that: Includes the following steps: S1. Text Collection and Preprocessing: Collect track circuit fault texts and process them into standardized fault texts. ; S2, Dual-view text feature extraction and SMOTE algorithm: Extracting standardized fault text from the output of S1. The semantic and word order features are analyzed, and then fused together. After optimization using the SMOTE algorithm, a fault text feature set is output. ; S3. Fault Text Classification Based on FEML Model: Construct a three-layer FEML model and classify the fault text feature set output from S2. Input the FEML model, and after parameter optimization, parallel training of base learners and integration with random forest meta-learners, output the major category labels corresponding to the fault texts; S4. Fault text clustering based on Dirichlet multinomial mixture model: For the major category labels output by S3, combine TF-IDF weights and Dirichlet multinomial mixture model to extract and output the fine-grained fault types and causes under the major category labels. S5. Construct a visual knowledge graph of fault causes: Use Neo4j to store the fault types and causes output by S4 and construct a visual knowledge graph. Then, use Cypher query language to perform multi-dimensional correlation analysis and output the analysis results of fault types and causes.

2. The method for analyzing the causes of track circuit faults based on text record data mining according to claim 1, characterized in that: S1 specifically includes the following steps: The collected track circuit fault texts were cleaned and standardized by referring to a railway professional thesaurus, transforming non-standard texts into standard fault texts. .

3. The method for analyzing the causes of track circuit faults based on text record data mining according to claim 2, characterized in that: S2 specifically includes the following steps: S21. Standard fault text output by S1 An improved mutual information algorithm is used to extract semantic dimension features, and a distributed memory paragraph vector model is used to extract word order dimension features, thus constructing a dual-view feature extraction framework. Among them, through word frequency To calculate semantic features in the semantic dimension of text The frequency of occurrence is calculated using the following formula: （1）； in, Semantic features; semantic features Number of times it appears; Total number of semantic features; word frequency semantic features Distribution frequency and semantic features in faulty text datasets With word frequency Proportional; In addition, through word frequency and average word frequency Calculate semantic features In the major categories of faults The frequency of occurrence in [the text] is calculated using the following formula: （2）； （3）； in, , semantic features In the major categories of faults The number of times it appears in; Total number of fault categories; semantic features In the major categories of faults The frequency of occurrence in [the context of troubleshooting] and in the major categories of troubleshooting The frequency of occurrence of faults outside the major categories is inversely proportional to the frequency of category words. and average word frequency Numerical and semantic features Representative categories of faults Proportional; S22, Based on word frequency , word frequency Average word frequency The system uses metrics to filter out faulty texts whose metric values ​​are significantly higher than other metrics, identifying them as high-frequency faulty texts. Then, the selected high-frequency fault texts The data structure is optimized using the SMOTE algorithm, and a fault text feature set is output. .

4. The method for analyzing the causes of track circuit faults based on text record data mining according to claim 3, characterized in that: S3 specifically includes the following steps: S31. Based on the FEML model, the fault text feature set output by S22... The data were mapped to numerical labels and then divided into training and test sets in a 7:3 ratio. S32. Based on the training set, use the first layer of the FEML model, employ K-fold cross-validation to train the model performance, and combine grid search cross-validation to optimize the base learner parameters. S33. Based on the optimized parameters of the first layer of the FEML model, the second layer of the FEML model is used to train three base learners in parallel: gradient boosting decision tree, linear support vector machine and ridge regression classifier, to capture the features of fault text from different perspectives and output intermediate results. S34. Using the third layer of the FEML model, the intermediate results output from the second layer are input into the random forest meta-learner for integration, and the final fault text is output. The corresponding category tags and their corresponding fault types are: Track line cause A, Power supply cause B, Equipment cause C, Environmental cause D, Construction and human-related cause E, and System and software cause F.

5. The method for analyzing the causes of track circuit faults based on text record data mining according to claim 4, characterized in that: S4 specifically includes the following steps: S41, Final fault text based on the output of S34 For the corresponding category labels, a dynamic feature weight adjustment mechanism is introduced, where the feature weights... The specific calculation formula is as follows: （4）； The final fault text after S3 processing The characteristics are , It is a feature The weight, It is a feature In the final fault text Corresponding category tags Weight value; S42. Adjust the dynamically weighted features. Input the Dirichlet multinomial mixture model to perform fine-grained clustering mining on the fault category data, and extract and output the fine-grained fault type and cause information under the corresponding category label.

6. The method for analyzing the causes of track circuit faults based on text record data mining according to claim 5, characterized in that: S5 specifically includes the following steps: S51. Based on the Neo4j graph database, the entities and their relationships of the fine-grained fault types and causal information output by S4 are integrated and stored as a graph structure knowledge base, and a visualization interface is built to present the entities and their relationships in the graph structure knowledge base in an intuitive form. S52. Retrieve the graph structure knowledge base generated in S51 using the Cypher query language, perform multi-dimensional correlation analysis based on fault type and cause information, and output the analysis results.

7. The method for analyzing the causes of track circuit faults based on text record data mining according to claim 3, characterized in that: In step S21, an improved mutual information algorithm is introduced for the semantic dimension. This algorithm uses average class frequency... Together with the inverted category frequency, they constitute a composite weighting factor, which dynamically adjusts the feature weights; The word order dimension establishes the connection between the grammatical structure of faulty text sequences and their context through a distributed memory paragraph vector model, thereby constructing a highly discriminative temporal feature vector.

8. The method for analyzing the causes of track circuit faults based on text record data mining according to claim 7, characterized in that: The improved mutual information algorithm uses inverted class frequency to measure the frequency of feature occurrence in non-target categories, thereby enhancing the representativeness of features for the target fault class.

9. The method for analyzing the causes of track circuit faults based on text record data mining according to claim 7, characterized in that: The Distributed Memory Phrase Vector Model is a neural network model for learning vector representations of text phrases. This model maps faulty text phrases to fixed-length feature vectors, captures faulty text information, and extracts word order features. To mitigate the impact of random fluctuations during feature vector training, the distributed memory paragraph vector model defines stable feature vectors for faulty text by independently repeating the training process multiple times and calculating the arithmetic mean of the feature vectors. Stable eigenvectors The specific calculation formula is as follows: （5）； in, For the first i The fault text is in the first Feature vectors generated in the next iteration of training. This represents the total number of iterations.

10. The method for analyzing the causes of track circuit faults based on text record data mining according to claim 6, characterized in that: The fine-grained fault types and cause information output by S42 are written to a CSV file using the LOAD CSV statement in Neo4j. The data in the CSV file is then imported into the Neo4j graph database, providing a structured data foundation for the construction of a knowledge graph of track circuit fault causes.

Images