Evidence-driven knowledge flow construction and intelligent response method for railway track engineering

Abstract

The invention relates to the technical field of engineering information management, in particular to an evidence-driven knowledge flow construction and intelligent response method for railway track engineering. The technical problem is that the railway track engineering knowledge management technology in the prior art has the problems of language isolation, knowledge system solidification and lack of context awareness in system decision support. According to the technical scheme, the evidence-driven knowledge flow construction and intelligent response method for the railway track engineering comprises the steps of mixed information source engineering data access and preprocessing, semantic atom extraction and cross-medium fusion, evidence-driven knowledge body dynamic evolution, full-life-cycle sequential logic verification and dynamic decision support of context awareness. Through cross-medium semantic alignment and fusion, an evidence-driven dynamic evolution mechanism and context-aware decision support, multi-source data semantics can be run through, dynamic knowledge evolution can be realized, and decision contexts can be actively understood and responded, so that crossing of railway track engineering management from passive query to active intelligent response is promoted.

Description

Technical Field

[0001] This invention relates to the field of engineering information management technology, and in particular to an evidence-driven knowledge flow construction and intelligent response method for railway track engineering. Background Technology

[0002] With the expansion of railway engineering construction scale and the increasing requirements for refined operation and maintenance, massive amounts of multi-source heterogeneous data are generated throughout the entire life cycle of the project, including design specifications, construction drawings, test reports, and sensor monitoring time series data. These data contain key knowledge supporting engineering decision-making, but current knowledge management technologies have significant limitations, which restrict the improvement of the level of intelligence.

[0003] Existing technologies first face the problem of "semantic isolation." Traditional methods often rely on single-modal analysis, such as using keyword-based natural language processing to parse text or using image processing to identify components in drawings. These methods lack the ability to analyze and integrate deep semantic relationships between text, images, and time-series data. For example, textual descriptions of "rail corrugation," their corresponding rail inspection vehicle waveforms, and vibration acceleration time-series data for specific sections are usually stored and processed independently in existing systems. This leads to fragmented knowledge describing the same engineering entity or state, forming a "cross-modal semantic gap" that cannot provide a unified and complete view of evidence for decision-making.

[0004] Secondly, existing knowledge systems are mostly in a "static and fixed" state. Common knowledge graph construction methods are mostly statically generated based on existing data at a specific point in time, with long update cycles and reliance on manual intervention. However, the state of railway track engineering (such as equipment wear and geometric deformation) is constantly evolving, with new inspection data, maintenance records, and case reports constantly being generated. Static knowledge systems cannot absorb this new evidence in real time, resulting in outdated knowledge base content that cannot reflect the latest on-site conditions and is not suitable for dynamic and evolving cognitive scenarios.

[0005] Furthermore, existing systems suffer from a lack of context awareness in their decision support capabilities. These systems often provide general knowledge retrieval results rather than conclusive recommendations tailored to specific tasks, roles, and business objectives. For example, in a disease remediation scenario, the focus and required knowledge granularity of design engineers, construction engineers, and maintenance engineers differ significantly. Existing systems lack a deep understanding of and proactive adaptation to the user's decision-making context, failing to generate focused, accurate, and actionable decision support information, making it difficult for knowledge services to directly empower frontline operations.

[0006] Therefore, to address the above issues, we propose an evidence-driven knowledge flow construction and intelligent response method for railway track engineering. Through cross-media semantic alignment and fusion, evidence-driven dynamic evolution mechanism, and context-aware decision support, this method can connect the semantics of multi-source data, realize dynamic knowledge evolution, and proactively understand and respond to decision-making contexts, thereby promoting the leap from passive querying to proactive intelligent response in railway track engineering management. Summary of the Invention

[0007] To overcome the problems of language isolation, rigid knowledge system, and lack of context awareness in the existing railway track engineering knowledge management technology.

[0008] The technical solution of this invention is: an evidence-driven knowledge flow construction and intelligent response method for railway track engineering, comprising the following steps: S1: Hybrid Source Engineering Data Access and Preprocessing: Accessing multi-source heterogeneous data from the entire lifecycle of railway track engineering, including at least engineering documents, design drawings, and sensor time-series data; performing format recognition and parsing on the data, and attaching spatiotemporal context information; S2: Semantic Atom Extraction and Cross-Media Fusion: For the multi-source heterogeneous data, textual semantic features are extracted using a domain adaptive model, and image features are extracted using a visual network to form a preliminary semantic atom representation; based on the normative semantic objects in the preset domain knowledge body, semantic alignment and information fusion are performed on semantic atoms from different media to generate a unified semantic representation. S3: Evidence-driven dynamic evolution of knowledge body: Construct a multi-dimensional evidence weight evaluation model, which at least considers the credibility of data sources, the freshness of evidence, and semantic consistency; Based on the evidence weight evaluation model, perform reliability evaluation and conflict resolution on the semantic atoms, and incrementally update and maintain the domain knowledge body according to the evaluation results to form a dynamically evolving knowledge base; S4: Full lifecycle temporal logic verification: Based on the temporal constraint rules of each stage of the full lifecycle of railway track engineering, perform temporal consistency verification on the semantic atoms and relationships with time attributes in the domain knowledge body, identify and mark logical errors that violate the temporal constraint rules; S5: Context-aware dynamic decision support: Receives user task queries, combines user roles and engineering constraints to generate a constraint-aware task context vector; Based on the task context vector, dynamically extracts relevant semantic subgraphs from the dynamically evolving domain knowledge body, and uses graph reasoning technology to generate serialized decision suggestions and support information; The feedback data generated after the application of the decision support information generated in step S5 is fed back to the evidence weight evaluation model in step S3 as new evidence, driving the continuous evolution of the domain knowledge body and forming a closed-loop knowledge flow of "data-knowledge-decision-feedback".

[0009] As a preferred option, the extraction of text semantic features using a domain-adaptive model in step S2 specifically involves: using a language model that has been pre-trained on a corpus in the field of railway track engineering to perform deep semantic analysis on engineering documents, logs, and specification texts to extract engineering concepts, attributes, and their relationships.

[0010] Preferably, the conflict resolution in step S3 specifically involves: when the evidence weight evaluation model identifies a new semantic atom that logically contradicts existing knowledge in the domain knowledge body, the multi-dimensional evidence weight evaluation model is used to judge the comprehensive reliability score of both sides of the contradiction, and a voting mechanism or path backtracking mechanism based on the reliability score is adopted to correct, supplement or eliminate the corresponding knowledge in the domain knowledge body.

[0011] Preferably, the multi-dimensional evidence weighting evaluation model in step S3 adopts a weighted fusion mechanism based on multi-source evidence. Its evaluation process comprehensively considers the source credibility, time freshness, topological importance of the semantic atom in the domain knowledge body, and semantic consistency with other semantic atoms.

[0012] Preferably, the semantic alignment and information fusion in step S2 specifically involves: measuring the semantic correlation between text features and image features by calculating semantic correlation based on interactive attention; using the standardized semantic objects in the domain knowledge body as alignment anchors, fusing multi-media features with correlation higher than a preset threshold to form an enhanced semantic representation.

[0013] Preferably, the dynamic extraction of relevant semantic subgraphs in step S5 specifically involves: using the task context vector as a guide, calculating the relevance weights of each node and edge in the domain knowledge body to the task context through an attention mechanism, selecting key nodes and associated paths from the domain knowledge body based on the relevance weights, and generating a contextualized semantic subgraph focused on the current task.

[0014] Preferably, the step S5 of generating the constraint-aware task context vector specifically involves: parsing the semantics of the user's task query, matching and calculating the similarity between it and a preset engineering domain constraint rule library; combining the matching similarity with the current satisfaction state of each constraint rule, weightedly fusing the semantic representation of the constraint rule to generate the task context vector that simultaneously encodes the task intent and engineering constraints.

[0015] As a preferred embodiment, the timing consistency verification in step S4 specifically involves: constructing timing constraint rules based on the chronological order, minimum time interval, and parallel or mutually exclusive relationships defined for each stage of railway track engineering; traversing the semantic atoms and relationships with timestamps or stage markers in the domain knowledge body, checking whether their temporal logic violates the timing constraint rules, and marking and quantifying the instances of violation.

[0016] Preferably, the method further includes: integrating and presenting the serialized decision suggestions and supporting information, as well as the semantic subgraph on which the reasoning is based and the related multi-source evidence chain, through a visual interface.

[0017] Preferably, the method is applicable to intelligent compliance review, maintenance and repair decision support, or safety risk early warning application scenarios in railway track engineering.

[0018] The beneficial effects of this invention are: This invention constructs a closed-loop knowledge flow consisting of "multi-source fusion, evidence evaluation, dynamic evolution, temporal verification, and context awareness." Specifically, through cross-media semantic alignment and fusion, it breaks down semantic barriers between text, images, and data, forming a unified knowledge view. Through an evidence-driven dynamic evolution mechanism, the knowledge base can autonomously absorb new evidence, resolve contradictions, and update in real time along with the engineering process, maintaining the freshness of knowledge. Through context-aware decision support, it can understand the user's specific task scenario, accurately extract relevant subgraphs from the dynamic knowledge base, and perform reasoning to generate targeted and interpretable decision suggestions, achieving a leap from passive querying to proactive intelligent response. This invention provides a knowledge processing method for the intelligent management of the entire lifecycle of railway track engineering. Attached Figure Description

[0019] Figure 1 The diagram illustrates the steps of the evidence-driven knowledge flow construction and intelligent response method for railway track engineering according to the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Please see Figure 1 This invention provides an embodiment of an evidence-driven knowledge flow construction and intelligent response method for railway track engineering, comprising the following steps: S1: Hybrid Source Engineering Data Access and Preprocessing: Accessing multi-source heterogeneous data from the entire lifecycle of railway track engineering, including at least engineering documents, design drawings, and sensor time-series data; performing format recognition and parsing on the data, and attaching spatiotemporal context information; S2: Semantic Atom Extraction and Cross-Media Fusion: For multi-source heterogeneous data, a domain-adaptive model is used to extract textual semantic features, and a visual network is used to extract image features to form a preliminary semantic atom representation; based on the normative semantic objects in the pre-defined domain knowledge body, semantic alignment and information fusion are performed on semantic atoms from different media to generate a unified semantic representation. S3: Evidence-driven dynamic evolution of knowledge body: Construct a multi-dimensional evidence weight evaluation model, which at least considers the credibility of data sources, the freshness of evidence, and semantic consistency; Based on the evidence weight evaluation model, perform reliability evaluation and conflict resolution on semantic atoms, and incrementally update and maintain the domain knowledge body according to the evaluation results to form a dynamically evolving knowledge base; S4: Full lifecycle temporal logic verification: Based on the temporal constraint rules of each stage of the railway track engineering lifecycle, perform temporal consistency verification on semantic atoms and relationships with time attributes in the domain knowledge body, identify and mark logical errors that violate the temporal constraint rules; S5: Context-aware dynamic decision support: Receives user task queries, combines user roles and engineering constraints to generate constraint-aware task context vectors; Based on task context vectors, dynamically extracts relevant semantic subgraphs from dynamically evolving domain knowledge bodies, and uses graph reasoning technology to generate serialized decision suggestions and support information; In this process, the feedback data generated after the application of the decision support information generated in step S5 is used as new evidence to flow back to the evidence weight evaluation model in step S3, driving the continuous evolution of the domain knowledge body and forming a closed-loop knowledge flow of "data-knowledge-decision-feedback".

[0022] Furthermore, in step S2, a domain-adaptive model is used to extract semantic features from the text. Specifically, a language model that has been pre-trained on a corpus in the railway track engineering domain is used to perform deep semantic analysis on engineering documents, logs, and specification texts to extract engineering concepts, attributes, and their relationships. While the general pre-trained language model has learned extensive semantic knowledge on a general corpus, it lacks a deep understanding of engineering terminology, expression habits, and contextual logic. By using a domain corpus for further pre-training, the model's parameters can be adjusted to make its internal representation more consistent with the engineering context. This allows for more accurate extraction of engineering concepts (such as "turnout" and "seamless track"), attributes (such as "gauge" and "superelevation"), and their relationships (such as "located in" and "belong to") from complex texts. This effectively solves the problems of ambiguous professional terms and low-frequency word recognition, improving the accuracy and domain adaptability of knowledge extraction.

[0023] Furthermore, in step S3, conflict resolution specifically involves the following: When the evidence weight evaluation model identifies a logical contradiction between a new semantic atom and existing knowledge in the domain knowledge body, the system judges the comprehensive reliability score of both contradictory parties based on the multi-dimensional evidence weight evaluation model. A voting mechanism or path backtracking mechanism based on the reliability score is then used to correct, supplement, or eliminate the corresponding knowledge in the domain knowledge body. The workflow involves the system tracing the evidence chains (source, time, other supporting knowledge, etc.) of each contradictory knowledge and calculating their comprehensive credibility. A voting mechanism based on the reliability score can be used, i.e., adopting the party with higher credibility and updating the status of the party with lower credibility (e.g., marking it as "pending verification" or "historical version"); or a path backtracking mechanism can be used to check for intermediate reasoning errors or missing evidence. This mechanism enables the knowledge base to adaptively handle information contradictions, achieve self-correction and purification of knowledge, and ensure the logical consistency of the knowledge base during dynamic updates.

[0024] Furthermore, in step S3, the multi-dimensional evidence weight evaluation model adopts a weighted fusion mechanism based on multi-source evidence. Its evaluation process comprehensively considers the source credibility, time freshness, topological importance in the domain knowledge body, and semantic consistency with other semantic atoms of the semantic atom. For example, the overall reliability score The weighted sum can be calculated using the following general form:

[0025] in, This indicates a rating based on the authority of the data source. This indicates a score based on the freshness of the evidence (newer evidence usually has higher weight). The score represents the consistency with other relevant semantic atoms in the knowledge base. This represents a score based on the importance of the semantic atom in the topology of the domain knowledge body (such as a node on a critical path); The corresponding adjustable weight coefficients satisfy the normalization condition; by quantifying and synthesizing multi-dimensional evidence, this model provides a dynamically changing credibility measure for each knowledge unit, which is a quantitative basis for driving knowledge evolution.

[0026] Furthermore, in step S2, semantic alignment and information fusion specifically involve: calculating semantic correlation based on interactive attention to measure the degree of semantic correlation between text features and image features; for example, for a track structure image and a text describing "sleeper cracks", the system will calculate the similarity between the features of each visual region in the image and the features of the text; the principle is to use a neural network to learn a common semantic space in which different modal features describing the same entity should be close to each other. Using the normative semantic objects in the domain knowledge body as alignment anchors, multi-media features with a correlation higher than a preset threshold are fused to form an enhanced semantic representation. For example, the visual feature region of "crack" with a high correlation is bound to the text description "crack" and jointly associated with the normative entity "Type II concrete sleeper" to form an enhanced semantic representation that includes both visual feature description and text attribute description. This effectively solves the cross-modal semantic gap and realizes multi-dimensional complementarity and enhancement of knowledge.

[0027] Furthermore, step S5 dynamically extracts relevant semantic subgraphs. Specifically, guided by the task context vector, the relevance weights of each node and edge in the domain knowledge body to the task context are calculated using an attention mechanism. Based on these relevance weights, key nodes and related paths are selected from the domain knowledge body to generate a contextualized semantic subgraph focused on the current task. The process involves using the task context vector as a query and the features of nodes in the knowledge graph as keys. Attention scores are calculated using dot products or neural networks, reflecting the relevance of the node's knowledge to the current task. Based on the relevance weights, key nodes and related paths are selected from the vast domain knowledge body, automatically "cropping" a small-scale, highly relevant contextualized semantic subgraph focused on the current task. This method avoids the inefficiency of global search in massive amounts of knowledge and achieves precise delivery and focus of knowledge services.

[0028] Furthermore, in step S5, a constraint-aware task context vector is generated. Specifically, the semantics of the user's task query are parsed, and the query is matched and similarity is calculated with a preset engineering domain constraint rule library. Combining the matching similarity with the current satisfaction status of each constraint rule, the semantic representation of the constraint rule is weighted and fused to generate a task context vector that simultaneously encodes the task intent and engineering constraints. For example, if a user queries "roadbed grouting reinforcement during the rainy season," the system will match constraint rules such as "safety specifications for construction in severe weather" and "rain protection requirements for procedures," and combine current meteorological data to determine the satisfaction level of these rules, ultimately generating a task context vector that encodes complex information such as "grouting reinforcement," "rainy season," and "needs to meet specific safety and process constraints." This gives the subsequent knowledge retrieval and reasoning process a built-in compliance check perspective.

[0029] Furthermore, the timing consistency verification in step S4 specifically involves: constructing timing constraint rules based on the defined sequence, minimum time interval, and parallel or mutually exclusive relationships of each stage of railway track engineering; traversing semantic atoms and relationships with timestamps or stage markers in the domain knowledge body to check whether their temporal logic violates the timing constraint rules, and marking and quantifying the violations; for example, if it is found that the time of "bridge track laying completed" is earlier than the time of "bridge bearing installation completed" in the knowledge base, it is identified as a violation of the "installation before track laying" process timing constraint; the system can mark the violations as violations and use the following formula for quantification to assess the severity of the conflict:

[0030] Here, Δmin is the minimum interval required by the specification, and σ is the tolerance parameter; this mechanism ensures the logical rationality of engineering knowledge in the temporal dimension.

[0031] Furthermore, the method also includes integrating and presenting the serialized decision recommendations and supporting information, as well as the semantic subgraphs on which the reasoning is based and the related multi-source evidence chains, through a visual interface; for example, when recommending "replacing a damaged rail", the method simultaneously displays the images of the rail's previous flaw detection reports (Evidence 1), the most recent track geometry inspection over-limit data table (Evidence 2), and the traffic load curve of the section (Evidence 3); this provides users with transparent basis for decision-making, enhancing the interpretability of the results and user trust.

[0032] Furthermore, the method is applicable to intelligent compliance review, maintenance and repair decision support, or safety risk early warning application scenarios in railway track engineering; by configuring different domain knowledge bodies and constraint rule bases, the closed-loop knowledge flow architecture of this method can adapt to a variety of specific engineering business needs.

[0033] Through the above steps, this invention constructs a closed-loop knowledge flow of "multi-source fusion - evidence evaluation - dynamic evolution - temporal verification - context awareness". Specifically: by cross-media semantic alignment and fusion, the semantic barriers between text, images, and data are broken down, forming a unified knowledge view; through an evidence-driven dynamic evolution mechanism, the knowledge base can autonomously absorb new evidence, resolve contradictions, and update in real time along with the engineering process, maintaining the freshness of knowledge; through context-aware decision support, it can understand the user's specific task scenario, accurately extract relevant subgraphs from the dynamic knowledge base and perform reasoning, generating targeted and interpretable decision suggestions, realizing a leap from passive query to proactive intelligent response; this invention provides a knowledge processing method for the intelligent management of the entire life cycle of railway track engineering.

[0034] Example 2 Optionally, this embodiment provides a basic implementation flow of the method.

[0035] First, an initial domain knowledge body is constructed or integrated. This knowledge body is preferably implemented in the form of a knowledge graph, where the nodes represent standard entities in railway track engineering (such as "Type III sleeper", "CRTS II type slab track", "turnout", etc.) and the edges represent the relationships between entities (such as "located in", "belongs to", "has attributes", etc.). This knowledge graph can be initially constructed based on existing structured or semi-structured data such as design specifications and standard drawing sets.

[0036] Step S1: Access and preprocessing of mixed source engineering data; This step establishes a unified data access layer; it connects new data streams from the engineering site and management system through multiple interface protocols: Sensor timing data: The system accesses the network of track inspection vehicles, dynamic inspection vehicles, and ground monitoring sensors deployed on the line via the MQTT protocol to receive real-time timing data streams of track geometry, vehicle dynamics, and environment. Engineering documents: Newly uploaded design change orders, construction logs, and inspection reports (PDF / WORD format) are asynchronously retrieved from project management systems (such as PMS) and document management systems via API calls or file listening. Design drawings and site images: Access CAD design drawings (DWG format), engineering BIM model slices, and digital photos taken by manual inspections via file interface; After access, the system first performs format recognition and parsing: for text / document data, it performs OCR (if it is a scanned document) and text extraction; for images, it decodes them into a pixel matrix; for time-series data, it parses their data packet structure; then, it adds spatiotemporal context information: it automatically extracts or associates timestamps (such as report date, data collection time), spatial locations (such as line mileage markers, work site names), and equipment / component IDs from data metadata or content; for example, for a photo of a sleeper crack taken on "09-27", it automatically labels the associated line section "K205+300 to K205+350" and the possible component type "prestressed concrete sleeper"; Step S2: Semantic atom extraction and cross-media fusion; The preprocessed data is then distributed to different feature extraction modules: Text semantic atom extraction: The extracted text (such as the descriptive paragraph in the detection report) is input into a domain-adaptive pre-trained text understanding model (e.g., a model based on the Transformer architecture); the model transforms the text into a vector representation containing contextual semantics and identifies entity, attribute, and relation triplets; for example, from the text "A nuclear damage of about 15mm in length was found on the left rail at K205+320", the semantic atoms are extracted as follows: entity "rail (left rail, K205+320)", attribute "damage type: nuclear damage", and attribute "length: 15mm". Image feature extraction: Engineering drawings and site photos are input into a convolutional neural network (CNN) for visual feature extraction; the network outputs high-level feature maps or feature vectors to characterize visual patterns in the image; for drawings, components such as "sleepers" and "fasteners" and their spatial layout can be identified; for photos, visual features such as "cracks", "rust", and "missing parts" can be identified. Cross-media fusion: The system uses the canonical entities within the knowledge graph established in step S1 as alignment anchors; for example, the extracted "rail (left rail, K205+320)" in the text attempts to associate it with the "rail" entity and its specific instances in the knowledge graph; at the same time, it calculates the similarity (such as cosine similarity) between the "crack" visual feature vector extracted by CNN and the "nuclear injury" semantic vector in the text in the shared semantic space; if the similarity exceeds the threshold, it is determined that the two features from different modalities describe the same fact (here, damage), and the visual feature vector is fused as an additional attribute ("visual evidence feature") of the damage semantic atom to generate a unified semantic representation that links canonical entities, text descriptions, and visual features simultaneously; Step S3: Evidence-driven dynamic evolution of knowledge bodies; This step introduces a multi-dimensional evidence weighting evaluation model to assess the reliability of the newly generated or updated semantic atoms in step S2; this model calculates a comprehensive reliability score R(e); for example, a simplified linear weighted model is as follows: R(e) = w_source * S(e) + w_time * T(e) + w_consistency * C(e) in: e: The semantic atom to be evaluated; S(e): Source credibility score; for example, data from "national-level testing center reports" scores higher than "team inspection records"; T(e): Time freshness score; generally, newer evidence scores higher, and can be calculated using a time decay function; C(e): Semantic consistency score; evaluates the logical consistency degree of this atom with existing relevant facts in the knowledge graph (such as the historical states of the same components, the general states of similar components). w_source, w_time, w_consistency: Weight coefficients for each dimension respectively, and their sum is 1, which can be set according to engineering experience. The system decides whether and how to update the knowledge graph based on the R(e) score; for a new atom, if the score is higher than the acceptance threshold, it is inserted into the graph; if there is a logical conflict between the new atom and the existing atoms in the graph (such as a contradiction between the new and old states), conflict resolution is initiated: compare the R(e) scores of the conflicting parties and the evidence chains supporting them, and preferentially retain or fuse the party with higher reliability, and may mark the other party as "historical version" or "doubtful"; this process realizes the incremental update and maintenance of the domain knowledge body. Step S4: Full life-cycle time-series logic verification. This step builds in a time-series constraint rule library that defines the logical constraints between different stages of railway engineering (survey, design, construction, operation and maintenance, etc.); for example: "The completion time of construction drawing design" must be earlier than "The start time of on-site construction"; "The start time of joint commissioning and testing" must be later than "The completion time of track laying" and the interval is not less than N days. The system scans all the nodes and relationships with timestamps in the knowledge graph regularly or after knowledge update (such as "Activity A, end time: T1", "Activity B, start time: T2", "Relationship: B starts after A is completed"); checks whether these time information violates the constraints in the rule library through a time-series reasoning engine; once a violation is found (such as T2 < T1), a time-series conflict exception record is generated, marking the relevant nodes and relationships, and an alarm can be sent through the interface to prompt the engineering management personnel to verify. Step S5: Situation-aware dynamic decision support. When a user (such as an operation and maintenance engineer) inputs a task query at the front end of the system, such as "Analyze the reasons and proposed measures for the deterioration of track geometry near K205+300": Generate a task situation vector: The system analyzes the query text, identifies the core entities ("K205+300", "track geometry") and intentions ("analyze reasons", "proposed measures"); combines the user's role ("operation and maintenance engineer") and predefined business constraints (such as "maintenance skylight time limit", "cost control requirements"), and encodes this information into a multi-dimensional constraint-aware task situation vector. Dynamic extraction of semantic subgraphs: Using the context vector as a "probe", the system runs a graph attention network (GAT) on a dynamically updated knowledge graph; the network calculates the relevance weights of each node and edge in the graph to the context vector; based on the weights, the system automatically "crops" a focused contextualized semantic subgraph, which may include: track structure entities near K205+300, recent geometric detection over-limit data, historical maintenance records of this section, related track environment information (such as curve radius, gradient), and similar cases, etc.

[0037] Decision-making recommendations are generated: On the extracted sub-map, graph neural networks (GNNs) are used for multi-hop reasoning and relational reasoning; for example, it is inferred that "geometric deterioration" may be highly correlated with "specific type of sleeper damage", which may in turn originate from "specific traffic load" and "historical maintenance process"; finally, the system serializes the reasoning path and conclusion into decision recommendations, such as: "1. It is recommended to conduct special flaw detection on the sleepers near K205+300; 2. If the damage is confirmed, it is recommended to replace them using XX process during the next maintenance window; 3. Monitor the traffic volume of this section in the long term and consider adjusting the preventive maintenance cycle." Closed-loop feedback: After the engineer adopts and implements the suggestion, the relevant maintenance records, new test data and other feedback data flow into step S1 as new data sources. After being processed and evaluated, these data are used to verify or correct the knowledge on which the previous reasoning was based (for example, verifying that "a certain maintenance process is effective for this type of damage"), thereby driving the knowledge body into the next round of evolution cycle, forming a continuously optimized closed-loop knowledge flow of "data 1 → knowledge 1 → decision 1 → feedback data 2 → knowledge 2 → ...".

[0038] Example 3 Optionally, this embodiment further elaborates on the preferred implementation details of the technical features of deep integration and dynamic evolution based on embodiment 2.

[0039] Specific implementation details of the domain-adaptive model: The construction of a domain-adaptive model is crucial; firstly, a large-scale corpus of railway track engineering data is collected, including: Unstructured texts: design specifications, construction organization designs, academic papers, technical regulations (such as the "Railway Track Design Code"), failure case reports, etc. Semi-structured text: engineering logs, test report templates, bills of materials, etc.; Next, a general pre-trained language model (such as BERT, RoBERTa, or ERNIE) is selected as the base. Using the aforementioned domain corpus, a continuous pre-training strategy is employed to train it. The training task typically includes masked language modeling (MLM), which enables the model to learn the contextual semantics of engineering terminology (e.g., "fastener" may refer to specific components such as "elastic clip" or "bolt" in different contexts). After training, compared to the original general model, this model shows a significant improvement in accuracy and recall when extracting professional concepts and relationships such as "seamless track stress relief" and "turnout tightness inspection".

[0040] Specific implementation details regarding evidence evaluation and cross-media fusion: Evidence weighting evaluation models can employ more complex nonlinear fusion methods; for example, using a multilayer perceptron (MLP) as the evaluator: R(e) = MLP( S(e), T(e), C(e), P(e) ) Where P(e) represents the topological importance score; it is obtained by calculating the centrality index of the semantic atom node in the graph (such as eigenvector centrality, between centrality), and nodes on the critical path are usually more important; In cross-media semantic alignment, the interactive attention mechanism operates as follows: Given a text feature sequence T = {t1, t2, ..., tm} and an image region feature sequence V = {v1, v2, ..., vn}; Calculate an m x n similarity matrix M, where the element M_{ij} = sim(t_i, v_j), where sim can be the cosine similarity or calculated through a small neural network; Softmax normalization is applied to the rows and columns of M respectively to obtain the attention weights of text to image A^{t->v} and image to text A^{v->t}. Based on attention weights, calculate the text-aware image features v'_i = sum_j(A^{v->t}_{ij} * t_j) and the image-aware text features t'_j = sum_i(A^{t->v}_{ji} * v_i); Finally, by using a gating mechanism or splicing method, v_i and v'_i, t_j and t'_j are fused to obtain a deeply fused cross-modal feature representation, which is then associated with the knowledge graph anchor points.

[0041] Specific implementation details of temporal conflict detection quantization: The quantization scoring formula for timing conflict detection can be specifically expressed as: ConflictScore(e_a, e_b) = max( 0, [t(e_b) - t(e_a) - Δ_min_required] / σ_t ) e_a, e_b: Two semantic atoms (activities or events) with timing dependencies; t(e_a), t(e_b): The occurrence times (or completion / start times) of e_a and e_b; Δ_min_required: The minimum required time interval from the completion of e_a to the start of e_b as required by engineering specifications; if it is required that e_b starts immediately after the completion of e_a, then Δ_min_required = 0; if it is required to perform the next process after N days of maintenance, then Δ_min_required = N; σ_t: The time tolerance parameter, used to normalize the time difference to a reasonable scoring scale; The meaning of this formula is: Only when the start time of e_b is earlier than the completion time of e_a plus the minimum required interval (i.e., t(e_b) < t(e_a) + Δ_min_required), a positive conflict score will be generated; the larger the conflict score, the more serious the violation of the timing constraint; the system can set a threshold to give priority warnings to events with high conflict scores.

[0042] Example 3 Optionally, based on Example 2, this example details the preferred implementation manner of the technical features of intelligent decision support.

[0043] Regarding the specific implementation of the conflict resolution strategy: When a knowledge conflict occurs, the system uses a collaborative voting and traceability mechanism based on reliability scoring to resolve it; the specific process is as follows: Conflict detection: Identify two or more sets of conflicting semantic atoms {e_conflict} regarding the same entity or fact; Evidence pool collection: For each atom e_i in the conflict set, collect all direct and indirect evidence supporting it to form an evidence pool Evid(e_i); the evidence includes its original data source, the timestamp at the time of extraction, other highly credible atoms supporting it, etc.; Comprehensive scoring calculation: For each e_i, not only calculate its own reliability score R(e_i), but also calculate the overall strength score R_evid(e_i) of its evidence pool, for example, taking the weighted average of the evidence scores in the pool; Voting and Decision-Making: Compare (R(e_i), R_evid(e_i)) of all conflicting atoms; if one side's score is significantly higher than the others (e.g., exceeding the threshold), then that side is determined to "win", its knowledge is updated to the graph, and the other conflicting atoms are marked as "overturned", and the credibility of the evidence associated with them is downgraded; Path backtracking: If the scores of all parties are close and a simple decision cannot be made, path backtracking is initiated; weak links in the chain of evidence are examined, such as the source of a key piece of evidence being questionable or its origin being outdated; manual intervention is prompted to review this link, or a new data collection task (such as retesting) is initiated for this point of contention, and the new evidence is used as the decisive factor.

[0044] Specific implementation details regarding dynamic subgraph and context vector generation: Detailed process of task context vector generation: After parsing the user query q, semantic matching is performed with each rule in the constraint rule base {c_1,c_2, ..., c_n}, and the similarity Sim(q, c_i) is calculated; simultaneously, the system evaluates the satisfaction degree Sat(c_i) (between 0 and 1) of each rule based on the current project status (queried from the knowledge graph); the final context vector C is generated by the following formula: C = Σ_i [ Sim(q, c_i) * Sat(c_i) * v(c_i) ] Where v(c_i) is the semantic vector representation of rule c_i; the generated C contains both the task intent and the state of satisfaction of each relevant constraint in the current context. The attention-weighted node importance algorithm used in dynamic subgraph extraction: For a node v in the knowledge graph, its importance I(v) to the current context C is calculated through the propagation of its neighbor information. I(v) = σ( W * [ h_v || Σ_{u∈N(v)} α_{uv} * h_u ] ) in: h_v, h_u: Feature vectors of node v and its neighbor u; N(v): The set of neighbors of node v; α_{uv}: attention weights, α_{uv} = softmax( LeakyReLU( a^T * [W*h_v || W*h_u]) ), where a and W are learnable parameters; || represents vector concatenation, and σ is the activation function; This formula makes the importance of a node depend not only on its own characteristics, but also on the information of its neighbors that is related to the context C; the system selects nodes with I(v) higher than the threshold and their associated edges to form a decision subgraph.

[0045] Regarding the specific implementation of visualization: The decision support interface not only displays the final text recommendations but also provides an interactive evidence tracing map; taking the decision recommendation for K205+300 in Example 2 as an example, the interface displays: Core conclusion section: Lists the sequential decision-making steps; Evidence Graph Area: An interactive graphical view with a central conclusion (e.g., "Recommendation to replace sleepers") surrounded by radial links displaying multi-source evidence nodes supporting that conclusion. Node 1 (Detection Data): Linked to the "09-27 Track Inspection Vehicle Level III Over-limit Report" segment, displaying the TQI value; Node 2 (Image Evidence): Thumbnail showing "Photos of sleeper cracks taken on-site on September 27th", click to enlarge; Node 3 (historical history): Linked to "The sleepers here have been in service for 10 years, and the last major overhaul was recorded 5 years ago"; Node 4 (Standard Basis): Links to the clause on "Standard for Judging Cracks and Damage to Concrete Sleepers" in the "Rules for Repairing Railway Lines"; Node 5 (similar case): Linked to "K180+500 similar cracks and post-replacement effect tracking record"; Users can click on any evidence node to view its original data, reliability score, and citation history; this presentation method enhances the transparency and credibility of the decision-making process.

[0046] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. An evidence-driven knowledge flow construction and intelligent response method for railway track engineering, characterized by: Includes the following steps: S1: Hybrid Source Engineering Data Access and Preprocessing: Accessing multi-source heterogeneous data from the entire lifecycle of railway track engineering, including at least engineering documents, design drawings, and sensor time-series data; The data is formatted and parsed, and spatiotemporal context information is appended. S2: Semantic Atom Extraction and Cross-Media Fusion: For the multi-source heterogeneous data, a domain adaptive model is used to extract textual semantic features, and a visual network is used to extract image features to form a preliminary semantic atom representation. Based on the predefined domain knowledge body of the normative semantic objects, semantic atoms from different media are semantically aligned and information is fused to generate a unified semantic representation. S3: Evidence-driven dynamic evolution of knowledge body: Construct a multi-dimensional evidence weight evaluation model, which at least considers the credibility of data sources, the freshness of evidence, and semantic consistency; Based on the evidence weight evaluation model, perform reliability evaluation and conflict resolution on the semantic atoms, and incrementally update and maintain the domain knowledge body according to the evaluation results to form a dynamically evolving knowledge base; S4: Full lifecycle temporal logic verification: Based on the temporal constraint rules of each stage of the full lifecycle of railway track engineering, perform temporal consistency verification on the semantic atoms and relationships with time attributes in the domain knowledge body, identify and mark logical errors that violate the temporal constraint rules; S5: Context-aware dynamic decision support: Receives user task queries, combines user roles and engineering constraints to generate a constraint-aware task context vector; Based on the task context vector, dynamically extracts relevant semantic subgraphs from the dynamically evolving domain knowledge body, and uses graph reasoning technology to generate serialized decision suggestions and support information; In this process, the feedback data generated after the application of the decision support information generated in step S5 is fed back to the evidence weight evaluation model in step S3 as new evidence, driving the continuous evolution of the domain knowledge body and forming a closed-loop knowledge flow of "data-knowledge-decision-feedback".

2. The evidence-driven knowledge flow construction and intelligent response method for railway track engineering according to claim 1, characterized in that: Step S2, which involves using a domain-adaptive model to extract semantic features from text, specifically involves using a language model that has been pre-trained on a corpus in the field of railway track engineering to perform deep semantic analysis on engineering documents, logs, and specification texts to extract engineering concepts, attributes, and their relationships.

3. The evidence-driven knowledge flow construction and intelligent response method for railway track engineering according to claim 1, characterized in that: The conflict resolution mentioned in step S3 specifically involves: when the evidence weight evaluation model identifies a new semantic atom that logically contradicts existing knowledge in the domain knowledge body, the model judges the comprehensive reliability score of both contradictory parties based on the multi-dimensional evidence weight evaluation model, and uses a voting mechanism or path backtracking mechanism based on the reliability score to correct, supplement, or eliminate the corresponding knowledge in the domain knowledge body.

4. The evidence-driven knowledge flow construction and intelligent response method for railway track engineering according to claim 1, characterized in that: The multi-dimensional evidence weighting evaluation model described in step S3 adopts a weighted fusion mechanism based on multi-source evidence. Its evaluation process comprehensively considers the source credibility, time freshness, topological importance of semantic atoms in the domain knowledge body, and semantic consistency with other semantic atoms.

5. The evidence-driven knowledge flow construction and intelligent response method for railway track engineering according to claim 1, characterized in that: The semantic alignment and information fusion described in step S2 are as follows: by calculating the semantic relevance based on interactive attention, the semantic relevance between text features and image features is measured; using the standard semantic objects in the domain knowledge body as alignment anchors, multi-media features with a relevance higher than a preset threshold are fused to form an enhanced semantic representation.

6. The evidence-driven knowledge flow construction and intelligent response method for railway track engineering according to claim 1, characterized in that: The dynamic extraction of relevant semantic subgraphs in step S5 specifically involves: using the task context vector as a guide, calculating the relevance weights of each node and edge in the domain knowledge body to the task context through an attention mechanism, selecting key nodes and associated paths from the domain knowledge body based on the relevance weights, and generating a contextualized semantic subgraph focused on the current task.

7. The evidence-driven knowledge flow construction and intelligent response method for railway track engineering according to claim 1, characterized in that: The step S5 of generating a constraint-aware task context vector specifically involves: parsing the semantics of the user's task query, matching and calculating the similarity between it and a preset engineering domain constraint rule library; combining the matching similarity with the current satisfaction state of each constraint rule, weighted fusion of the semantic representation of the constraint rule, and generating the task context vector that simultaneously encodes the task intent and engineering constraints.

8. The evidence-driven knowledge flow construction and intelligent response method for railway track engineering according to claim 1, characterized in that: The timing consistency verification in step S4 specifically involves: constructing timing constraint rules based on the chronological order, minimum time interval, and parallel or mutually exclusive relationships defined for each stage of railway track engineering; traversing the semantic atoms and relationships with timestamps or stage markers in the domain knowledge body, checking whether their temporal logic violates the timing constraint rules, and marking and quantifying the instances of violation.

9. The evidence-driven knowledge flow construction and intelligent response method for railway track engineering according to claim 1, characterized in that: The method further includes integrating and presenting the serialized decision suggestions and supporting information, as well as the semantic subgraph and related multi-source evidence chains on which the reasoning is based, through a visual interface.

10. The evidence-driven knowledge flow construction and intelligent response method for railway track engineering according to any one of claims 1-9, characterized in that: The method is applicable to intelligent compliance review, maintenance and repair decision support, or safety risk early warning application scenarios in railway track engineering.

Images