Substation engineering construction progress deviation monitoring and analysis method based on digital twinning

CN122819231APending Publication Date: 2026-09-25GUANGDONG TENGFENG ELECTRIC POWER CONSTR CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610687389.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-05-19
Publication Date
2026-09-25

AI Technical Summary

Technical Problem

对于突发的人力、设备、材料或环境变化,缺乏动态调整或迁移能力,难以适应施工阶段的多变性和复杂性

Benefits of technology

(1)通过构建具有三重属性的语义锚点池并引入受限的多跳语义传播机制,有效克服了传统变电工程施工管理系统中知识表征静态化、规则僵化的问题,实现了对施工工艺逻辑的动态适配与局部演化。相较于现有系统依赖预设固定流程或全局知识图谱重构的方式,本方案在不重建整体语义结构的前提下,仅基于时空约束模板的连续违背触发语义漂移判定,并通过邻近工序中语义兼容性最高的替代锚点实现要素权重向量的自适应调整,显著提升了系统在复杂多变现场环境下的响应灵敏度与鲁棒性;该机制避免了端到端特征融合带来的高计算开销与模型耦合问题,使得语义更新过程轻量化、可追溯且无需外部大模型支持,尤其适用于边缘侧资源受限的工程场景,大幅降低对云端算力和网络带宽的依赖,提高了系统部署的灵活性与实用性。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122819231A_ABST
    Figure CN122819231A_ABST
Patent Text Reader

Abstract

The application relates to a power transformation engineering construction progress deviation monitoring and analyzing method based on digital twinning, aiming to realize standardized semantic anchor point extraction of multi-source heterogeneous construction data, dynamic drift detection and adaptive updating. Through natural language processing and domain ontology alignment technology, process, environmental constraint and resource dependency triples are extracted and a standardized semantic anchor point pool is constructed; in combination with the field BIM model log and the environmental monitoring data, the dynamic checking and the continuity drift determination of the space-time constraint of the semantic anchor points are realized. If semantic drift occurs, the compatibility evaluation and the execution stability weighted screening replacement anchor point are used to dynamically adjust the correlation weight, so that the time sequence adaptability of the anchor point is realized. The scheme effectively improves the fracture positioning and deviation diagnosis capability of the semantic relationship in the construction process, and enhances the closed-loop self-evolution effect of semantic modeling.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of digital twin and semantic modeling technology for power engineering construction progress, and in particular to a method for monitoring and analyzing deviations in the construction progress of substation engineering based on digital twins. Background Technology

[0002] Currently, in the field of power engineering, especially in the construction progress management and digital twin application of substation projects, mainstream technical solutions mainly revolve around rule-driven progress monitoring, BIM (Building Information Modeling) integration, real-time data acquisition from sensors and the Internet of Things, and relationship modeling based on knowledge graphs. Technological development trends focus on deep fusion of multi-source data, intelligent anomaly detection, end-to-end causal analysis, and automated attribution of progress deviations.

[0003] Currently, the industry widely uses knowledge graph processing of static construction standards, work instructions, and structured sensor data. By pre-setting work processes and parameter constraints, and leveraging BIM model progress synchronization and rule-based alarm systems, preliminary warnings of major risk nodes and schedule deviations can be achieved. Some innovative research has begun to introduce big data-driven multimodal fusion, end-to-end neural networks, or large-scale pre-trained models (LLMs) for deviation cause inference, and to utilize multi-factor coupling analysis to identify abnormal patterns in complex environments. However, these solutions generally employ highly coupled, computationally intensive architectures that heavily rely on massive model parameters or cloud resources, which contradicts the high concurrency, frequent dynamic disturbances, and limited edge computing power of actual substation construction environments.

[0004] In the fields of power digital twins, semantic modeling, and schedule deviation analysis, representative technologies such as intelligent decision-making based on dynamic knowledge graphs, anomaly detection methods driven by BIM-IoT fusion data, and large-scale construction status time sequence analysis systems can effectively monitor progress and trace some causes in scenarios with large project scale, clear standard processes, and relatively fixed environmental factors. However, when faced with dynamic environmental disturbances such as constantly changing weather, material supply, mechanical failures, or temporary organizational adjustments during substation construction, traditional semantic modeling methods have limitations.

[0005] Previous technical solutions had the following core shortcomings and unmet needs: First, existing deviation identification models are mostly based on static knowledge graphs or rule templates, and the modeling of semantic anchor points (i.e., process-type nodes and their constraint attributes) is relatively rigid. They lack the ability to dynamically adjust or migrate in the face of sudden changes in manpower, equipment, materials or environment, making it difficult to adapt to the variability and complexity of the construction phase.

[0006] Second, many bias attribution processes rely on traversing the global knowledge graph or calling multimodal feature fusion and causal reasoning subnets, resulting in high computational consumption, slow response speed, and heavy dependence on cloud computing and bandwidth. Edge application configurations are difficult to meet, hindering the implementation of real-time intelligent decision-making on-site.

[0007] Third, existing methods have semantic coverage blind spots when identifying the causes of schedule deviations in dynamic construction environments. Unforeseen environmental disturbances can easily cause knowledge graph nodes to become "inactive," making it difficult to automatically attribute some key causes. This necessitates manual intervention based on the experience of on-site personnel, thus limiting the system's intelligence level.

[0008] Fourth, multi-source heterogeneous data (such as BIM logs, equipment sensors, meteorological sequences, text reports, etc.) lack efficient and unified semantic mapping and dynamic association mechanisms. Existing unified feature vectors or multimodal fusion engines improve integration on the one hand, but introduce complexity and computational bottlenecks; on the other hand, related technologies are often incompatible with high-frequency changes and asynchronous workflows on construction sites, leading to increased model failure and misjudgment risks.

[0009] Therefore, there is an urgent need for a semantic modeling method that can dynamically perceive changes in the construction phase, achieve lightweight and time-adaptive relationships among multi-source elements. Specifically, it should possess the following capabilities: autonomously perceive disturbances in the on-site construction environment without increasing the cost of fine-tuning large models, relying on multimodal fusion, constructing complex coupling factor networks, or consuming high-performance cloud resources; and accurately model and attribute the dynamic relationships of multi-dimensional elements such as people, machines, materials, and environment. Simultaneously, semantic anchor points should be able to flexibly migrate with changes in construction phases and on-site constraints, improving the intelligence, robustness, and adaptability of schedule deviation cause identification. This not only solves the limitations of traditional static models in dealing with dynamic and complex environments but also provides an efficient and low-cost support foundation for the implementation of digital twins and smart construction in substation engineering, meeting the industry's urgent need for a highly reliable, transferable, and edge-friendly semantic modeling and intelligent attribution system. Summary of the Invention

[0010] This application provides a method for monitoring and analyzing deviations in the construction progress of substation projects based on digital twins, aiming to solve one of the problems or issues of the existing technology mentioned in the background section.

[0011] The method for monitoring and analyzing construction progress deviations in substation projects based on digital twins provided in this application specifically includes: S1: Extract structured semantic units from the preset engineering document source and construct an initial semantic anchor point pool containing basic semantic tags, spatiotemporal constraint templates and related element weight vectors; S2: Based on the initial semantic anchor pool, collect multi-source heterogeneous data and map it to the corresponding spatiotemporal constraint template to generate a temporal state sequence; S3: Perform continuity verification on the time-series state sequence. If it is determined that the spatiotemporal constraint state of the basic semantic tag continues to deviate from the spatiotemporal constraint template bound to the basic semantic tag, then trigger the semantic drift determination process and generate a drift anchor point identifier to be processed. S4: Based on the identifier of the drift anchor point to be processed, determine the alternative anchor point in the process logic chain of the current construction stage, and dynamically adjust the weight vector of the associated elements bound to the drift anchor point to be processed in combination with the historical execution stability index of the alternative anchor point to generate dynamic semantic anchor points. S5: Extract the set of dynamic semantic anchor points at the time when the construction progress deviation of the substation project occurs, calculate the semantic fit deviation value between each dynamic semantic anchor point and the actual observed value of its bound element, and convert it into a semantic distance mapping result. S6: Based on the semantic distance mapping results and the process sequence relationship between each dynamic semantic anchor point, construct a semantic adaptation chain break map; S7: If an anchor segment that meets the preset severity condition is detected in the semantic adaptation chain break map, the local cached construction anomaly pattern library is called to perform matching reasoning and generate an explanation of the deviation cause chain. S8: Write the newly generated dynamic semantic anchors and their drift paths into the incremental anchor evolution log, and record the dynamic semantic anchor identifier, drift triggering conditions, replacement anchor identifier, weight adjustment direction vector and process context snapshot to complete the closed-loop update of semantic modeling.

[0012] The method for monitoring and analyzing deviations in the construction progress of substation projects based on digital twins provided in this application has the following beneficial effects: (1) By constructing a semantic anchor pool with three attributes and introducing a restricted multi-hop semantic propagation mechanism, the problem of static knowledge representation and rigid rules in the traditional substation construction management system is effectively overcome, and dynamic adaptation and local evolution of construction process logic are realized. Compared with the existing system's reliance on preset fixed processes or global knowledge graph reconstruction, this solution, without rebuilding the overall semantic structure, only triggers semantic drift judgment based on continuous violations of spatiotemporal constraint templates, and achieves adaptive adjustment of element weight vectors through the alternative anchor with the highest semantic compatibility in adjacent processes, which significantly improves the system's response sensitivity and robustness in complex and changeable field environments. This mechanism avoids the high computational overhead and model coupling problem caused by end-to-end feature fusion, making the semantic update process lightweight, traceable, and without the need for external large model support. It is especially suitable for engineering scenarios with limited edge resources, greatly reducing the dependence on cloud computing power and network bandwidth, and improving the flexibility and practicality of system deployment.

[0013] (2) A fracture map generation method based on the joint modeling of semantic fit deviation level and process sequence relationship is proposed, which breaks through the technical bottleneck of traditional schedule deviation analysis relying on numerical residual statistics or full map traversal causal reasoning, and realizes efficient identification and interpretable tracing of construction anomaly propagation path. Unlike conventional methods that require running complex causal search algorithms or calling large-scale external knowledge bases for attribution analysis, this scheme uses active anchor point set as slice unit, and uses preset discretized semantic distance mapping rules to transform the deviation between actual observation value and bound elements into readable levels such as "complete mismatch" and "partial offset". Combined with the process sequence relationship, the semantic chain fracture segment is quickly located, and for segments with ≥3 consecutive "complete mismatch" anchor points, the local cached anomaly pattern library is activated to perform lightweight pattern matching and output the causal chain explanation. The whole process does not require real-time access to external safety specification library or cross-modal data alignment, which not only ensures the engineering interpretability of the reasoning results, but also significantly shortens the anomaly diagnosis delay, and solves the problems of high diagnosis delay, fuzzy attribution and high resource consumption that are common in existing schemes.

[0014] (3) An incremental anchor evolution log mechanism is designed to realize the closed-loop accumulation of construction knowledge and cross-project strategy migration, effectively alleviating the long-standing cold start problem and experience reuse difficulties in the field of engineering construction. By automatically recording the drift path, triggering conditions, substitution relationships and context snapshots of newly generated anchors every day, the system completes knowledge accumulation without storing original sensitive data, forming a transferable lightweight strategy asset; subsequent projects can load the historical anchor evolution trajectory based on similar process contexts, quickly initialize the semantic anchor pool and pre-set potential drift directions, thereby improving the adaptation speed and decision-making quality of new projects. This mechanism constructs a decentralized, loosely coupled, and self-evolving construction cognition architecture, getting rid of the dependence on LLM inference engine and centralized data center, and realizing the continuous evolution and horizontal reuse of construction wisdom under the premise of ensuring data privacy and system independence.

[0015] In summary, this solution constructs a digital twin construction management system with adaptive evolution capabilities, an efficient anomaly tracing mechanism, and sustainable knowledge accumulation characteristics through a lightweight dynamic modeling framework driven by semantic anchors. This not only significantly improves the system's stability and response accuracy under dynamic disturbances, but also fundamentally optimizes resource utilization efficiency and engineering decision-making transparency. It provides a high-performance, low-overhead, and highly interpretable intelligent solution for highly complex and constrained construction scenarios such as substation engineering. Attached Figure Description

[0016] Figure 1 This is the main flowchart of a method for monitoring and analyzing deviations in the construction progress of substation projects based on digital twins; Figure 2This is a sub-flowchart of a method for monitoring and analyzing construction progress deviations in substation engineering based on digital twins; Figure 3 This is another sub-flowchart of the substation engineering construction progress deviation monitoring and analysis method based on digital twins. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below, examples of which 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. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0019] like Figure 1 As shown, this application provides a method for monitoring and analyzing deviations in the construction progress of substation projects based on digital twins, specifically including: S1: Extract structured semantic units from substation engineering construction standards and specifications, typical work instructions and historical completion reports, and construct an initial semantic anchor point pool containing basic semantic tags, spatiotemporal constraint templates and related element weight vectors, as the benchmark data source for subsequent drift detection.

[0020] S2: Based on the initial semantic anchor pool, collect on-site BIM model update logs, process inspection data, meteorological measurement sequences and mobile terminal inspection texts at the daily granularity, and map the collected multi-source heterogeneous data to the corresponding spatiotemporal constraint templates to generate time-series state sequences.

[0021] S3: Perform continuity verification on the time-series state sequence. If it is determined that the spatiotemporal constraint template bound to a certain basic semantic tag is not satisfied twice in a row, the semantic drift determination process is triggered and a drift anchor point identifier to be processed is generated.

[0022] S4: Based on the drift anchor point identifier to be processed, retrieve the alternative anchor point with the highest semantic compatibility in the process logic chain of the current construction stage, and dynamically adjust the weight vector of the original associated element in combination with the historical execution stability index of the alternative anchor point to generate a dynamic semantic anchor point with time-adaptive properties.

[0023] S5: When a schedule deviation occurs, extract the set of active dynamic semantic anchors at the time of the deviation, calculate the semantic fit deviation between each dynamic semantic anchor and the actual observed value of its bound element, and convert the numerical residual into a discrete level semantic distance mapping result.

[0024] S6: Based on the semantic distance mapping results and the process sequence relationship between each dynamic semantic anchor point, construct a semantic adaptation chain break map to represent the semantic relationship break state, so as to locate anchor point segments that continuously appear at the level of complete mismatch.

[0025] S7: For anchor segments with three or more consecutive complete mismatch levels in the semantic adaptation chain break map, call the locally cached construction anomaly pattern library for matching reasoning to generate a deviation cause chain explanation containing typical deviation patterns.

[0026] S8: Write the newly generated dynamic semantic anchors and their drift paths into the incremental anchor evolution log, and record the dynamic semantic anchor identifier, drift triggering conditions, replacement anchor identifier, weight adjustment direction vector and process context snapshot to complete the closed-loop update of semantic modeling.

[0027] Step S1: Extract structured semantic units from substation engineering construction standards and specifications, typical work instructions, and historical completion reports to construct an initial semantic anchor pool containing basic semantic tags, spatiotemporal constraint templates, and related element weight vectors. This pool serves as the benchmark data source for subsequent drift detection. Specifically, this includes: S1.1: Multi-source heterogeneous data acquisition and processing are performed on substation engineering construction standard and specification texts, typical work instruction documents and historical completion report archives. Entity recognition algorithms in the field of natural language processing are used to extract key construction process terms, environmental constraint values ​​and resource dependency descriptions to generate a standardized set of original semantic triples.

[0028] Multi-source heterogeneous data acquisition and processing are performed on substation engineering construction standard specifications, typical work instructions, and historical completion report archives. A paragraph segmentation and metadata extraction component based on a document parsing engine is invoked to parse the document content from different sources into text and table data structures that can be further analyzed, based on chapter indexes and time tags. The parsed text block data structures are input into a named entity recognition algorithm model in the field of natural language processing. Construction-specific entity categories are defined, including "construction procedure terms," ​​"environmental constraint values," and "resource dependency descriptions." A terminology table for substation engineering is loaded into the model dictionary to achieve accurate annotation of entities in the documents that conform to the domain definition. Context-window-based dependency parsing is performed on the annotated construction procedure terms to extract the semantic paths between the procedure terms and their corresponding environmental constraints and resource dependencies, forming a chain-like association information of procedure → constraint → resource. For environmental constraint entities, a numerical parsing module is invoked to convert the raw text values ​​containing units and dimensions into standard numerical formats and add unit identifiers, ensuring that the constraints can be directly used in subsequent calculations. Semantic role labeling is performed on resource dependency descriptions to identify resource categories (equipment, manpower, materials, etc.), resource characteristics (model, qualification, batch, etc.), and resource status (available, under maintenance, delayed, etc.), and these are transformed into structured attribute fields. The aforementioned chain-like association information is then encapsulated into a set of raw semantic triples (<process label, environmental constraint value, resource dependency description>) according to triple format rules. A unique sequence index is assigned to facilitate subsequent mapping and weight calculation. Through structured encapsulation, the data from the previous step is transformed into a standardized set of raw semantic triples, enabling rapid extraction of key construction information from multi-source heterogeneous documents and providing a solid data foundation for subsequent semantic normalization mapping.

[0029] S1.2: Based on the original set of semantic triples, perform semantic normalization mapping processing, use domain ontology alignment technology to map unstructured process terms into unified basic semantic tags, and transform environmental constraint values ​​into computable spatiotemporal constraint template parameters to output a standard semantic unit sequence with a unique identifier.

[0030] Based on the original set of semantic triples generated in step S1.1, the preset semantic ontology mapping library for the substation engineering field is called, and semantic normalization mapping calculation is performed using the process noun field as the search key, outputting a unique basic semantic tag number corresponding to the ontology node.

[0031] For the numerical fields of environmental constraints, the constraint type definition table in the domain ontology is referenced. The non-SI numerical values ​​are converted into SI standard units through unit conversion rules. Then, the parameterized template construction function is called based on the constraint type to fill the numerical values ​​into the template variable slots to form computable spatiotemporal constraint template parameters.

[0032] After word segmentation and dependency parsing, the resource dependency description field is matched with the set of related elements defined in the domain ontology to generate an index of related elements that is bound to the basic semantic tags and spatiotemporal constraint templates, ensuring that the three types of attributes have a unified identification system.

[0033] The basic semantic label number, instantiation spatiotemporal constraint template parameters, and associated element index are integrated into a standard semantic unit, and a unique identifier is generated by hashing to ensure the stability and traceability of the identifier during subsequent serialization and transmission.

[0034] By using domain ontology alignment technology, the results of the previous step are transformed into a sequence of standard semantic units with unique identifiers, thereby achieving cross-source consistent modeling and computable feature generation of construction semantic entities.

[0035] S1.3: Perform association strength quantification calculation on each standard semantic unit in the standard semantic unit sequence, and use statistical co-occurrence frequency analysis to mine the coupling frequency of each basic semantic tag in the historical completion report with related elements such as the status of lifting machinery, the qualifications of high-altitude workers and wind speed threshold, so as to generate a weight vector of related elements representing the original coupling strength.

[0036] When performing association strength quantification calculation on a standard semantic unit sequence with a unique identifier, the input conditions are the basic semantic labels that have completed semantic normalization mapping and their bound spatiotemporal constraint templates, combined with the records of related elements such as equipment status, high-altitude worker qualifications, and wind speed thresholds that have undergone entity recognition processing in the historical completion report text.

[0037] The associated element parsing module is invoked to split each construction record in the historical completion report into timestamp pairs of the occurrence events of basic semantic tags and the occurrence events of associated elements. The event timestamp alignment algorithm is used to establish temporal co-occurrence pairs of tags and associated elements.

[0038] Based on the statistical co-occurrence frequency analysis method, frequency counting processing is performed to accumulate the number of co-occurrences of each basic semantic tag and each associated element in historical data, and calculate their respective proportions in the total number of records.

[0039] By employing normalization operations, co-occurrence counts of different orders of magnitude are converted into intensity values ​​of a uniform scale, generating the original coupling strength matrix. In this matrix, rows represent basic semantic labels, columns represent associated elements, and element values ​​are normalized co-occurrence intensities.

[0040] Based on the coupling strength matrix, a correlation coefficient adjustment mechanism is applied to correct the coefficients of the correlation elements that occur frequently and have low stability in a certain stage, so as to avoid the deviation caused by extreme frequencies.

[0041] Using a weighted aggregation method, the corrected intensity values ​​are combined into a related element weight vector according to preset weights. This weight vector represents the original coupling strength between the basic semantic tags and each related element.

[0042] By using correlation strength quantification calculation and weight vector construction, the standard semantic unit sequence from the previous step is transformed into a correlation element weight vector with computable coupling characteristics, thereby achieving an accurate characterization of semantic anchor point correlation and providing quantitative input for subsequent 3D attribute encapsulation.

[0043] S1.4: Based on the basic semantic tags, spatiotemporal constraint template parameters and associated element weight vectors, perform three-dimensional attribute encapsulation processing, and use an object-oriented data structure construction method to bind the above three types of attributes to the same semantic entity index to form a candidate semantic anchor instance with complete attribute definition.

[0044] S1.5: Perform consistency verification and deduplication optimization on the candidate semantic anchor instance set, use hash fingerprint comparison algorithm to remove redundant and duplicate candidate semantic anchor instances and correct attribute conflicts, so as to finally generate an initial semantic anchor pool for subsequent drift detection benchmark.

[0045] Step S2: Based on the initial semantic anchor pool, collect on-site BIM model update logs, process inspection data, meteorological measurement sequences, and mobile terminal inspection texts at a daily granularity, and map the collected multi-source heterogeneous data to the corresponding spatiotemporal constraint templates to generate a time-series state sequence. Specifically, this includes: S2.1: Acquire multi-source heterogeneous raw data such as daily granular BIM model update logs, process inspection data, meteorological measurement sequences, and mobile terminal inspection texts. Use a rule-based regular expression extraction algorithm to perform field cleaning and format standardization on the multi-source heterogeneous raw data to generate a standardized process event set with a unified timestamp index.

[0046] The input conditions for acquiring multi-source heterogeneous raw data, including daily granular BIM model update logs, process inspection data, meteorological measurement sequences, and mobile terminal inspection texts, are defined as follows: BIM model update logs contain structured component status change records; process inspection data contains semi-structured process completion status and acceptance results; meteorological measurement sequences are derived from real-time numerical data such as wind speed, temperature, and humidity collected by on-site sensors; and mobile terminal inspection texts are natural language records entered by operators. A unified time baseline is constructed for the multi-source heterogeneous raw data. The system time synchronization module is called to bind timestamps accurate to the minute to all data records, and offset adjustments are made according to data acquisition priority to eliminate multi-source contradictions when timestamp conflicts occur. A rule-based regular expression extraction algorithm is used to perform matching filtering on fields containing symbols and special characters in the BIM logs. The rule pattern can be defined as a combination expression of engineering component numbers and status keywords. Component IDs and status values ​​are extracted through group capture to form structured fields. When performing field cleaning on process inspection data, a pre-built keyword library for inspection items is used to scan the descriptive fields in segments. Successfully matched segments are mapped to the corresponding inspection item codes, while unmatched segments are marked as fields for manual review. Numerical type validation and outlier removal are performed on meteorological measurement sequences. A threshold detection algorithm is used to remove measurements exceeding reasonable physical ranges, and missing values ​​are restored through linear interpolation to ensure the continuity of meteorological data. When standardizing the format of mobile terminal inspection text, regular expressions are used to match entities such as dates, equipment numbers, and work statuses. Natural language descriptions are converted into uniformly encoded enumerated values, and noise removal is performed on the text content to eliminate redundant symbols and punctuation. Through this chained processing, data from different sources and in different formats are all transformed into a unified field structure and bound to a unified timestamp index. This outputs a standardized set of process events with standard field definitions and time uniformity, achieving unbiased alignment of multi-source heterogeneous data on the same timeline.

[0047] S2.2: Based on the standardized process event set, call the pre-built initial semantic anchor pool, and use the semantic tag matching algorithm to associate each event in the standardized process event set with the corresponding basic semantic tag to generate an event stream to be mapped with basic semantic tag annotations.

[0048] Based on a standardized set of process events with a unified timestamp index, all basic semantic tags and their associated index structures in the pre-built initial semantic anchor pool are loaded, and a mapping retrieval interface from event fields to semantic tags is established to realize the event-level semantic anchor call function.

[0049] For each event record in the standardized process event set, read the process keyword vector in the event field, call the semantic tag matching algorithm to calculate the semantic similarity score between the vector and each basic semantic tag in the initial semantic anchor pool, and generate a similarity score list between the event and the tag.

[0050] Based on the similarity score list, the maximum value selection logic is executed to select the basic semantic tag with the highest score that exceeds the preset matching threshold as the only matching object of the event. The matching result is then bound to the event identifier to generate an labeled structured event unit.

[0051] For events where multiple scores simultaneously exceed a preset matching threshold, the domain priority rule base is invoked to prioritize candidate tags, and the final matching tag is calculated using weighted averages. This avoids misassignment of tags due to ambiguity and ensures domain consistency.

[0052] All labeled structured event units are arranged in the original timestamp index order and encapsulated to form a stream of events to be mapped with basic semantic labels, providing a complete semantic labeling foundation for subsequent spatiotemporal constraint template mapping.

[0053] By combining semantic tag matching algorithm and optimization logic, the standardized process event set of the previous step is transformed into a stream of events to be mapped that are arranged in the order of events and have domain-consistent annotations, thus achieving an accurate mapping effect from the original events to the semantic tag annotations.

[0054] S2.3: For the event stream to be mapped with basic semantic tags, extract the spatiotemporal constraint template bound to each basic semantic tag, and use numerical normalization and logical Boolean operation methods to map the meteorological measurement sequence and process duration data to the variable slots of the spatiotemporal constraint template to generate an instantiated constraint condition set containing specific parameter values.

[0055] S2.4: Based on the instantiated constraint set, execute the constraint satisfaction determination logic, compare the actual observed value with the threshold range of the spatiotemporal constraint template, and generate a binary state label sequence representing whether each basic semantic label meets the constraint requirements on the day.

[0056] Based on each constraint parameter group in the instantiated constraint set, the actual observation data corresponding to the basic semantic labels are read, including numerical records such as meteorological measured variables and construction procedure execution time, and the synchronization of the comparison is ensured by precise alignment through timestamps.

[0057] During the comparison process, the threshold range determination module is invoked to calculate the numerical difference between the actual observed value and the upper and lower limits of the bound spatiotemporal constraint template. The difference calculation result serves as the direct input parameter for constraint satisfaction determination.

[0058] Boolean logic operations are used to map each difference calculation result to a binary state value of "satisfied" or "not satisfied". Differences within the threshold range are generated as true value labels, and differences outside the threshold range are generated as false value labels.

[0059] For multiple constraint variables in a multidimensional constraint template, a global decision rule is executed. The binary state values ​​of each single variable are combined into an overall state label using a logical AND operation, indicating whether the basic semantic label satisfies the constraint conditions as a whole in the current time segment.

[0060] All basic semantic labels are arranged in label index order and encapsulated into a binary state label sequence, which serves as the standardized input dataset for subsequent continuity verification.

[0061] By using constraint satisfaction determination and state label generation processing, the instantiated constraint condition set of the previous step is transformed into a binary state label sequence representing the semantic adaptation status of daily construction, thereby realizing accurate semantic encoding of the on-site execution status and providing a directly usable judgment benchmark for subsequent drift detection.

[0062] S2.5: The binary state marker sequence is sequentially spliced ​​and serialized along the time axis, and the discrete binary state marker sequence is integrated into continuous time segments using a sliding window aggregation mechanism to generate the final time-series state sequence for continuity verification.

[0063] like Figure 2 As shown, step S3: Perform continuity verification on the temporal state sequence. If it is determined that the spatiotemporal constraint template bound to a certain basic semantic tag is not satisfied twice consecutively, then trigger the semantic drift judgment process and generate a drift anchor point identifier to be processed. Specifically, this includes: S3.1: Obtain the basic semantic labels arranged by daily granularity and their corresponding spatiotemporal constraint template execution state records in the time-series state sequence. Use the sliding time window algorithm to extract and align the execution state records of two consecutive time steps to generate a dual-time-period state comparison vector containing the current time and the previous time-series time.

[0064] Before executing this sub-step on the time-series state sequence, the input conditions are basic semantic labels arranged by daily granularity and their corresponding spatiotemporal constraint template execution state records, derived from a continuous time segment sequence generated by S2.5, with each time segment containing a binary state label and a unified timestamp index. The input sequence is loaded into a memory buffer in ascending order of timestamp. A sliding time window algorithm manager is called to instantiate a window object, setting the window span parameter to two time steps and binding the time axis index to form a window traversal controller. The window traversal controller is applied to the state sequence, synchronously reading the state vectors of the current and previous times during each window sliding process. An index alignment mechanism is used to ensure the consistency of the order of the state records in the basic semantic label dimension between the two time periods, and misaligned records are filled in by interpolation or zero-padding. The aligned current state and the previous state are concatenated into a dual-time period matrix. A column vector mapping mechanism is used to divide the matrix into multi-dimensional comparison units according to the basic semantic labels, and the corresponding spatiotemporal constraint template execution state value is retained in each comparison unit. All multi-dimensional comparison units are merged into a unified dual-time period state comparison vector set, and a verification module is called to check the correctness of the vector encoding format and template binding relationship. This chain-like derivation process converts the binary state markers of continuous time segments from the previous step into dual-time-segment state comparison vectors, thus enabling the continuous judgment input required for subsequent drift trigger signal detection.

[0065] For example, in a substation construction management system, the time-series state sequence contains 10 basic semantic tags. Each tag is bound to a spatiotemporal constraint template execution status record as a binary value, with a time span of 7 days. The window span is set to 2 time steps, traversing the paired records of day 3 and day 2, day 4 and day 3, etc., of the time-series state sequence. For the pairing of day 3 and day 2, the basic semantic tag "hoisting operation" is recorded as an unmet state value of 0 for both days, while other tags such as "prefabricated module placement" and "casing test" show state differences of 1 and 0 respectively over the two days. After index alignment, this pairing is formed into a dual-time-period matrix containing 10 tags. The first column of the matrix represents the state value of day 2, and the second column represents the state value of day 3. The matrix is ​​then divided into 10 comparison units according to the tag dimension. The 10 comparison units are merged into a dual-time-period state comparison vector set, and data structure verification is performed. Finally, a structured dual-time-period state comparison vector dataset is output, which enables the subsequent Boolean logic NOT operation in S3.2 to accurately evaluate the continuous failure situation and significantly improve the stability and accuracy of drift determination.

[0066] S3.2: Based on the dual-time-period state comparison vector, the Boolean logic NOT operation rule is used to perform reverse discrimination processing on the spatiotemporal constraint template satisfaction status within each time step, so as to generate a discretized failure flag bit sequence representing a single constraint failure.

[0067] Based on the dual-time-phase state comparison vector generated in step S3.1, the input condition is the set of binary state label values ​​for each basic semantic label at the current and previous times, along with the corresponding spatiotemporal constraint template execution state records. The state values ​​at each time step in the dual-time-phase state comparison vector are processed element-wise using Boolean NOT operations to extract the inverse attributes of the satisfaction determination results, forming failure candidate labels that are logically opposite to the original state values. A threshold filtering operation is performed on the inverse attribute values ​​at each time step, retaining only labels where the original satisfaction determination value is true, the inverse attribute value is also true, and the corresponding spatiotemporal constraint template has a measured value exceeding the threshold range at that time, thus eliminating false failures caused by missing observation data. The filtered inverse labels are bound to the basic semantic label index to construct a discretized failure flag sequence, where each flag position in the sequence corresponds to a unique combination of a time step and a specific semantic label. A batch assembly mechanism based on vectorized indexes is used to output the failure flag sequence as a fixed-length structured array, serving as the direct input for continuous pattern matching detection. By using Boolean logic reverse discrimination and threshold filtering, the dual-time state comparison vector result from the previous step is transformed into discretized failure flag sequence data, thereby achieving accurate location of single constraint failure.

[0068] For example, in a substation construction monitoring scenario, the input conditions include the state value records of the basic semantic tag "lifting operation" over two consecutive days: the state value on the first day is 1 (constraint satisfied), and the state value on the second day is 0 (constraint not satisfied); the corresponding spatiotemporal constraint template parameters are a wind speed threshold of ≤5m / s, and the measured wind speeds are 4m / s and 8m / s respectively. A Boolean NOT operation is performed on the dual-time-period state comparison vector, resulting in a reverse value of 0 for the first day and 1 for the second day. Threshold filtering is applied, discarding the reverse value for the first day and retaining the reverse value for the second day, as its corresponding measured wind speed of 8m / s exceeds the threshold range. The retained reverse flags are bound to the "lifting operation" tag index, forming a failure flag sequence [0,1]. A fixed-length array encapsulation structure is used to output the failure flag sequence for continuous mode detection. In this embodiment, by processing the chain to filter out non-failure states, the single failure event on the second day is accurately identified, significantly improving the robustness of subsequent continuous violation detection.

[0069] S3.3: Perform continuous pattern matching detection on the discretized failure flag sequence, and use the sequence consistency check algorithm to identify whether there are continuous failure segments where the failure flag bits of two adjacent time steps are true at the same time, so as to generate a potential drift trigger signal representing continuous violation events.

[0070] During the continuous pattern matching detection process on the discretized failure flag sequence, the input object is a binary state sequence formed after Boolean logic inverse discrimination. This sequence is arranged according to time steps, and the value range of each flag bit is limited to true or false, with a true value representing a single constraint failure. Based on the state sequence indexed by the time axis, a sliding detection window for pattern comparison is constructed, with a window length set to two time steps to ensure coverage of state pairs of adjacent nodes. For each window position, a sequence consistency check algorithm is called to sequentially read the values ​​of the two flag bits within the window and compare their matching patterns one by one with a predefined set of continuous failure modes. The matching mode set contains patterns such as... A pattern is defined as consecutive true pairs, indicating that both observations have failed. During the detection process, when both flag bits within a window meet the expected true value conditions, a continuous failure segment index is immediately generated in the corresponding time period, and the potential drift trigger signal is recorded using this index as the key. A pattern matching statistical counting method is used to perform a full traversal cumulative detection of the entire flag bit sequence, forming a set of trigger signals containing all consecutive violation events. By coupling the consistency verification logic with the trigger signal generation mechanism, failure events at adjacent time steps can be quickly captured, providing accurate event location data for subsequent drift confidence assessment.

[0071] S3.4: Based on the potential drift trigger signal, invoke the predefined drift confidence assessment mechanism to perform weighted fusion calculation on the historical stability weights of the basic semantic labels, so as to generate a valid drift determination result after confidence filtering.

[0072] The system obtains the potential drift trigger signal sequence and the corresponding historical stability weight index table of basic semantic labels generated based on S3.3, and calls the parameter configuration file of the predefined drift confidence assessment mechanism to determine the weight fusion coefficient matrix required for calculation. The potential drift trigger signals and the historical stability weight index table are vectorized to align the time dimension of the trigger signals with the historical distribution dimension of the stability weights in a unified data space, forming a multidimensional state tensor for computation. Based on the multidimensional state tensor, a weighted fusion calculation is performed. A linear combination model is used to perform element-wise operations on the historical stability weights and the current drift trigger signal strength according to the set fusion coefficients, as shown in the following formula: Where C is the drift confidence value after fusion. The historical stability weight of the i-th basic semantic label. The signal strength of the tag in the current trigger signal sequence is multiplied and then summed by index to obtain the total confidence score. The obtained confidence score is then compared with a predefined confidence threshold table for interval determination. Based on the threshold distribution rules, a set of drift event indices with confidence scores higher than the minimum effective determination threshold is selected, thus achieving the filtering process from all trigger signals to effective drift determination. The filtered drift event index set is bound to the corresponding basic semantic tag identifier, generating a valid drift determination result verified by confidence. Through a drift confidence evaluation mechanism and weighted fusion calculation method, the potential drift trigger signals in step S3.3 are transformed into valid drift determination results after statistical stability correction, improving the accuracy and robustness of drift identification.

[0073] For example, during the main transformer installation phase of a substation, the potential drift triggering signal sequence contains four basic semantic labels: hoisting operation, bushing installation, oil handling, and high-voltage test. The corresponding weight values ​​in the historical stability weight index table are 0.82, 0.75, 0.68, and 0.91, respectively, and the current triggering signal strength vectors are 0.65, 0.80, 0.60, and 0.50, respectively. The drift confidence assessment mechanism configuration file is called, and the fusion coefficient is set to 1 (equal weight fusion). The data is aligned, and the product of the first label is calculated using the formula: 0.82 × 0.65 = 0.533; the second is 0.75 × 0.80 = 0.600; the third is 0.68 × 0.60 = 0.408; and the fourth is 0.91 × 0.50 = 0.455. The summation yields the confidence level C = 0.533 + 0.600 + 0.408 + 0.455 = 1.996. The confidence score is compared with the minimum valid judgment threshold of 1.50 in the threshold table. Drift events that meet the condition are all four labels. The final output of valid drift judgment results is the set containing the above four labels. The application results show that this method can still accurately identify other operation drifts affected by process dependence when the hoisting operation is canceled due to the continuous excessive wind speed on site, which significantly improves the accuracy and timeliness of deviation cause identification.

[0074] S3.5: Based on the valid drift determination result, perform a unique identifier allocation operation to map the corresponding basic semantic label to a specific drift anchor point identifier to be processed, so as to complete the conversion from continuous state verification to specific drift object positioning and output the final drift anchor point identifier to be processed.

[0075] like Figure 3 As shown, step S4: Based on the identifier of the drift anchor point to be processed, retrieve the alternative anchor point with the highest semantic compatibility in the process logic chain of the current construction stage, and dynamically adjust the weight vector of the original associated elements in combination with the historical execution stability index of the alternative anchor point to generate a dynamic semantic anchor point with time-adaptive properties. Specifically, this includes: S4.1: Obtain the identifier of the drift anchor point to be processed and the process logic chain data of the current construction stage. Use the graph traversal algorithm to perform semantic compatibility matching calculation on the adjacent process nodes in the process logic chain, and generate a semantic compatibility mapping table containing a set of candidate alternative anchor points and corresponding compatibility score values.

[0076] The process logic chain data for the current construction phase is a structured data set describing the steps of each process within this phase and the logical connections between them. It clearly records which processes must be executed before or after other processes, as well as the direct dependencies between processes. When it is necessary to find alternatives for a drifting anchor point, the system uses the connections defined in this data to search and match only among neighboring process nodes that have a direct dependency on the current anchor point, ensuring that all candidate alternatives found are logically feasible in terms of construction process.

[0077] The process acquires the drift anchor point identifier and the process logic chain data of the current construction stage. A graph traversal algorithm based on node relationships is used to perform structured parsing of the process logic chain, extracting a set of neighboring process nodes that are adjacent to the drift anchor point identifier and have direct process dependencies. For each node in the set of neighboring process nodes, the bound basic semantic tags and associated element descriptions are read. Tag vectorization processing is performed based on a preset domain ontology semantic matrix, transforming unstructured semantic information into a computable semantic feature vector. The cosine similarity calculation method is used to calculate the similarity between the drift anchor point feature vector and the feature vector of each neighboring node, and a compatibility score is generated using the following formula: Where S is the compatibility score, and V is the compatibility rating. p V is the feature vector of the drift anchor point. c Let k be the feature vector of the candidate node, k be the feature dimension index, and d be the dimension of the semantic feature vector. For the compatibility score, a normalization module is called to map it to the [0,1] interval for unified measurement in subsequent multi-objective evaluation. After the normalized score is generated, all candidate alternative anchors and their corresponding score values ​​are encapsulated into a semantic compatibility mapping table and stored in temporary memory for use in step S4.2. Through graph traversal and vectorized matching calculation, the drift anchor identifiers from the previous step are transformed into a set of candidate alternative anchors and a standardized compatibility score that can be used for stability evaluation, achieving accurate retrieval and preliminary screening of semantic alternative anchors.

[0078] For example, during the construction phase of a 220kV substation project, the drifting anchor point to be processed is identified as "lifting operation." The process logic chain contains six adjacent process nodes that directly depend on it, namely "hydraulic jacking," "prefabricated cabin placement," and "steel structure welding." Using the domain ontology semantic matrix, the feature vector of "lifting operation" is vectorized into a 12-dimensional feature vector, for example, the value corresponding to dimension 3 is 0.85, and the value corresponding to dimension 7 is 0.64; the feature vector of "hydraulic jacking" is vectorized into 12 dimensions, with a value of 0.72 in dimension 3 and 0.60 in dimension 7. Substituting into the compatibility scoring formula, the compatibility score between this node and the drifting anchor point is calculated to be 0.918, which becomes 0.91 after normalization. The calculation is repeated for all candidate nodes, finally obtaining a set of candidate alternative anchor points with a compatibility score range of 0.68 to 0.94, and output in the form of a semantic compatibility mapping table. In subsequent steps, the system uses this mapping table to screen alternative anchor points with excellent stability indicators, enabling time-adaptive adjustment of anchor points under construction disturbances and significantly improving the accuracy of deviation cause identification.

[0079] S4.2: Based on the set of candidate alternative anchor points in the semantically compatible mapping table, the historical execution records of each candidate alternative anchor point are extracted by calling the locally cached historical completion report database. The execution volatility index of each candidate alternative anchor point is calculated by using statistical variance analysis, and a historical execution stability quantitative sequence representing the degree of construction stability is generated.

[0080] The historical completion report database is a locally cached structured data source used to provide historical process execution data associated with each candidate alternative anchor point. The system extracts historical execution records covering multiple construction stages through a field indexing mechanism. After time series reconstruction and standardization, statistical variance analysis is used to calculate the volatility index of its execution status, thus providing a data foundation for generating a historical execution stability quantification series that quantifies the degree of construction stability.

[0081] Based on the candidate alternative anchor set in the semantically compatible mapping table, the historical execution records of each candidate alternative anchor are extracted from the locally cached historical completion report database. A field indexing mechanism is used to locate the historical process execution data corresponding to the basic semantic tags of the candidate anchors, ensuring that the extracted data covers multiple construction stages in the time dimension to reflect long-term trend characteristics. The extracted historical execution records undergo time series reconstruction processing, employing a unified time step and event state encoding method to standardize execution states from different sources and periods into a highly comparable sequence format, eliminating the impact of data source differences on subsequent analysis. For the reconstructed time series data, statistical variance analysis is used to quantify the fluctuation degree of the execution state of each candidate alternative anchor, and the variance index is used to characterize construction stability by representing the magnitude of event state changes within the construction cycle. The variance calculation formula is used. in, For the first In the first time step The state values ​​of the candidate alternative anchor points The total number of time steps. This serves as the volatility metric for the anchor point. It is the arithmetic mean of the state value of the j-th candidate anchor point over n time steps.

[0082] The calculated volatility indicators are encapsulated into a historical execution stability quantification sequence according to the anchor point identifiers, and each indicator is appended with a time span and data reliability label to support the input requirements of the subsequent multi-objective weighted evaluation function. Through the above processing method, the candidate alternative anchor point set from the previous step is transformed into a measurable historical execution stability quantification sequence, realizing a structured quantitative characterization of the degree of construction stability.

[0083] For example, based on five candidate alternative anchor points contained in a semantically compatible mapping table, the local cached database of historical substation project completion reports is accessed to extract the daily execution status records for each anchor point over the past 12 months. A status code of 1 indicates normal execution on that day, and 0 indicates execution interruption. The data corresponding to different anchor points are aligned to a unified step-size sequence at a daily granularity, with each sequence having a length of 365. Using the analysis of variance formula, the status values ​​in the sequence are treated as individual samples. First, the mean is calculated, then the sum of squared differences between each status value and the mean is calculated, and finally, the sum of squared differences is divided by n-1 to obtain the volatility index. If the execution status of an anchor point has an annual volatility index below 0.1, it is marked as a high-stability anchor point; if the volatility index is between 0.1 and 0.4, it is marked as medium-stability; and if it is above 0.4, it is marked as low-stability. This ultimately forms a historical execution stability quantification sequence containing five anchor points, each with its volatility value, stability level, and data reliability label. This quantification sequence significantly improves the effectiveness of construction stability assessment in subsequent optimal alternative anchor point selection.

[0084] S4.3: Based on the compatibility score in the semantic compatibility mapping table and the execution volatility index in the historical execution stability quantification sequence, construct a multi-objective weighted evaluation function to perform the optimal solution search operation, select the target alternative anchor point with the highest comprehensive score, and generate the optimal alternative anchor point identifier.

[0085] Based on the compatibility score in the semantic compatibility mapping table and the execution volatility index in the historical execution stability quantification sequence, an input parameter matrix for a multi-objective weighted evaluation function is constructed, with the compatibility score as the first objective weight vector and the execution volatility index as the second objective weight vector.

[0086] The input parameter matrix is ​​matched by row indexing based on the identifier sequence of each candidate alternative anchor point to ensure that the compatibility score and volatility value of the same candidate alternative anchor point correspond in the same row.

[0087] Normalization is performed on each row of data. The compatibility score is mapped to the interval [0,1] using a linear normalization method, and the volatility index is mapped to the interval [0,1] using a reverse normalization method to highlight anchor points with high stability.

[0088] Construct a multi-objective weighted evaluation function and calculate the comprehensive score using the following formula: Where S is the overall score. C represents the compatibility score weight, and C is the normalized compatibility score value. To apply stability weights, V is the normalized volatility value.

[0089] The comprehensive scores of each candidate alternative anchor point in the formula are sorted by matrix, and the anchor point with the highest comprehensive score is identified by using a descending order.

[0090] The anchor point with the highest overall score is output as the optimal alternative anchor point, and a corresponding identifier mapping record is established to provide accurate alternative anchor point positioning for subsequent dynamic weight adjustment.

[0091] By integrating weighted vectors and calculating comprehensive scores, the compatibility score and volatility index generated in the previous step are transformed into unique identifiers for the optimal alternative anchor points, thus achieving optimal decision-making in selecting alternative anchor points.

[0092] For example, in semantic drift detection during the equipment installation phase of a substation, the compatibility score weight is set to 0.6, and the execution stability weight is set to 0.4. The compatibility scores of the five candidate alternative anchor points are 0.85, 0.78, 0.92, 0.80, and 0.75, respectively, and the execution volatility indices are 0.10, 0.15, 0.05, 0.12, and 0.20, respectively. The normalized C matrix values ​​are 0.85, 0.78, 0.92, 0.80, and 0.75, and the V matrix values ​​are 0.10, 0.15, 0.05, 0.12, and 0.20. Substituting into the evaluation function, the comprehensive scores obtained are 0.91, 0.872, 0.952, 0.888, and 0.85, respectively. The ranking results show that the candidate alternative anchor point with the highest comprehensive score is the third anchor point, and its identifier is the optimal alternative anchor point identifier. After being applied to subsequent dynamic weight adjustments, the accuracy of identifying the causes of construction progress deviations is significantly improved, the response time is shortened, and the system maintains high robustness under limited computing power.

[0093] S4.4: Based on the historical execution stability quantification sequence corresponding to the optimal alternative anchor point identifier, an adaptive gradient update strategy is used to iteratively correct and calculate the weight vector of the associated elements bound to the original drift anchor point identifier to generate a dynamic weight adjustment vector that reflects the latest construction disturbance characteristics.

[0094] Based on the historical execution stability quantization sequence corresponding to the selected optimal alternative anchor point identifier, the weight vector of the associated elements bound to the original drift anchor point identifier is loaded as the initial state input. The stability quantization sequence is aligned to the position of each element in the associated element weight vector according to the time index, forming a matching time-driven update matrix. The adaptive gradient update strategy module is called to calculate the element-level difference between the stability value in the update matrix and the original weight vector, obtaining the gradient direction vector reflecting the degree of execution fluctuation of each associated element. For the positive component in the gradient direction vector, the weight is incremented according to the stability improvement, and the increment step size is set to be proportional to the stability value; for the negative component in the gradient direction vector, the weight is decremented according to the stability decrease, and the decrement step size is set to be inversely proportional to the stability value. The iterative correction calculation unit is called to calculate the weight update result in each iteration cycle according to the following formula: Where w represents the original associated element weight vector, w' represents the updated weight vector, α represents the adaptive learning rate, whose value is determined by the mean and variance of the stability quantization sequence, and g represents the gradient direction vector. During the iteration process, the update speed is matched with the intensity of construction disturbance by dynamically adjusting the value of α. When the stability value fluctuates greatly, the value of α is appropriately increased to quickly respond to weight changes; when the stability value fluctuates little, the value of α is decreased to maintain a steady state of weights. After the iteration is completed, all updated elements are combined into a dynamic weight adjustment vector, which serves as the input for subsequent data structure reorganization.

[0095] By combining an adaptive gradient update strategy with stability values, the optimal alternative anchor point selection result from the previous step is transformed into a dynamic weight adjustment vector that reflects the current construction disturbance characteristics. This enables the temporal adaptive correction of the weights of related elements, thereby improving the responsiveness of dynamic semantic anchor points to environmental changes and the accuracy of cause identification.

[0096] For example, regarding the drift anchor point during the main transformer installation phase of a substation, the original associated element weight vector is [0.35, 0.40, 0.25], corresponding to the elements "crane machinery status", "high-altitude worker qualification", and "wind speed threshold", respectively. The historical execution stability quantification sequence of the optimal alternative anchor point is [0.90, 0.75, 0.60], and after time alignment, the gradient direction vector is [+0.10, -0.05, -0.15]. The adaptive learning rate α is calculated and set to 0.08 based on the mean of the stability sequence (0.75) and the variance (0.015). The iterative correction formula is executed as follows: First element w' = 0.35 + 0.08 × 0.10 = 0.358, Second element w' = 0.40 + 0.08 × (-0.05) = 0.396, Third element w' = 0.25 + 0.08 × (-0.15) = 0.238. After the iteration is completed, the dynamic weight adjustment vector is [0.358, 0.396, 0.238]. In the on-site construction disturbance simulation test, it was verified that the adjustment vector significantly improves the diagnostic response to the delay of hoisting operation caused by excessive wind speed after the dynamic anchor point is generated, and can accurately capture the causes of multi-factor coupling deviation in a low computing power environment.

[0097] S4.5: Integrate the basic semantic tags, spatiotemporal constraint templates, and dynamic weight adjustment vectors of the original drift anchor point identifiers to be processed, perform data structure reorganization and version serialization operations, and generate dynamic semantic anchor point instances with temporal adaptability and context awareness.

[0098] Step S5: When a schedule deviation occurs, extract the set of active dynamic semantic anchors at the time of the deviation, calculate the semantic fit deviation between each dynamic semantic anchor and the actual observed value of its bound feature, and convert the numerical residuals into discrete-level semantic distance mapping results. Specifically, this includes: S5.1: Obtain the process time slice data at the moment of deviation occurrence, filter out the dynamic semantic anchors in the active state based on the spatiotemporal constraint template in the dynamic semantic anchor set, and generate an active anchor subset containing basic semantic labels and related element weight vectors to lock the key monitoring objects in the current construction scenario.

[0099] The input conditions include process time slice data at the time of deviation occurrence and a dynamic semantic anchor set generated by step S4. Each anchor in this set contains a basic semantic label, a spatiotemporal constraint template, and a weight vector of associated elements. After receiving the process time slice data, a timestamp-accurate alignment process is performed. The start and end times of the slice data are matched with the time window required by the spatiotemporal constraint template bound to each anchor in the dynamic semantic anchor set to determine the suitability of the template with the current slice time conditions. For anchors that have passed the time condition matching, the environmental condition parameters (such as meteorological thresholds, personnel qualification status, etc.) defined in their spatiotemporal constraint template are read and combined with the corresponding measured fields in the slice data for Boolean logic judgment to calculate the environmental condition satisfaction flag. For anchors with a true satisfaction flag, a data entity containing its basic semantic label and associated element weight vector is generated and stored in the active state cache. Unique identifier verification is performed on all data entities in the active state cache to ensure that the anchor data matched in subsequent steps does not produce duplication or conflict. Through the above filtering and attribute binding process, the dynamic semantic anchor point set from the previous step is transformed into an active anchor point subset containing basic semantic tags and related element weight vectors, thereby enabling the locking of key monitoring objects in the current construction scenario.

[0100] For example, in the time slice data of the main transformer installation process in a substation project, the start and end time is from 08:00 to 16:00. The measured wind speed on site is 6.5 m / s, the temperature is 18℃, and the pass rate of the high-altitude operation personnel qualification review is 100%. There is an anchor point in the dynamic semantic anchor point set with the basic semantic label "lifting operation". Its spatiotemporal constraint template requires a duration ≥ 4h, a wind speed ≤ 8 m / s, and a personnel qualification review pass rate ≥ 90%. After matching the spatiotemporal constraint template of this anchor point with the time slice, the duration condition is met. The environmental condition parameters are read and compared with the measured values. The wind speed condition and the personnel qualification condition are met, generating a satisfaction flag. The associated element weight vector of this anchor point is [0.45, 0.35, 0.20], corresponding to the weights of the lifting machinery status, personnel qualification, and meteorological conditions, respectively. This anchor point is bound to the basic semantic label and weight vector and stored in the active state cache. After verification by a unique identifier and no conflict, this anchor point is finally included in the active anchor point subset. In the subsequent S5.2 step, this subset is used as input for the weighted matching operation, which can effectively improve the accuracy of identifying the causes of deviations.

[0101] S5.2: For each dynamic semantic anchor point in the active anchor point subset, extract the measured sequence data of its bound related elements in the actual construction environment, and use a multidimensional vector alignment algorithm to perform a weighted matching operation between the measured sequence data and the weight vector of the related elements to generate an original numerical residual vector that represents the difference between theoretical expectations and actual state.

[0102] For each dynamic semantic anchor in the active anchor subset, read the multi-dimensional measured sequence data of the associated elements bound to it in the actual construction environment. The data sources include the operating status of mechanical equipment collected by sensors at the construction site, the qualification records of operators, environmental parameters monitored by meteorological monitoring, and the inventory logs of materials in place, and keep the timestamps of each data channel strictly aligned.

[0103] The measured sequence data are grouped according to the physical dimensions of the associated elements. Normalization transformation is performed on the components with consistent dimensions to eliminate the differences in numerical scale under different unit systems and ensure the comparability of subsequent matching operations.

[0104] The multidimensional vector alignment algorithm is invoked to match the normalized measured sequence vector with the weight vector of the associated elements bound to the dynamic semantic anchor point in one dimension. The product operation is performed according to the weight coefficient and accumulated to form the theoretical expected value vector. During the matching process, interpolation compensation is performed for time delay factors in different dimensions to ensure alignment accuracy.

[0105] The deviation components between the actual state vector and the theoretical expected value vector are extracted through difference operations. Residual calculations are then performed on the deviation values ​​for each dimension, using the following formula: in, Let i be the residual value of the i-th dimension. For the measured value components, These are the weighting coefficients for related elements. This is a theoretically expected reference value.

[0106] The residual values ​​of all dimensions are combined in the original dimension order to form an original numerical residual vector that represents the difference between the theoretical expectation and the actual state, providing input for subsequent semantic distance mapping.

[0107] By using multidimensional vector alignment and residual calculation processing, the active anchor point subset selected in the previous step is transformed into a raw numerical residual vector that can quantify the deviation of the construction state, thus realizing the semantic-to-numerical mapping basis in the identification of deviation causes.

[0108] For example, in a substation construction project, the active anchor point subset includes two items: "lifting operation" and "high-voltage test". The associated elements of the lifting operation include crane amplitude status, operator qualification level, and wind speed conditions. The measured sequence data are amplitude status 0.85, qualification level 4, and wind speed 12 m / s, respectively; the corresponding weight vectors are 0.6, 0.25, and 0.15, and the theoretical expected reference values ​​are 0.9, 5, and 8 m / s, respectively. When normalizing the wind speed dimension, the measured values ​​and reference values ​​are unified to a dimensionless scale. After matching operations, the residual values ​​are calculated, where the residual for the crane amplitude dimension is 0.31, the residual for the operator qualification dimension is 2.75, and the residual for the wind speed dimension is 10.8. The combined residual values ​​form the original numerical residual vector [0.31, 2.75, 10.8]. This vector directly reflects the multidimensional state of the difference between the key elements at the construction site and the theoretical expectations, providing a highly reliable data foundation for subsequent deviation calculation and semantic level mapping. The application effect is that the real-time performance of dynamic semantic anchor point deviation detection is significantly improved and the misjudgment rate is reduced.

[0109] S5.3: Based on the original numerical residual vector, apply the preset semantic distance mapping rule to normalize the scalar transformation of each dimension component in the original numerical residual vector, eliminate the influence of different physical dimensions on the deviation assessment, and generate a dimensionless comprehensive semantic deviation index to quantify the overall mismatch degree of a single anchor point.

[0110] For the generated original numerical residual vector, the difference values ​​of the associated elements corresponding to each dimension are read, and a normalization processing pipeline based on the preset semantic distance mapping rule is established to uniformly convert the residual components containing different physical dimensions into comparable scalar forms.

[0111] Each component in the numerical residual vector is input into the normalization module, and the dimension coefficient matrix matched by the feature category is called to scale the measured and theoretical differences to the dimensionless range, ensuring that multidimensional features have equal evaluation weights under a unified scale.

[0112] Vector aggregation is performed on the normalized components, and the weighted mean method is used to calculate the comprehensive semantic deviation. The weight coefficients are quantitatively assigned based on the importance ranking table of the related elements within the dynamic semantic anchor point.

[0113] The comprehensive semantic deviation index is calculated using the following formula: in, Let be the weight coefficient of the i-th element. This represents the normalized residual components.

[0114] The obtained comprehensive semantic deviation index is recorded in the state attribute of the current dynamic semantic anchor point, and used as the direct input for subsequent threshold determination and discrete level mapping, realizing the technical transformation from multi-dimensional original differences to a single comprehensive mismatch index.

[0115] Through the above normalization and weighted aggregation methods, the original numerical residual vector from the previous step is transformed into a dimensionless comprehensive semantic deviation index, which enables accurate quantification of the overall mismatch degree of a single anchor point and provides a quantitative basis with a unified scale for subsequent semantic state mapping.

[0116] For example, in a substation construction deviation detection scenario, the original numerical residual vector is [4.5, -0.8, 12], corresponding to differences in the qualifications of high-altitude workers, the condition of lifting machinery, and the ambient wind speed, respectively. For the difference in personnel qualifications, a dimensional coefficient of 0.1 is set; for the difference in equipment condition, a dimensional coefficient of 0.5 is set; and for the difference in wind speed, a dimensional coefficient of 0.05 is set. After normalization, the components are 0.45, -0.4, and 0.6, respectively. According to the anchor point attribute table, the weight coefficients for the three types of elements are 0.4, 0.35, and 0.25, respectively. The comprehensive semantic deviation is calculated using the formula, and the result is 0.21, indicating that the comprehensive mismatch of the dynamic semantic anchor point is in the low deviation range. This processing eliminates the interference of direct comparison between elements of different dimensions, making the deviation index uniformly comparable in subsequent threshold determination, and significantly improving the accuracy and robustness of semantic state mapping.

[0117] S5.4: Based on the comprehensive semantic deviation index, execute the multi-level threshold interval judgment logic to map the continuously distributed comprehensive semantic deviation index to the predefined discrete semantic state space, generate discrete level semantic distance labels with clear business meaning, and realize the semantic leap from quantitative calculation to qualitative diagnosis.

[0118] Based on a dimensionless comprehensive semantic deviation index as input, a predefined multi-level threshold interval set is invoked. The boundary value parameters of each interval are read and mapped to a discrete semantic state space identifier set in ascending order. Interval boundary value comparison processing is performed, comparing the comprehensive semantic deviation index with the lower and upper limits of each interval to generate a state index vector corresponding to the matching interval. The semantic state mapping table is invoked to map each index value in the state index vector to a business semantic state label, such as standard level labels for complete mismatch, partial offset, and temporary detour, forming a preliminary discrete level sequence. State label normalization processing is performed, using a lexical consistency check algorithm to verify the consistency of each label with the construction deviation diagnosis semantic system, eliminating conflicting state labels and correcting mapping anomalies. This generates a discrete level semantic distance label sequence with clear business meaning that can be used for process relationship breakage analysis.

[0119] By using a multi-level threshold interval determination and semantic state mapping method, the dimensionless comprehensive semantic deviation index of the previous step is transformed into a discrete level semantic distance label, realizing a semantic leap from quantitative deviation calculation to qualitative deviation diagnosis.

[0120] For example, in the scenario of diagnosing deviations in the construction progress of substation projects, the comprehensive semantic deviation index is set to a value range of 0 to 1, with predefined multi-level threshold intervals of [0, 0.25), [0.25, 0.5), [0.5, 0.75), and [0.75, 1], corresponding to the semantic status labels "complete match," "partial offset," "temporary detour," and "complete mismatch," respectively. When the calculated comprehensive semantic deviation index of a certain dynamic semantic anchor point is 0.82, interval comparison processing is performed, and its value falls into the interval [0.75, 1], which is mapped to the "complete mismatch" label. The calculation steps for the comprehensive semantic deviation index at this time are as follows: in, As a comprehensive semantic deviation index, For the number of related elements, The square of the normalized residual value in the i-th dimension is given. Input parameters include normalized residual values ​​of the three-dimensional associated elements: 0.7, 0.85, and 0.91. After summing the squares and dividing by m=3, we get D²=0.688. Taking the square root, we get D≈0.82, which falls within the complete mismatch interval. In different cases, when D=0.46, it is mapped as "partial offset," and when D=0.32, it is mapped as "temporary detour." The overall mapping output can be directly used in subsequent S5.5 to construct the semantic fit deviation dataset. In multiple batch validations, the consistency between the mapped labels and the manual diagnostic results in this step is significantly improved, ensuring the accuracy and robustness of the semantic transition process.

[0121] S5.5: Integrate the discrete-level semantic distance labels corresponding to all dynamic semantic anchors to construct a semantic fit deviation dataset containing anchor identifiers and mismatch level attributes. This dataset serves as the direct input for subsequent identification of semantic relationship break states, completing the key feature engineering processing in the deviation cause identification process.

[0122] Step S6: Based on the semantic distance mapping results and the process sequence relationship between each dynamic semantic anchor point, construct a semantic adaptation chain break map representing the semantic relationship breakage state to locate anchor point segments that continuously exhibit complete mismatch levels. Specifically, this includes: S6.1: Obtain the set of active dynamic semantic anchors and their corresponding semantic distance mapping results at the time of deviation occurrence. Use a directed graph construction algorithm to instantiate each dynamic semantic anchor as a graph node, and bind the basic semantic labels of the graph nodes and the semantic distance mapping results as node attribute vectors to generate an initial set of semantic nodes with state attributes.

[0123] Obtain the active dynamic semantic anchor set data and corresponding semantic distance mapping results output by step S5 at the time of deviation occurrence. Set the anchor identifier as the unique index and the mismatch level label as the state variable to establish a node attribute table to support the subsequent graph structure construction.

[0124] The directed graph construction algorithm is invoked to transform each dynamic semantic anchor point in the node attribute table into a graph node object. The basic semantic labels and semantic distance mapping results are simultaneously bound to the node attribute vector by object instantiation, ensuring that the node has searchable state attributes in the topology.

[0125] Based on the definition of node attribute vectors, attribute type declaration processing is performed on the graph nodes. The basic semantic label field is defined as a process attribute, and the semantic distance mapping result field is defined as a state attribute, so as to form a node description structure with dual attribute classification.

[0126] By using a node attribute index matching mechanism, the consistency between the spatiotemporal constraint template execution state of each node in the active dynamic semantic anchor set and its mapped discrete-level semantic distance result is verified. Node instances with attribute binding errors or abnormal states are removed to ensure the structural integrity and data accuracy of the initial semantic node set.

[0127] By using the node unique identifier generation rules, globally unique identification codes are assigned to the reserved nodes in sequence, and the identification codes are mapped to the identification field of the node object, so as to achieve the unique location of all nodes in the initial semantic node set in the subsequent process topology association.

[0128] By using the above-mentioned directed graph construction and attribute binding methods, the deviation dataset from the previous step is transformed into an initial semantic node set with state attributes, thereby achieving node object standardization and directed graph construction preparation before the generation of the process topology skeleton.

[0129] For example, in a substation construction deviation triggering scenario, the active dynamic semantic anchor set contains 12 anchor instance instances, and the semantic distance mapping result corresponding to each anchor is obtained from the dataset in step S5. The basic semantic labels include "main transformer hoisting," "casing test," and "oil handling," with state level labels of "complete mismatch," "partial offset," and "temporary detour," respectively. A directed graph construction algorithm is executed on the node objects, setting the node attribute vector structure to {process label: string type, state level: enumeration type}. In the attribute consistency verification stage, a node with a "complete mismatch" label is detected as not meeting the wind speed threshold in its spatiotemporal constraint template and is removed from the set. The final 11 nodes are assigned unique identifiers from N001 to N011. After instantiation, these nodes are stored in the initial semantic node set. The subsequent step S6.2 will directly execute topology connection rules on this set to form the process topology skeleton of the standard operation flow. In testing and verification, this method significantly improves the accuracy of node attribute matching and the reliability of fracture risk identification.

[0130] S6.2: Based on the basic semantic tags of each dynamic semantic anchor point in the initial semantic node set, retrieve the pre-stored substation construction process logic chain, extract the process sequence constraint relationship between adjacent dynamic semantic anchor points, and use topology connection rules to map the initial semantic node pairs with direct process dependency relationship into directed edges to generate a benchmark process topology skeleton representing the standard operation flow.

[0131] The substation engineering construction process logic chain is a knowledge model that digitizes and structures the standard construction process of substation engineering. The system defines unique identifiers for all standard process nodes and, crucially, clarifies the directional process sequence dependencies and constraints between these nodes. This pre-stored structured data provides the system with a benchmark process flow reference, enabling real-time monitored progress deviations—dynamic semantic anchors—to be precisely associated with specific process nodes through their semantic tags. Furthermore, based on the sequence constraints in this logic chain, it determines whether deviations should have existed and what direct process dependencies exist, thus providing a fundamental basis for constructing a benchmark process topology framework and identifying process disruption risks.

[0132] Based on the basic semantic tags of each dynamic semantic anchor in the initial semantic node set, the pre-stored substation construction process logic chain index table is read to generate a process relationship dataset containing process node identifiers and sequence constraint fields. Using the basic semantic tags of each dynamic semantic anchor as matching keys, a tag retrieval algorithm combining exact matching and fuzzy matching is called to locate the corresponding process nodes in the process relationship dataset, outputting a set of successfully matched node pairs. Adjacency search is performed on the located node pairs, filtering out node pairs with direct process dependencies based on the sequence constraint fields in the process relationship dataset, and removing nodes with no direct dependencies or only cross-stage dependencies. The filtered directly dependent node pairs are mapped according to topology connection rules, and a directed edge construction function is called to generate edge objects with directional attributes for each node pair, binding the unique identifiers of the start and end nodes to the two ends of the edge objects. All edge objects are topologically sorted to ensure that the process flow is consistent with the construction logic chain, forming a structured baseline process topology skeleton data structure. By using topological modeling based on the sequence of processes, the initial set of semantic nodes formed in the previous step is transformed into a baseline process topology skeleton with standard operation flow, thereby enabling visual modeling of the process sequence relationship of dynamic semantic anchor points.

[0133] For example, during the main transformer installation phase of a substation, the initial semantic node set includes three dynamic semantic anchors: "hoisting operation," "bushing test," and "oil handling." Their basic semantic tags correspond to nodes N1, N2, and N3 in the process logic chain, respectively. A pre-stored process logic chain index table records that nodes N1 and N2 have direct order constraints, nodes N2 and N3 have direct order constraints, and N1 and N3 are only indirectly related through N2. The tag retrieval algorithm uses a Levenshtein distance threshold of 0.85 for fuzzy matching, ensuring successful matching even with slight spelling differences in the field records. In the adjacency retrieval stage, node pairs with an order constraint value of 1 are considered direct dependencies, resulting in (N1, N2) and (N2, N3). The topology connection rule is defined as the directed edge construction logic from the starting node to the ending node, generating edge E1: N1 → N2 and edge E2: N2 → N3. Through topology sorting, the baseline process topology skeleton with the process flow N1 → N2 → N3 is obtained. In this skeleton, all edges and nodes are equipped with dynamic semantic anchor point identifiers and discrete level semantic distance labels, providing a structured data interface that can be directly invoked for subsequent fracture risk assessment.

[0134] S6.3: For each directed edge in the baseline process topology skeleton, read the semantic distance mapping result bound to its start node and end node, execute the semantic continuity verification logic, and if the semantic distance mapping result of the start node or end node is detected to be a complete mismatch level, then mark the directed edge as a break risk edge and generate a weighted process topology graph containing break risk identifiers.

[0135] For the set of directed edges in the baseline process topology skeleton, the attribute vectors of the start and end nodes corresponding to each directed edge are loaded, and the node attribute parsing module is called to extract the bound discrete level semantic distance label values. The extracted discrete level label values ​​are input into the semantic continuity determination unit, and the matching rules are used to verify whether there is a complete mismatch level between the label values ​​of the start and end nodes. The complete mismatch level determination is based on the correspondence between the level code and the "complete mismatch" business semantic in the pre-stored discrete level mapping rule table. A Boolean type continuity determination result vector is generated to identify the integrity status of each directed edge in semantic relationship. The Boolean type continuity determination result vector is aligned with the directed edge index table of the baseline process topology skeleton, and the fracture risk identification encoding module is called to assign fracture risk identification codes to the directed edges with true determination results. Weight assignment processing is performed on the directed edges assigned fracture risk identification codes. The weight value depends on the semantic fit deviation quantization value of the start and end nodes. The deviation quantization value is calculated by the following formula: in, (start) is the comprehensive semantic deviation index of the starting node. (end) represents the comprehensive semantic deviation index of the terminating node, which is averaged to generate a fracture risk weight value. This weight value, along with the fracture risk identifier, is bound to a directed edge data structure to construct a weighted process topology graph containing the fracture risk identifier and weight attributes. Through this processing method, the semantic distance mapping result from the previous step is transformed into topological data with fracture risk localization and risk intensity quantification capabilities, providing a highly reliable structured input for subsequent identification of consecutive mismatched node clusters.

[0136] For example, in the topological skeleton of the substation main transformer installation process logic chain, the starting node A (semantic label "equipment hoisting", discrete level label "complete mismatch", comprehensive semantic deviation index 2.4) and the ending node B (semantic label "positioning verification", discrete level label "partial offset", comprehensive semantic deviation index 1.8) are selected and input into the semantic continuity determination unit. According to the mapping rule table, node A is determined to be at the complete mismatch level, and the continuity determination result is true. Weight assignment processing is performed, substituting the index value of node A (2.4) and the index value of node B (1.8) into the formula to calculate the weight value of 2.1. The fracture risk identifier code "DR1" and the weight value 2.1 are bound to the structure of the directed edge from A to B, forming an edge object with risk attributes. After full-link execution, weight assignment and fracture risk encoding are performed on all edges in the process topology that meet the complete mismatch condition. The generated weighted process topology graph can significantly improve the accuracy of fracture risk location and provide structural consistency guarantee for subsequent mismatch node cluster retrieval.

[0137] S6.4: Based on the weighted process topology graph containing the fracture risk identifier, start the connected component search algorithm to traverse all node sequences marked as fracture risk edges, filter out the complete mismatch level node clusters that are continuous in the topology and have more than or equal to three nodes, and generate a set of candidate fracture subgraphs representing the substantial interruption of semantic relations.

[0138] The weighted process topology graph containing fracture risk identifiers is obtained as the input object. The initialization module in the connected component search algorithm is called to aggregate the start and end nodes of all fracture risk edges to form a candidate node set with risk attributes.

[0139] An adjacency list construction operation is performed on the candidate node set. The direct neighboring nodes of each candidate node are indexed into their adjacency list entries using the directed edge relationships defined in the topology graph, and the break risk indicators on the edges are recorded simultaneously to quickly filter out non-risk connections during traversal.

[0140] Based on the adjacency list, a combination of depth-first traversal (DFS) and breadth-first traversal (BFS) algorithms is called to traverse all nodes belonging to the marked fracture risk edges in turn. For each traversal path, the mismatch level distribution of nodes within the path is calculated and a level sequence of consecutive nodes is generated.

[0141] A continuity detection operation is performed on the grade sequence. The sliding window mechanism is used to check whether the number of nodes in the window is greater than or equal to three, and to determine whether the semantic distance mapping result of all nodes in the window is equal to the complete mismatch grade, so as to form a candidate sequence of node clusters that meets the continuous complete mismatch condition.

[0142] The candidate node cluster sequence is mapped back to the original weighted process topology graph. The subgraph extraction function is called to extract the subgraph structure containing the candidate node cluster and its internal fracture risk edges. The topological connectivity of the node cluster is used as the screening condition to generate a set of candidate fracture subgraphs that represent the substantial interruption of semantic relations.

[0143] By searching connected components and filtering continuous mismatch patterns, the fracture risk identifiers and topological skeletons from the previous step are transformed into a set of candidate fracture subgraphs with precise localization, thereby achieving a structured characterization of the interruption state of semantic relationships of dynamic semantic anchors.

[0144] For example, in the monitoring of the main transformer installation process in a substation, the input weighted process topology graph contains 15 dynamic semantic anchor nodes and 22 directed edges, of which the directed edges marked with fracture risk account for approximately 0.4%. Using an adjacency list, the adjacency relationships of risk nodes A, B, C, and D are stored in a sparse matrix form, with weight values ​​corresponding to the fracture risk level of the edges. A BFS traversal is used, starting from node A. The level sequence corresponding to the path A→B→C→D is [complete mismatch, complete mismatch, complete mismatch, partial offset]. The sliding window length is set to 3. When the window meets the condition of continuous complete mismatch and node count = 3 at A→B→C, it is selected as a candidate node cluster. Using this candidate cluster as the core, a subgraph containing A, B, C, and their risk edges is extracted from the original topology graph, forming a candidate fracture subgraph set. This set successfully triggers the "main transformer installation delay" mode in the local anomaly pattern library during subsequent matching, completing the key location in the deviation cause chain and improving the real-time performance and accuracy of deviation cause identification.

[0145] S6.5: Perform complete encapsulation on each candidate fracture subgraph in the candidate fracture subgraph set, integrate the dynamic semantic anchor point identifiers of its internal nodes, the semantic distance mapping result sequence and the process context snapshot of the boundary, and output the finally determined semantic adaptation chain fracture map as the direct input object for triggering the matching reasoning of the local cached construction anomaly pattern library.

[0146] Based on the input conditions of the candidate fracture subgraph set, the unique identifiers of all dynamic semantic anchor points within each candidate fracture subgraph and their corresponding semantic distance mapping results are extracted, and a node attribute vector matrix is ​​constructed to ensure the integrity of the internal state information.

[0147] The node attribute vector matrix is ​​serialized and encapsulated, and the discrete-level semantic distances of each node are arranged into a state sequence according to the process topology order to maintain the temporal consistency of the mismatch states within the fracture subgraph.

[0148] Read the boundary process nodes of the candidate fracture subgraph and their context positions in the process logic chain, call the process context snapshot generation module to extract the multi-dimensional state data of the two process stages before and after the fracture subgraph, compress and encode them into compact snapshot data blocks.

[0149] An association binding operation is performed between the node attribute vector matrix and the compact snapshot data block, incorporating the internal state and boundary context of the fracture subgraph into a structured encapsulation object, thereby achieving a joint expression of the fracture state and the process background.

[0150] The structured encapsulated object is uniquely encoded and assigned a version identifier, and stored as a standard record unit in the semantic adaptation chain break graph, ensuring that the local cached construction anomaly pattern library matching inference process can be directly triggered.

[0151] Through the above encapsulation and encoding processing, the candidate fracture subgraph from the previous step is transformed into a semantic adaptation chain fracture map with complete internal mismatch states and external process context, thereby achieving accurate positioning and traceable triggering of fracture states in deviation cause identification.

[0152] For example, in a substation transformer installation delay scenario, the candidate fracture subgraph contains five dynamic semantic anchor points, labeled DT001 to DT005, with a semantic distance mapping sequence of [complete mismatch, complete mismatch, complete mismatch, partial offset, complete mismatch]. The node attribute vector matrix is ​​stored in a 5-row × 2-column format, where the first column is the anchor point identifier and the second column is the discrete level label. The boundary process context includes the "lifting operation" process status before DT001 and the "oil handling" process status after DT005. The size of the snapshot data block after compression is 2KB. During the encapsulation process, the node attribute sequence and the snapshot data block form a structured encapsulation object through binding operations. This object is assigned the version identifier V2024.06 when generating the fracture graph record unit. When used to trigger anomaly pattern library matching, this fracture graph can directly retrieve and compare with the anomaly pattern library template, quickly matching the typical pattern "transformer installation delay → bushing test skipped → oil handling delay". In the verification test, the trigger matching time of the encapsulated recording unit was significantly shortened, and the accuracy of fracture state positioning was greatly improved.

[0153] Step S7: For anchor segments in the semantic adaptation chain break map that show three or more consecutive complete mismatch levels, the locally cached construction anomaly pattern library is invoked for matching inference to generate a deviation cause chain explanation containing typical deviation patterns. Specifically, this includes: S7.1: Obtain anchor point fragment data with three or more consecutive complete mismatch levels in the semantic adaptation chain fracture map, and use the graph substructure extraction algorithm to perform feature extraction processing on the node identifier sequence and semantic distance mapping results in the anchor point fragments to generate a standard fracture feature vector set representing the current construction interruption characteristics.

[0154] The input conditions include semantic adaptation chain fracture map data from the output of step S6, which contains dynamic semantic anchor fragments with three or more consecutive complete mismatch levels, node identifier sequences, and semantic distance mapping results. For this input, the processing first calls a graph substructure extraction algorithm based on the node identifier sequences to locate and extract the node sequence corresponding to each anchor fragment in the map according to preset process dependency constraints, while retaining its topological index position in the fracture map. The extracted node sequences are then attribute-aligned with the corresponding semantic distance mapping results, and the mismatch level is bound as a node state parameter to the node attribute matrix, forming a node feature structure with state labels. Multidimensional feature encoding is performed on the node feature structure, encoding the basic semantic labels as fixed-length categorical features, encoding the discrete level mapping results as ordinal features, and encoding the relative positions of nodes in the fragments as position index features, resulting in an original feature vector set. Normalization and feature aggregation are then performed on the original feature vector set, aggregating the feature vectors of each node using a graph convolution embedding method to form a comprehensive vector representation of the overall fracture characteristics of the fragment. A feature selection mechanism is employed to optimize the dimensionality of the comprehensive vector based on node state variance, level distribution entropy, and segment length weights. Low-contribution features are eliminated, while high-sensitivity features are retained to form a standard fracture feature vector set. Through this processing method, the fracture map data from the previous step is transformed into a standard fracture feature vector set with a unified encoding format and high-sensitivity features, achieving a structured abstraction of the current construction interruption state.

[0155] For example, in a substation main transformer installation delay scenario, the fracture map output by S6 contains a node identifier sequence [ND_034, ND_051, ND_062, ND_078], corresponding to semantic distance mapping results of [complete mismatch, complete mismatch, complete mismatch, partial offset]. The graph substructure extraction algorithm is used to locate this sequence according to the process dependency order, corresponding to the process chain of "main transformer installation → bushing test → oil handling → cooling system commissioning". Semantic labels are encoded as four-dimensional classification features, mismatch levels are mapped to ordinal feature values ​​3, 3, 3, 2, and position indices are encoded as 0, 1, 2, 3. After forming the original feature vector matrix, Z-score normalization is performed on each dimension to eliminate the influence of different dimensions. A graph convolution aggregation method is used to calculate the fragment comprehensive feature vector, generating a comprehensive vector with a node state variance of 1.27, a level distribution entropy of 0.81, and a fragment length weight of 1.0. In the feature selection process, a variance threshold of 0.5 and a distribution entropy threshold of 0.7 are set. Two types of highly sensitive features, semantic level and location index, are retained. The output standard fracture feature vector set is [[3,0],[3,1],[3,2],[2,3]]. This vector set can significantly improve the retrieval accuracy when matching the abnormal pattern library.

[0156] S7.2: Based on the standard fracture feature vector set, call the locally cached construction anomaly mode library, and use the multi-dimensional index retrieval mechanism to perform key-value matching operation on the pre-stored typical deviation mode templates in the library to filter out a set of candidate anomaly modes that are highly similar to the standard fracture feature vector set in terms of topological structure and mismatch level distribution.

[0157] S7.3: For each candidate abnormal pattern in the candidate abnormal pattern set, extract its associated historical deviation cause labels and process context constraints, and use the Bayesian probabilistic inference algorithm to calculate the posterior matching probability value between the current standard fracture feature vector set and each candidate abnormal pattern, so as to generate a pattern matching score sequence with quantified confidence.

[0158] During the matching inference process after obtaining the candidate abnormal pattern set, the candidate abnormal pattern set is used as the input object, and the standard fracture feature vector set formed in the current step and its numerical components of each dimension are loaded at the same time to ensure that the target of feature matching is consistent with the reference benchmark.

[0159] For each anomaly pattern in the candidate anomaly pattern set, historical deviation cause tags bound to the pattern are extracted from the locally cached construction anomaly pattern library, and its process context constraint parameters are read at the same time to form a pattern attribute set containing causal logic information and environmental limitation information. This pattern attribute set is used to define the prior conditions for subsequent probability calculations.

[0160] The standard fracture feature vector set is parameter-mapped to the pattern attribute set, and a joint matrix of the observed data vector and the prior label vector is constructed according to the correspondence of the attributes in each dimension, so as to perform conditional probability derivation in Bayesian probabilistic inference.

[0161] Setting prior probabilities in Bayesian inference operations To determine the frequency of candidate anomaly patterns in historical data, a likelihood function is defined. Given the observation probability of the current fracture feature vector set under this mode, in order to apply Bayes' theorem... Calculate the posterior matching probability value, where For the current fracture feature vector set, This is a candidate abnormal pattern.

[0162] The above probability calculation results are stored in the pattern matching scoring sequence, and a corresponding quantitative confidence score is generated for each candidate anomaly pattern, which serves as the direct basis for subsequent selection and screening.

[0163] By using Bayesian probabilistic inference, the candidate abnormal pattern set from the previous step is transformed into a pattern matching score sequence with quantifiable confidence, thereby achieving objective pattern evaluation based on probability weights.

[0164] S7.4: Perform a maximum value optimization operation based on the pattern matching score sequence to lock the target typical deviation pattern with the highest posterior matching probability value, and use causal chain reconstruction technology to align and fuse the process dependencies contained in the target typical deviation pattern with the actual temporal state of the current anchor segment to generate an initial draft explanation of the deviation cause chain.

[0165] Based on the quantified confidence data of the pattern matching scoring sequence, the scoring values ​​corresponding to all candidate abnormal patterns are read and numerical comparison operations are performed. The set of scoring values ​​is input into the maximum value selection component to calculate the target typical deviation pattern index with the highest posterior matching probability value. The process dependency relationship data structure of the target typical deviation pattern is called, and the node identifiers in the dependency relationship are associated and aligned with the node identifier sequence of the current anchor segment using the process node matching algorithm to form a preliminary matching mapping table. For each pair of matching nodes in the mapping table, the actual time-series state data of the current anchor segment is read, and the execution time series of the process dependency relationship and the actual time-series state are phase-adjusted using a time-series synchronization mechanism to ensure the alignment consistency of the two in the time dimension. The time-series aligned process dependency relationship and the semantic attribute vector of the current field state are fused, and the causal chain structure after fusion is calculated through the association relationship enhancement model to generate a structured draft chain containing causal pointing information. Through the above processing method, the pattern matching scoring sequence of the previous step is transformed into an initial deviation cause chain explanation draft after the process dependency relationship and field state alignment are fused, realizing the technical effect of candidate pattern locking and field data fusion construction.

[0166] For example, in a case of a delay in the installation of a substation main transformer, the pattern matching scoring sequence contains five candidate abnormal patterns with scores of 0.62, 0.58, 0.74, 0.69, and 0.55, respectively. After maximum value optimization, the candidate pattern with a score of 0.74 is selected as the target typical deviation pattern. The process dependency of this pattern is "main transformer installation → bushing test → oil handling," and the current anchor point segment node sequence is "main transformer installation mismatch, bushing test mismatch, oil handling mismatch." After alignment and matching, the actual timing status shows that the main transformer installation started at 14:00 and ended at 18:00. The bushing test was canceled due to unfavorable environmental conditions, and the oil handling was postponed to 10:00 the following day. The timing synchronization mechanism is invoked to adjust the preset execution times in the dependencies (main transformer installation is expected to end at 12:00, the bushing test will proceed immediately, and the oil handling is expected to be completed at 08:00 the following day). The adjustment result perfectly matches the actual observations on the timeline. By integrating the adjusted dependencies with the observed state, the resulting causal chain structure is "delay in main transformer installation leads to failure of bushing testing, which in turn causes delay in oil handling." This causal chain serves as the initial draft explanation of the deviation cause chain. During field verification, this draft explanation can accurately reflect the key delay chains in the construction process, significantly improving the accuracy and usability of deviation cause identification.

[0167] S7.5: Perform semantic normalization verification on the initial draft explanation of deviation cause chain, and use a natural language generation template to convert the structured causal logic in the draft into natural language description text with clear business meaning, so as to output the final explanation result of deviation cause chain containing typical deviation patterns and specific cause logic.

[0168] Step S8: Write the newly generated dynamic semantic anchors and their drift paths into the incremental anchor evolution log, recording the dynamic semantic anchor identifier, drift trigger condition, replacement anchor identifier, weight adjustment direction vector, and process context snapshot to complete the closed-loop update of semantic modeling. Specifically, this includes: S8.1: Obtain the set of dynamic semantic anchor points generated in the current construction phase and the corresponding semantic drift judgment results. Based on the preset log structure template, perform field mapping processing on the set of dynamic semantic anchor points to generate a standardized anchor point record unit containing dynamic semantic anchor point identifiers, basic semantic tags, and spatiotemporal constraint templates.

[0169] S8.2: Extract the original monitoring data sequence that triggers semantic drift determination from the standardized anchor point recording unit, and use the time-series event association algorithm to extract causal features from the original monitoring data sequence to generate a discretized trigger event descriptor that represents the drift triggering condition.

[0170] The aforementioned time-series event association algorithm is a computational method for automatically identifying key events and mining their causal relationships from time-series data. It transforms continuous raw data sequences into structured discrete event chains by analyzing temporal patterns, abrupt changes, and event intervals in the monitored data stream, and infers the underlying causal logic between events. This extracts a set of causal features that can explain state changes, providing a foundation for subsequently generating understandable drift-triggered descriptors.

[0171] The drift triggering condition refers to the specific data pattern or event combination extracted from the original monitoring data through a time-series event association algorithm that causes semantic drift of the dynamic semantic anchor. It is represented as a discretized triggering event descriptor, which clarifies the data-driven basis for the system to determine whether the anchor is invalid or needs to be updated. For example, continuous exceeding of the standard reading of a specific sensor, or the successive occurrence of abnormal status of multiple related processes within a time window, etc., are the decision conditions for initiating the subsequent anchor replacement and model update process.

[0172] S8.3: Retrieve the information of the alternative anchor points selected in the semantic drift determination process, calculate the semantic compatibility score between the original dynamic semantic anchor point and the alternative anchor point based on the process logic chain topology, and generate an associated jump pointer and weight adjustment direction vector pointing to the alternative anchor point identifier.

[0173] S8.4: Extract the multi-dimensional state data segment of the construction site at the moment the deviation occurs, and use the context snapshot compression algorithm to perform key feature dimensionality reduction processing on the multi-dimensional state data segment of the construction site to generate a compact process context snapshot data block that reflects the current process context environment.

[0174] S8.5: Integrate the standardized anchor point recording unit, discretized trigger event descriptor, associated jump pointer, weight adjustment direction vector and compact process context snapshot data block, and perform incremental data appending write operations to the incremental anchor point evolution log stored locally to complete the semantic modeling closed-loop update and output an anchor point strategy evolution file that can be migrated.

[0175] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0176] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0177] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for monitoring and analyzing construction progress deviations in substation engineering based on digital twins, characterized in that, Specifically, it includes: S1: Extract structured semantic units from the preset engineering document source and construct an initial semantic anchor point pool containing basic semantic tags, spatiotemporal constraint templates and related element weight vectors; S2: Based on the initial semantic anchor pool, collect multi-source heterogeneous data and map it to the corresponding spatiotemporal constraint template to generate a temporal state sequence; S3: Perform continuity verification on the time-series state sequence. If it is determined that the spatiotemporal constraint state of the basic semantic tag continues to deviate from the spatiotemporal constraint template bound to the basic semantic tag, then trigger the semantic drift determination process and generate a drift anchor point identifier to be processed. S4: Based on the identifier of the drift anchor point to be processed, determine the alternative anchor point in the process logic chain of the current construction stage, and dynamically adjust the weight vector of the associated elements bound to the drift anchor point to be processed in combination with the historical execution stability index of the alternative anchor point to generate dynamic semantic anchor points. S5: Extract the set of dynamic semantic anchor points at the time when the construction progress deviation of the substation project occurs, calculate the semantic fit deviation value between each dynamic semantic anchor point and the actual observed value of its bound element, and convert it into a semantic distance mapping result. S6: Based on the semantic distance mapping results and the process sequence relationship between each dynamic semantic anchor point, construct a semantic adaptation chain break map; S7: If an anchor segment that meets the preset severity condition is detected in the semantic adaptation chain break map, the local cached construction anomaly pattern library is called to perform matching reasoning and generate an explanation of the deviation cause chain.

2. The method for monitoring and analyzing construction progress deviations in substation engineering based on digital twins according to claim 1, characterized in that, The spatiotemporal constraint state continuously deviates from the spatiotemporal constraint template bound to the basic semantic tag, including: within at least two adjacent time steps, the spatiotemporal constraint state of the basic semantic tag does not satisfy the spatiotemporal constraint template it is bound to.

3. The method for monitoring and analyzing construction progress deviations in substation engineering based on digital twins according to claim 2, characterized in that, Step S3 specifically includes: Obtain the basic semantic tags arranged by daily granularity and their corresponding spatiotemporal constraint template execution state records in the time-series state sequence. Use the sliding time window algorithm to extract and align the execution state records of two consecutive time steps to generate a dual-time-period state comparison vector containing the current time and the previous time-series time. Based on the dual-time-period state comparison vector, the Boolean logic NOT operation rule is used to perform reverse discrimination processing on the spatiotemporal constraint template satisfaction status within each time step, generating a discretized failure flag bit sequence; A continuous pattern matching detection is performed on the discretized failure flag sequence, and a sequence consistency check algorithm is used to identify whether there are continuous failure segments with failure flags of at least two adjacent time steps that are simultaneously true, thereby generating a potential drift trigger signal. Based on the potential drift trigger signal, a predefined drift confidence assessment mechanism is invoked to perform weighted fusion calculation on the historical stability weights of the basic semantic labels, and a valid drift determination result is generated. Based on the valid drift determination result, a unique identifier allocation operation is performed to map the corresponding basic semantic label to a specific drift anchor point identifier to be processed.

4. The method for monitoring and analyzing construction progress deviations in substation engineering based on digital twins according to claim 1, characterized in that, Step S4 specifically includes: The process logic chain data of the current construction stage and the identifier of the drift anchor point to be processed are obtained. The semantic compatibility matching calculation of the adjacent process nodes in the process logic chain is performed using a graph traversal algorithm to generate a semantic compatibility mapping table containing a set of candidate alternative anchor points and corresponding compatibility score values. Based on the set of candidate alternative anchor points, the historical execution records of each candidate alternative anchor point are extracted by calling the locally cached historical completion report database. The execution volatility index of each candidate alternative anchor point is calculated by statistical variance analysis, and a historical execution stability quantification sequence is generated. Based on the compatibility score in the semantic compatibility mapping table and the execution volatility index in the historical execution stability quantification sequence, a multi-objective weighted evaluation function is constructed to perform an optimal solution search operation, select the target alternative anchor with the highest comprehensive score, and generate the optimal alternative anchor identifier. Based on the historical execution stability quantification sequence corresponding to the optimal alternative anchor point identifier, an adaptive gradient update strategy is used to iteratively correct and calculate the weight vector of the associated elements bound to the original drift anchor point identifier to generate a dynamic weight adjustment vector. By integrating the basic semantic tags, spatiotemporal constraint templates, and dynamic weight adjustment vectors of the original drift anchor point identifiers to be processed, data structure reorganization and version serialization operations are performed to generate the dynamic semantic anchor point instance.

5. The method for monitoring and analyzing construction progress deviations in substation engineering based on digital twins according to claim 4, characterized in that, The semantic compatibility mapping table contains a set of candidate alternative anchors and corresponding compatibility scores.

6. The method for monitoring and analyzing construction progress deviations in substation engineering based on digital twins according to claim 4, characterized in that, Step S6 specifically includes: A set of dynamic semantic anchor points and their corresponding semantic distance mapping results at the time when the deviation in the construction progress of the electrical engineering is obtained. Each dynamic semantic anchor point is instantiated into a graph node using a directed graph construction algorithm. The basic semantic labels of the graph nodes and the semantic distance mapping results are bound into node attribute vectors to generate an initial set of semantic nodes. Based on the basic semantic tags of each dynamic semantic anchor point in the initial semantic node set, the pre-stored substation construction process logic chain is retrieved, the process sequence constraint relationship between adjacent dynamic semantic anchor points is extracted, and the initial semantic node pairs with direct process dependency relationship are mapped to directed edges using topology connection rules to generate the baseline process topology skeleton. For each directed edge in the baseline process topology skeleton, if the semantic distance mapping result of the starting node or ending node of the directed edge is a complete mismatch level, then the directed edge is marked as a break risk edge, and a weighted process topology graph is generated. Based on the weighted process topology graph, the connected component search algorithm is started to traverse all node sequences marked as fracture risk edges, and the clusters of nodes with complete mismatch level that are continuous in the topology and have more than or equal to three nodes are selected to generate a set of candidate fracture subgraphs. Each candidate fracture subgraph in the candidate fracture subgraph set is fully encapsulated, and its internal nodes' dynamic semantic anchor point identifiers, semantic distance mapping result sequences, and boundary process context snapshots are integrated to output the semantic adaptation chain fracture map.

7. The method for monitoring and analyzing construction progress deviations in substation engineering based on digital twins according to claim 6, characterized in that, After step S7, the following is included: S8: Write the newly generated dynamic semantic anchors and their drift paths into the incremental anchor evolution log, and record the dynamic semantic anchor identifier, drift trigger condition, replacement anchor identifier, weight adjustment direction vector and process context snapshot to complete the closed-loop update of semantic modeling.

8. The method for monitoring and analyzing construction progress deviations in substation engineering based on digital twins according to claim 1, characterized in that, The engineering document sources include: substation engineering construction standards and specifications, typical work instructions, and historical completion reports.

9. The method for monitoring and analyzing construction progress deviations in substation engineering based on digital twins according to claim 1, characterized in that, The multi-source heterogeneous data includes: on-site BIM model update logs, process inspection data, meteorological measurement sequences, and mobile terminal inspection texts.

10. The method for monitoring and analyzing construction progress deviations in substation engineering based on digital twins according to claim 1, characterized in that, The preset severity condition includes: the appearance of three or more anchor segments with a complete mismatch level in the semantic adaptation chain break map.