An engineering quality cause force quantification analysis method and system

Abstract

The application relates to the technical field of data processing and knowledge graph, in particular to an engineering quality reason power analysis method and system. The method comprises the following steps: acquiring multi-modal engineering data; performing classification preprocessing on the multi-modal engineering data to obtain standardized multi-modal feature data; performing semantic analysis on the standardized multi-modal feature data to construct a causal relationship structure; extracting a structured causal feature set according to topological information of the causal relationship structure; performing weighted calculation on each cause event in the structured causal feature set to obtain an initial weighted score of each cause event; quantitatively correcting the initial weighted score of each cause event based on external environmental factors to obtain a final score of each cause event; and outputting a causal effect value proportion of each engineering participant and external environmental factor based on the final score of each cause event. The application realizes objective quantitative attribution of engineering quality reasons through multi-modal data fusion and causal knowledge graph driving.

Description

Technical Field

[0001] This invention relates to the fields of data processing and knowledge graph technology, and in particular to a quantitative analysis method and system for engineering quality causal factors. Background Technology

[0002] In recent years, with the continuous expansion of the construction scale of my country's construction industry and the increasing diversity of participating entities, the number of engineering quality disputes caused by the collaborative work of multiple entities and the coupling effect of multiple factors has shown a significant upward trend. In the practice of judicial appraisal of engineering quality, causal force refers to the degree of contribution of the quality risk behavior of each responsible party to the damage to the engineering quality. Causal force is also known as the degree of causal contribution. The objective quantification of causal force is the core challenge in the appraisal of engineering quality disputes.

[0003] Currently, the determination of causation in engineering quality judicial appraisals mainly relies on the professional experience and subjective judgment of the appraisers, which has the following technical defects: First, in the face of complex engineering quality problems involving multiple coupled factors and intertwined causal chains, appraisers find it difficult to conduct systematic, multi-dimensional causal tracing analysis, leading to a significant deviation between the liability division results and the actual causal logic; Second, the parties involved have different understandings of the boundaries of rights and responsibilities and the performance obligations under the construction contract, and the appraisal process lacks a unified and objective quantitative basis for judgment. Subjective factors account for a high proportion in the appraisal discretion process, which can easily lead to biased conclusions, affecting the fairness and consistency of the appraisal results, and thus adversely affecting the credibility of engineering quality judicial appraisals and the efficiency of case handling.

[0004] Chinese patent application CN116452073A discloses an intelligent diagnostic method and system for the causal force of engineering quality judicial appraisal. The technical solution includes: acquiring appraisal data; cleaning and segmenting the appraisal data into short text sequences of equal length, comparing their similarity with behavioral descriptions in a quality risk behavior database, and eliminating text sequences with a similarity of less than 50%; inputting the pre-processed text sequences into a pre-trained LSTM neural network model to obtain the quality risk behavior categories and correlation feature vectors of each responsible party, with quality risk behavior categories categorized as technical, non-management, and non-standard; determining the responsibility proportion weight matrix of each responsible party using the analytic hierarchy process (AHP) through an expert consultation database; and finally calculating the responsibility proportion of each party based on the correlation feature vectors and the responsibility proportion weight matrix. However, the technical limitations of this patent are: it only processes a single text modality and cannot utilize images or structured data; its LSTM model can only output statistical correlation rather than causal contribution, and the weights rely on subjective expert assignment, making it difficult to eliminate confounding factors and obtain the net causal effect of each responsible party based on causal inference.

[0005] In summary, existing technologies for diagnosing the causal forces in engineering quality forensic appraisals still have the following shortcomings: they cannot comprehensively utilize multimodal appraisal data such as text, images, and structured data to reconstruct causal chains; they lack the ability to structurally model the causal relationship between quality-risk behaviors and engineering quality defects; they are difficult to make unbiased estimates of the net causal contributions of each responsible party in the presence of confounding factors; and they have not been able to break free from the reliance on subjective expert assessment. Therefore, there is an urgent need for an intelligent diagnostic method and system that can objectively quantify the causal forces in engineering quality forensic appraisals. Summary of the Invention

[0006] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0007] According to a first aspect of the present invention, a quantitative analysis method for the causes of engineering quality problems is provided, the method comprising the following steps:

[0008] Acquire multimodal engineering data, which includes text data, image data, and structured data;

[0009] The multimodal engineering data is classified and preprocessed to obtain standardized multimodal feature data.

[0010] Based on an engineering domain knowledge base, the standardized multimodal feature data is semantically parsed to construct a causal relationship structure that characterizes the causal associations between engineering participants, triggering events, and deviations from entity states.

[0011] Based on the topological information of the causal relationship structure, a set of structured causal features for each triggering event is extracted.

[0012] Based on a preset quantification parameter system, each causal event in the structured causal feature set is weighted and calculated to obtain an initial weighted score for each causal event; wherein, at least some parameters in the quantification parameter system are determined based on the topological information of the causal relationship structure.

[0013] The initial weighted scores of each triggering event are quantitatively adjusted based on external environmental factors to obtain the final score of each triggering event.

[0014] Based on the final scores of each inducing event, the causal effect ratio of each project participant and external environmental factor is output as the quantitative analysis result of the cause force of project quality.

[0015] According to a second aspect of the present invention, a quantitative analysis system for engineering quality causal factors is provided, comprising:

[0016] The multimodal data acquisition module is used to acquire multimodal engineering data, which includes text data, image data, and structured data.

[0017] The multimodal preprocessing module is used to classify and preprocess the multimodal engineering data to obtain standardized multimodal feature data.

[0018] The causal relationship construction module is used to perform semantic parsing on the standardized multimodal feature data based on the engineering domain knowledge base, and to construct a causal relationship structure to characterize the causal association between engineering participants, inducing events and deviations from entity states.

[0019] The causal feature extraction module extracts a set of structured causal features for each triggering event based on the topological information of the causal relationship structure.

[0020] The first calculation module is used to perform weighted calculations on each causal event in the structured causal feature set based on a preset quantification parameter system to obtain an initial weighted score for each causal event; wherein at least some parameters in the quantification parameter system are determined according to the topological information of the causal relationship structure.

[0021] The second calculation module is used to quantify and correct the initial weighted scores of each triggering event based on external environmental factors, so as to obtain the final score of each triggering event.

[0022] The attribution module is used to output the proportion of causal effect values ​​of each project participant and external environmental factor based on the final score of each inducing event, as the quantitative analysis result of the cause force of project quality.

[0023] This invention eliminates the reliance on subjective expert empowerment by using a multimodal data fusion and a causal knowledge graph topology-driven quantitative parameter system, thereby achieving objective quantitative attribution of engineering quality causal forces and improving the accuracy and interpretability of responsibility quantification.

[0024] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart of a quantitative analysis method for engineering quality causes provided in an embodiment of the present invention;

[0027] Figure 2This is a structural block diagram of an engineering quality cause force quantitative analysis system provided in an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0030] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0031] First Embodiment

[0032] This embodiment provides a quantitative analysis method for causal forces in engineering quality. Causal force refers to the degree of causal contribution of specific behaviors or events (including actions and omissions) of various engineering stakeholders during the construction or implementation of an engineering entity to the final deviation from the entity's final state, i.e., quality defects. This method can be applied to scenarios such as engineering quality dispute analysis, engineering quality accident investigation, and responsibility quantification in engineering judicial appraisal. It is used to automatically extract the causal relationship between triggering events and deviations from the entity's state from multimodal engineering data, and to quantify the causal effect proportion of each participating entity and external environmental factors.

[0033] like Figure 1 As shown, the method includes the following steps:

[0034] S100, acquires multimodal engineering data.

[0035] The multimodal engineering data includes text data, image data, and structured data. Text data includes, but is not limited to, engineering-related documents such as engineering contracts, construction logs, supervision reports, testing reports, complaints, and court transcripts; image data includes, but is not limited to, visual materials such as on-site photos of engineering quality defects, construction process photos, and engineering CAD drawings; and structured data includes, but is not limited to, quantitative information with clearly defined field structures such as engineering acceptance record forms, material testing data, and construction schedule plans.

[0036] Multimodal engineering data can be collected from engineering quality-related business systems. These business systems include, but are not limited to: engineering project management platforms, construction process recording systems, quality inspection information systems, supervision report management systems, contract and legal document databases, and engineering drawing and image archive management systems.

[0037] The data collection methods include at least one of the following:

[0038] (1) Connect with the information systems of engineering quality supervision agencies, testing units or construction companies through data interfaces to obtain structured engineering data in real time or periodically;

[0039] (2) Receive unstructured text data, including contract documents, construction logs, supervision reports, test reports, complaints, court transcripts, etc., by importing files;

[0040] (3) Obtain image data, including on-site photos of engineering quality defects, construction process photos, and engineering CAD drawings, through image acquisition equipment or batch import of image files;

[0041] (4) Read structured data tables such as project acceptance record tables, material testing data, and construction schedule plans through database connection.

[0042] The collected multimodal engineering data are classified and stored according to their original format and semantic categories, and a data index is established for subsequent unified classification preprocessing.

[0043] S200, perform classification preprocessing on the multimodal engineering data to obtain standardized multimodal feature data.

[0044] Specifically, the classification preprocessing includes the following steps:

[0045] S201, the text data is subjected to text normalization processing, and word segmentation and semantic fragment segmentation are performed using a large language model to obtain the first processed data.

[0046] The text normalization process includes: removing noise characters, standardizing the encoding format, and completing the identification and name standardization of professional entities related to engineering quality. In this embodiment, a general open-source large language model is used to perform the above processing, such as the Qwen series model, and the text is segmented into semantically complete fragments. A semantically complete fragment refers to a text unit with independent semantic units, defined by natural semantic boundaries such as periods, paragraphs, and chapters; the specific length of the fragment is adaptively determined based on the original text structure, without needing to pre-set a fixed number.

[0047] S202, the image data is processed by extracting defect area features and engineering construction node information through a multimodal visual language model, and then converted into structured text description to obtain the second processed data.

[0048] In this embodiment, a general multimodal visual language model, such as the Qwen-VL series model, is used to analyze the image. The defect area features refer to the visual attributes of the areas in the image that reflect deviations from the normal state of the engineering entity, such as cracks, leaks, honeycomb texture, and deformation. The engineering construction node information refers to information in the image reflecting specific construction procedures or engineering parts, such as foundation pouring, main structure capping, and equipment installation. The structured text description refers to organizing image features into a structured data format with predetermined fields and values. For example, for a site photograph showing cracks in a concrete wall, the model can generate the following structured text description: {“Object Type”: “Concrete Wall”, “Defect Type”: “Crack”, “Location”: “Between the 2nd and 3rd floors of the east exterior wall”, “Estimated Length”: “Approximately 1.2m”, “Estimated Width”: “Approximately 0.5mm”, “Associated Construction Node”: “3 days after the second-floor wall was poured”}. The above example is only to illustrate the form of the structured text description and does not constitute a limitation on the scope of protection.

[0049] S203, the structured data is numerically normalized and missing values ​​are filled in, and the semantic description of the table is extracted using a large language model to obtain the third processed data.

[0050] Specifically, numerical fields in the structured data are scaled to the [0,1] interval using a max-min normalization method; missing values ​​are handled using preset strategies, such as mean imputation, median imputation, or model prediction imputation based on similar samples. Simultaneously, a general open-source large language model, such as the Qwen series models, is used to extract semantic descriptions of the tables. These semantic descriptions include the meanings of column names, the logical relationships between row records, and the association information between the tables and engineering quality issues, thereby enhancing the contextual expressiveness of the structured data.

[0051] S204, perform multimodal feature fusion on the first processed data, the second processed data, and the third processed data to obtain the standardized multimodal feature data.

[0052] In this embodiment, the multimodal feature fusion adopts a feature concatenation method: the first processed data, the second processed data, and the third processed data are respectively encoded into vector representations of fixed length, and then concatenated in a preset order (e.g., text features, image features, structured features) to form a single feature vector, which serves as standardized multimodal feature data.

[0053] In other embodiments, an attention mechanism can be used to weightedly fuse the three types of features, or a multimodal transformer model can be used to achieve cross-modal information interaction. The above fusion methods are all conventional techniques in the art, and the specific network structure and parameters can be selected and configured by those skilled in the art based on the actual data scale and computing resources; these will not be elaborated upon here.

[0054] It should be noted that the Qwen series large language models and multimodal visual language models used in this embodiment are all open-source general-purpose models, and their internal network structures and training methods are well-known technologies in the field, and will not be elaborated here. The technical contribution of this application lies in applying these models to the preprocessing of multimodal engineering quality data, specifically:

[0055] For text-based data, the original text is used as the model input, and the model's natural language understanding ability is used to output segmented semantic fragments and identified engineering quality professional entities.

[0056] For image-based data, the original image is used as the model input, and the model's visual understanding and multimodal alignment capabilities are used to output structured text describing defect areas and construction node information.

[0057] For structured data, the serialized representation of the data table is used as the model input, and the semantic description text of the table is output.

[0058] The above application methods do not depend on the internal implementation details of a specific model. After reading this specification, those skilled in the art can use any general model with text understanding, visual understanding or multimodal understanding capabilities to implement the technical solutions of this application.

[0059] S300, based on the engineering domain knowledge base, performs semantic parsing on the standardized multimodal feature data to construct a causal relationship structure to characterize the causal association between engineering participants, triggering events and entity state deviations.

[0060] In this embodiment, the causal relationship structure is specifically implemented as a causal knowledge graph. The construction process of the causal knowledge graph is as follows:

[0061] S301 defines node types. These node types include three categories: project participant nodes, causal event nodes, and entity state deviation nodes. Project participant nodes correspond to various parties or personnel involved in construction, design, supervision, and development; causal event nodes correspond to specific behaviors or events that may lead to quality problems, such as substandard materials, incorrect procedures, or insufficient maintenance; and entity state deviation nodes correspond to specific abnormal conditions in the project entity, such as cracks, settlement, or insufficient strength.

[0062] S302, Extract relationship edges between nodes. Based on an engineering domain knowledge base, a combination of keyword matching and semantic rule extraction is used to extract three types of relationship edges from the standardized multimodal feature data.

[0063] In this embodiment, the engineering domain knowledge base refers to a pre-constructed structured knowledge set specifically designed for storing knowledge related to the field of engineering quality. This knowledge includes, but is not limited to, the following:

[0064] (1) Standard terms and entity names related to engineering quality, including the types of engineering participants such as construction unit, supervision unit, design unit, construction unit, types of engineering quality defects such as cracks, leakage, settlement, insufficient strength, construction technology and process names, etc.;

[0065] (2) The causal chain of common quality problems, that is, the causal relationship between known inducing events and deviations in physical state in the field of engineering quality is recorded in a structured form, such as "material substandard → insufficient bearing capacity", "inadequate curing → concrete cracks", "reinforcing bar spacing deviation → structural strength reduction", etc.

[0066] (3) Construction process standards and standard procedure sequence, such as typical procedure chains such as "formwork erection → steel bar binding → concrete pouring → curing → formwork removal";

[0067] (4) Rules for the division of rights and responsibilities in legal contracts, used to help identify the participation relationship between the project participants and the causative events;

[0068] (5) Preset parameters such as keyword list, causal pattern library, hierarchical threshold and correction coefficient.

[0069] Engineering knowledge bases can be stored and accessed using data structures such as databases, graph databases, or knowledge graphs. The construction of the knowledge base includes extracting and organizing knowledge from engineering quality-related standards, technical regulations, academic literature, historical appraisal cases, and expert experience, and then populating and updating it through manual review or semi-automated knowledge extraction tools. In this embodiment, the specific construction method of the engineering knowledge base is a conventional technique in the field and will not be elaborated here. The technical contribution of this application lies in applying this knowledge base to the semantic parsing of multimodal engineering quality data, the construction of causal relationship structures, and the calibration of quantitative parameters.

[0070] The specific methods for extracting the three types of relation edges are as follows:

[0071] (1) Extraction of participating relation edges

[0072] Participation relationship edges are used to connect project participant nodes with triggering event nodes, indicating that a project participant has implemented or should have implemented a triggering event. The extraction method is as follows:

[0073] First, a named entity recognition model is used to identify project participants (such as "construction unit," "supervision unit," "design unit," "construction unit," etc.) and action / event descriptions such as "construction not according to drawings," "delayed material inspection," and "insufficient curing time" from standardized multimodal feature data. Then, dependency parsing is used to identify the subject-predicate or agent-action relationships between the participants and actions. Simultaneously, a pre-defined participation relationship trigger word list, including sentence patterns such as "implemented by," "responsible for," "conducted by," and "confirmed by," is used to extract the participation relationships between the participants and events through regular expression matching or semantic role labeling. For implicit participation relationships that cannot be directly matched, prior knowledge of rights and responsibilities in the engineering domain knowledge base is used for reasoning and completion; for example, based on the event of "concrete pouring," the "construction unit" is inferred as the participant.

[0074] (2) Extraction of causal relationship edges

[0075] Causal edges connect causal event nodes to entity state deviation nodes, indicating that a causal event directly or indirectly causes a deviation in the entity's state. The extraction method is as follows:

[0076] First, a dependency parsing-based method is used to identify causal cue words such as "cause," "lead to," "cause," "due to," "because," "cause," and "generate" from the first processed data obtained from textual data (S201 processing). Centering on these causal cue words, dependency parsing is used to locate cause and effect clauses. Candidate causal events are extracted from the cause clauses, and candidate entity state deviations are extracted from the effect clauses. For cases without explicit causal cue words, matching is performed using predefined causal patterns in the engineering quality domain, such as "(material) substandard → (structure) insufficient bearing capacity" and "(curing) inadequate → (concrete) cracks." Furthermore, for structured text descriptions generated during the image data preprocessing stage, such as {"defect type": "cracks," "associated construction node": "3 days after second-floor wall pouring"}, implicit causal relationships such as "insufficient curing time" leading to "cracks" can be directly extracted and fused with the causal chains extracted from the text. Causal relationships from multiple evidence sources are fused using a confidence-weighted method to improve extraction accuracy.

[0077] (III) Extraction of temporal relation edges

[0078] Temporal relationship edges are used to connect nodes of different triggering events, representing the chronological order or temporal dependency of the triggering events. The extraction method is as follows:

[0079] First, a time information extraction model is used to identify time expressions such as "May 10, 2023," "3 days after pouring," and "before the supervisor's acceptance" from textual data, as well as sequence conjunctions such as "after," "then," "before," and "subsequently." A time information extraction model, such as a rule-based time parser or a pre-trained temporal relationship recognition model, is used to convert relative time expressions into absolute time sequences. For structured data such as construction logs and schedule plans in multimodal data, the start / end time fields of the work processes are directly extracted to construct the sequence between processes. For construction node information in image data, based on node types such as "foundation pouring" and "main structure capping," and combined with engineering common sense, their relative temporal positions within the long construction period are determined. For cases lacking explicit time information, standard work process sequences such as "formwork erection → rebar tying → concrete pouring → curing → formwork removal" from the engineering domain knowledge base are used for reasoning and completion. Finally, directed edges are constructed between causal events according to their chronological order, forming temporal relationship edges.

[0080] In this embodiment, the keyword list, causal pattern library, and standard process sequence relied upon by the above extraction method are all pre-stored in an engineering domain knowledge base. The named entity recognition model can employ a deep learning model based on BERT-BiLSTM-CRF, dependency parsing can utilize existing open-source tools such as LTP and Stanford Parser, and temporal information extraction can employ a rule-based temporal parser or a pre-trained temporal relation recognition model. The specific implementation methods described above are conventional techniques in this field, and their internal principles will not be elaborated upon here.

[0081] S400, extract the structured causal feature set of each triggering event based on the topological information of the causal relationship structure.

[0082] Specifically, based on the constructed causal relationship structure, i.e., the causal knowledge graph, structured causal features are extracted for each triggering event node to form a feature set. The extraction process is as follows:

[0083] (1) Topological importance characteristics

[0084] The following three specific metrics are used to quantify the structural importance of triggering events in causal networks:

[0085] Causal Out-degree: The number of causal edges that directly point from the triggering event node to the entity state deviation node. This metric reflects the direct impact of the triggering event on the entity state deviation.

[0086] Causal in-degree: The number of causal edges pointing to the triggering event node, i.e., how many other triggering events or factors led to the occurrence of this triggering event. This indicator reflects the degree to which the triggering event is an intermediate node in the causal chain.

[0087] Causal path betweenness: The frequency with which the triggering event node appears in all causal paths, i.e., the proportion of directed paths from any triggering event to the entity state deviation node. The higher the betweenness, the more critical the triggering event is, and the more likely it is to be located at the intersection of multiple causal paths.

[0088] Among the above metrics, causal out-degree and causal in-degree can be obtained directly from the edge statistics of the graph, while causal path betweenness needs to be calculated by traversing all directed paths from the starting point to the ending point.

[0089] (2) Characteristics of causal paths

[0090] The causal transmission chain properties used to characterize the deviation between the triggering event and the final entity state include the following specific metrics:

[0091] Longest causal distance: The number of edges on the longest directed path from the causal event node to the entity state deviating from the node along the causal relationship edge. The longer the distance, the more complex the causal chain, and the more likely there are multiple levels of propagation.

[0092] Branching factor: The average number of branches that branch off from the triggering event node during causal propagation, which is the logarithmic mean of the number of descendants of the node. The larger the branching factor, the wider the influence of the triggering event.

[0093] Convergence factor: The number of different inducing events that lead to the same entity's state from a deviating node. This metric is used to subsequently assess the cumulative effect of multiple factors.

[0094] (3) Contextual features

[0095] The following specific indicators are used to characterize the relationship between precipitating events, engineering participants, and the time-series environment:

[0096] Number of responsible parties: The number of different project participants connected to the triggering event node through participation relationship edges. This indicator reflects the diversity of responsible parties.

[0097] Temporal position encoding: The sequential index of the triggering event in the linear sequence formed by the temporal relationship edges is normalized to the [0,1] interval according to time sequence. This index is used to quantify the early or late stage of the event.

[0098] Number of preceding triggers: The number of preceding trigger event nodes pointed to by the temporal relationship edges, reflecting whether the event depends on the occurrence of other events.

[0099] The three types of features mentioned above together constitute a structured causal feature set for each triggering event. This feature set is stored in the form of feature vectors, with each triggering event corresponding to a feature vector. Each dimension of the feature vector corresponds to the specific indicators mentioned above. This feature vector is directly used for subsequent calculations and mapping of quantification dimensions, including parameter deviation, state influence magnitude, causal path dependence, and process timing constraints. For example:

[0100] The parameter deviation can be mapped to the ratio of the causal out-degree in the topological importance feature to the canonical threshold in the domain knowledge base. Specifically, the calculation method is as follows: Let the causal out-degree of a certain triggering event be d. out This refers to the number of causal edges originating from this node and pointing to entities whose states deviate from the node. The canonical threshold d for the behavior type corresponding to this triggering event is obtained from the engineering domain knowledge base. std For example, for the causal event of "delayed material inspection," its standard deviation threshold can be set to 1, representing the amount of state deviation directly caused under normal circumstances. The formula for calculating the parameter deviation is min(d out / d std ,1), that is, when the ratio exceeds 1, take 1.

[0101] The magnitude of state influence can be mapped to a weighted combination of the longest causal distance and the branching factor in the causal path features. The specific calculation method is as follows: Let the longest causal distance be Lmax, which is the number of edges on the longest directed path from the causal event node to the entity state it can reach; let the branching factor be B, which is the average number of branches in the causal propagation process starting from this node, i.e., the logarithmic mean of the number of descendant nodes. After normalizing both, they are summed according to preset weights. For example, the normalized longest distance L′ = Lmax / Lref, where Lref is a preset reference distance, for example, 5, limited to [0,1]; the normalized branching factor B′ = min(B / Bref,1), where Bref is a preset reference branching factor, for example, 3. The magnitude of state influence is equal to α•L′ + β•B′, where α and β are preset weights, with default values ​​of α = 0.6 and β = 0.4, which can be adjusted according to the engineering domain knowledge base.

[0102] Causal path dependence can be mapped to causal path betweenness.

[0103] The temporal constraint degree of a process can be mapped to a combination of the temporal position encoding and the number of antecedent factors in the contextual features. The specific calculation method is as follows: Temporal position encoding T pos This is the normalized index value of the triggering event in the linear sequence formed by the temporal relationship edges, with 0 corresponding to the earliest event and 1 corresponding to the latest event. The number of preceding triggers, N. pre This represents the number of preceding causal event nodes that are linked by temporal relationship edges to the current causal event. Normalize the number of preceding causal events: N pre =min(N) pre / N ref ,1), where N ref The preset reference quantity is, for example, 3. The process timing constraint is equal to γ•T. pos +δ•N pre ′, where γ and δ are preset weights, with default values ​​of γ=0.5 and δ=0.5, indicating that the timing and the number of preceding dependencies are equally important.

[0104] S500, based on a preset quantization parameter system, performs weighted calculations on each causal event in the structured causal feature set to obtain the initial weighted score of each causal event.

[0105] In this embodiment, at least some parameters in the quantization parameter system are determined based on the topological information of the causal relationship structure. The preset quantization parameter system includes:

[0106] (a) Topological weight parameters

[0107] Topological weight parameters are used to quantify the structural importance of each triggering event in the causal relationship structure, and are configured individually for each triggering event. They can be determined based on the position of the triggering event node in the causal path and the number of associated edges. Specifically, this includes the following sub-steps:

[0108] Step 1: Confidence Assessment of Causal Path

[0109] A confidence level is assessed for each causal path in the causal relationship structure, the confidence level being calculated based on at least one of the following factors:

[0110] (1) Matching strength of knowledge extraction rules.

[0111] The matching strength of knowledge extraction rules measures the reliability of the knowledge extraction rules used to extract the causal path, including keyword matching, syntactic pattern matching, etc.

[0112] In one illustrative embodiment, the matching strength is a weighted average of the keyword matching score and the syntactic pattern fit.

[0113] The keyword matching score is determined by counting the number of preset keywords matched during the causal path extraction process. These keywords include causal prompts such as "lead to," "cause," and "due to," as well as engineering quality terms such as "maintenance," "crack," and "insufficient strength." The number of matched keywords is divided by the total number of words in the keyword library, resulting in a keyword matching score between 0 and 1. For example, if the keyword library contains 50 words, and 10 of these words are matched during a causal path extraction, the keyword matching score is 0.2.

[0114] The method for determining the syntactic pattern fit is as follows: divide the number of syntactic features actually matched when extracting the causal path by the total number of features required by the preset standard pattern. For example, the standard pattern requires three features: subject (causal event), predicate (causal verb), and object (entity state deviation). If the extraction result fully satisfies these three elements, the syntactic pattern fit is 1; if the subject is missing (only the predicate and object are identified), the fit is two-thirds, approximately 0.67.

[0115] Weighted combination: The two scores above are added together according to preset weights to obtain the matching strength of the knowledge extraction rule. In one example, the weight of the keyword matching score is set to 0.4, and the weight of the syntactic pattern matching score is set to 0.6.

[0116] (2) The number and consistency of multimodal data sources

[0117] The number and consistency of multimodal data sources are used to measure the degree of evidence supporting the same causal path in different modal data such as text, images, and structured data, as well as the logical consistency between descriptions from different sources.

[0118] In one illustrative embodiment, the metric is obtained by multiplying the data source quantity score by the consistency score.

[0119] The method for determining the data source quantity score is as follows: Identify the types of data sources that support the causal path. Data sources are divided into three categories: textual data (such as explicit descriptions in construction logs and supervision reports), image data (such as causal relationships implicit in structured descriptions extracted through visual language models), and structured data (such as time-series relationships reflected in acceptance records). Count the number of data source types with supporting evidence, divide this number by 3, and obtain the data source quantity score. For example, if only textual data supports the causal path, the score is one-third (approximately 0.33); if both text and image data support it, the score is two-thirds (approximately 0.67); and if all three types of data support it, the score is 1.

[0120] The consistency score is determined as follows: Causal parameters extracted from each data source, including the time, location, and severity of the triggering event, as well as the type and location of entity state deviations, are standardized and represented as parameter vectors. For cases with multiple data sources, the cosine similarity between the parameter vectors of every two data sources is calculated, and then the average of all pairwise similarities is taken. If there is only one data source (no other sources for comparison), the consistency score defaults to 1. For example, if the parameter similarity between text and image is 0.9, the similarity between text and structured parameters is 0.8, and the similarity between image and structured parameters is 0.85, then the average of these three (0.85) is taken as the consistency score.

[0121] Combination method: Multiply the data source quantity score by the consistency score to obtain the multimodal support metric. This multiplication method reflects the logic that "a large number of data sources that are consistent with each other" are required to obtain a high score.

[0122] (3) Prior support in the knowledge base of the engineering domain

[0123] Prior support is used to measure the degree to which a causal path matches known causal relationships in existing knowledge, standards, and historical cases in the field of engineering quality.

[0124] In one illustrative embodiment, prior support is a weighted combination of semantic matching score and accuracy score.

[0125] The semantic matching score is determined as follows: The causal path to be evaluated is represented as a triplet consisting of the type of the inducing event, optional intermediate events, and the type of entity state deviation. A word embedding model such as BERT or Word2Vec, pre-trained on an engineering quality domain corpus, is used to convert the descriptions of the inducing events and state deviations into high-dimensional vectors. Then, the cosine similarity of this vector with the vector of each standard causal chain in the engineering domain knowledge base is calculated, and the maximum value is taken as the semantic matching score. For example, if the path to be evaluated is "insufficient maintenance leads to shrinkage cracks," and its semantic similarity with the standard chain "insufficient maintenance → shrinkage cracks" in the knowledge base is 0.92, then the semantic matching score is 0.92.

[0126] The accuracy score is determined by comparing the node sequence of the path to be evaluated with the node sequence of the standard path in the knowledge base. The length of the longest common subsequence (LCS) is calculated, and then divided by the longer of the two sequences. For example, if the standard path has 4 nodes (A→B→C→D) and the path to be evaluated has 3 nodes (A→C→D), the LCS is A→C→D (length 3). Therefore, using the standard path length of 4 as the denominator, the accuracy score is 3 / 4 = 0.75. If the path to be evaluated is longer, then the length of the path to be evaluated is used as the denominator.

[0127] Weighted combination: The semantic matching score and the accuracy score are added together according to preset weights. In one example, the semantic matching score is weighted at 0.6 and the accuracy score is weighted at 0.4.

[0128] After calculating the matching strength of the knowledge extraction rules, the number and consistency of multimodal data sources, and the prior support in the engineering domain knowledge base, the three factors are weighted and summed according to preset weights to obtain the comprehensive confidence of each causal path. In one example, the weights of the three indicators are all set to one-third (equal weight). If an indicator cannot be calculated due to missing data, the weight corresponding to that indicator is reset to zero, and the remaining weights are renormalized proportionally.

[0129] Step 2: Causal Path Pruning

[0130] The causal relationship structure is pruned based on the confidence level, eliminating causal paths with confidence levels below a preset pruning threshold. This preset threshold can be set based on engineering experience (e.g., 0.6) or dynamically adjusted through cross-validation. In specific applications, it can be adjusted according to accuracy requirements: for scenarios with high accuracy requirements, such as forensic identification, the threshold can be increased to 0.7; for general analysis, the threshold can be decreased to 0.4.

[0131] The causal relationship structure after pruning has more reliable topological information, and the topological weight parameters determined based on this structure are more objective and robust.

[0132] Step 3: Determine the topology weight parameters

[0133] Based on the topological information of the pruned causal relationship structure, the topological weight parameters for each triggering event are determined. Specifically, for each triggering event node, at least one of the following indicators for that node in the pruned graph is calculated, and these indicators are weighted and fused to obtain the topological weight parameters:

[0134] The out-degree of a node is the number of causal edges pointing to the state of the entity that deviate from the node.

[0135] Node betweenness centrality refers to the frequency with which a node appears in all causal paths;

[0136] The level depth of a node in a causal chain refers to the number of edges along the longest path from that node to the entity state that deviates from the node.

[0137] The fusion weights of the above indicators can be pre-set by an engineering domain knowledge base, or learned from historical cases through machine learning methods.

[0138] (ii) Environmental weight parameters and gradient correction coefficients

[0139] The environmental weight parameter is used to quantify the baseline intensity of the impact of external environmental factors on project quality, while the gradient correction coefficient is used to dynamically adjust the amplification or weakening effect of environmental factors on the scores of causal events according to the severity of the environmental factors. External environmental factors include at least one of the following: weather factors, force majeure factors, and changes in the construction environment.

[0140] 1. Configuration of environmental weight parameters

[0141] A unified environmental weight parameter is configured for various external environmental factors, such as weather factors, force majeure factors, and changes in the construction environment. This parameter is preset by the engineering knowledge base based on historical data and expert experience.

[0142] 2. Matching rules for gradient correction coefficients

[0143] The gradient correction coefficient is matched based on a preset threshold range where the comprehensive quantitative score of external environmental factors falls, specifically including:

[0144] (1) Calculation of comprehensive quantitative score

[0145] In one illustrative embodiment, the comprehensive quantitative score of external environmental factors is calculated by weighted summation based on the feature values ​​of external environmental factors across multiple preset environmental dimensions. The preset multiple environmental dimensions may include the following three dimensions:

[0146] Dimension 1: Intensity of Factor Influence

[0147] Factor intensity is used to quantify the severity of the impact of external environmental factors on project quality or the degree to which they deviate from the normal range.

[0148] The quantification methods for the intensity of different types of external environmental factors vary, but all follow the following general principle: compare the actual observed value with the preset normal range corresponding to the factor, and convert it into a feature value between 0 and 1 according to a preset mapping rule. The preset normal range, preset warning threshold, and mapping rule are all stored in an engineering domain knowledge base and can be preset according to the engineering type and the characteristics of the environmental factors.

[0149] Rainfall Factors: The basic metric is the rainfall per unit time, such as millimeters per hour or cumulative rainfall (millimeters). The actual rainfall is compared to a preset allowable rainfall threshold for normal construction. If the actual rainfall is below this threshold, the intensity is assigned a preset minimum value, such as 0; if it falls between the preset threshold and the preset warning value, it is linearly interpolated to a preset low-intensity range, such as 0.3 to 0.7; if it exceeds the preset warning value, it is mapped to a preset high-intensity range, such as 0.7 to 1.0. In an illustrative example: if the allowable daily rainfall for normal construction is 10 mm, the actual daily rainfall is 30 mm, and the warning value is 50 mm, then the estimated intensity is approximately 0.5. It should be noted that the above thresholds, warning values, and mapping ranges are examples; in practical applications, they can be preset and adjusted based on the local climate conditions and engineering specifications using an engineering knowledge base.

[0150] Temperature factor: The deviation of the actual temperature from the suitable temperature range required by the construction specifications, such as the lower limit to the upper limit, is used as a measure. The deviation value is calculated by taking the maximum of the following three: the difference between the lower limit of the suitable temperature and the actual temperature, the value 0, and the difference between the actual temperature and the upper limit of the suitable temperature. Specifically, when the actual temperature is lower than the lower limit of the suitable temperature, the former is taken as the difference; otherwise, it is taken as 0. When the actual temperature is higher than the upper limit of the suitable temperature, the latter is taken as the difference; otherwise, it is taken as 0. Dividing the deviation value by the preset maximum allowable deviation value yields an intensity value between 0 and 1. The larger the deviation, the higher the intensity. The upper limit, lower limit, and maximum allowable deviation value of the suitable temperature range are all preset by the engineering knowledge base according to the project type and construction specifications.

[0151] Geological condition variation factors: The relative percentage deviation between actual geological parameters, such as foundation bearing capacity and groundwater level changes, and design parameters is used as a metric. The deviation percentage is calculated as: |actual value - design value| / design value. The deviation percentage is divided by a preset deviation threshold and then limited to the [0,1] interval to obtain the action intensity value. This deviation threshold is preset by an engineering knowledge base based on geological conditions and engineering requirements.

[0152] The characteristic value range of the factor's effect intensity is 0 to 1, with a larger value indicating a stronger effect of the environmental factor.

[0153] Dimension 2: Duration

[0154] Duration is used to quantify the length of time that external environmental factors affect a project.

[0155] In one illustrative embodiment, the characteristic value of duration is calculated as follows:

[0156] First, based on the type of project and the type of environmental factor, a preset baseline duration is obtained from the engineering domain knowledge base. For example, for rainfall factors, the baseline duration can be preset to 1 day; for temperature anomaly factors, the baseline duration can be preset to 3 days. The baseline duration represents the reference duration for which the environmental factor is considered to have "a certain impact" under normal circumstances.

[0157] Secondly, the actual duration is compared with the baseline duration, and the actual duration is converted into a feature value between 0 and 1 according to a preset mapping rule. The mapping rule can be a piecewise linear function, and the specific segmentation method and the corresponding output value range are preset by an engineering domain knowledge base.

[0158] In an illustrative example, the following rules can be set:

[0159] If the actual duration is less than or equal to half of the baseline duration, the feature value is mapped to a lower lower limit, such as 0.1, indicating a short-lived effect.

[0160] If the actual duration is between half and twice the baseline duration, the feature value increases linearly from the lower limit to a higher reference value, such as 0.8.

[0161] If the actual duration exceeds twice the baseline duration, the feature value continues to increase linearly from the reference value, but the upper limit does not exceed 1.0.

[0162] For example, assuming a baseline duration of 1 day and an actual duration of 1.5 days (between 0.5 and 2 days), the eigenvalue obtained by linear interpolation according to the above rules is approximately 0.48. If the actual duration is 3 days (exceeding 2 days), the eigenvalue is approximately 0.9. It should be noted that the baseline duration, segmentation threshold, and mapping output value mentioned above are examples. In practical applications, these parameters can be preset and adjusted based on the project type, environmental characteristics, and historical data statistics using an engineering domain knowledge base.

[0163] As a simplified implementation, the actual duration can be divided by a preset maximum reference duration, such as 7 days, which can be set according to the project type and specifications. The ratio is then limited to the [0,1] range to obtain a characteristic value of the duration. This maximum reference duration is also stored in the engineering domain knowledge base and can be configured as needed. The characteristic value of the duration ranges from 0 to 1; a larger value indicates a longer impact period.

[0164] Dimension 3: Scope of Impact on the Project

[0165] The scope of impact on the project is used to quantify the proportion or level of the project area affected by external environmental factors within the entire project.

[0166] In one illustrative embodiment, the characteristic values ​​affecting the scope of the project are calculated as follows:

[0167] First, based on the location information of the engineering quality problems, such as the affected area recorded in the construction log, the on-site description in the supervision report, and the defect distribution identified in the image data, estimate the percentage of the affected area relative to the total area or total quantity of the entire project. If a precise percentage cannot be obtained, a graded quantification method can be used to divide the affected area into multiple preset levels.

[0168] The estimated percentage or level is converted into a feature value between 0 and 1 according to a preset mapping rule. The mapping rule and its corresponding output value range are preset by an engineering domain knowledge base.

[0169] In an illustrative example, the following hierarchy and mapping rules can be defined:

[0170] Local effects, such as those affecting a single component or an area smaller than a preset area threshold, are mapped to a lower lower limit feature value, such as 0.2.

[0171] Regional impact, such as a construction section or multiple adjacent components: mapped to a moderate eigenvalue, such as 0.5;

[0172] Large-area impact, such as an entire floor or multiple construction sections: mapped to a higher characteristic value, such as 0.8;

[0173] Overall impact on the entire project or most of the area: mapped to the maximum value, such as 1.0.

[0174] For example, if a heavy rainfall causes water accumulation in a foundation pit, affecting only a localized area at the bottom of the pit, it can be classified as a localized impact, with an impact range characteristic value of 0.2. It should be noted that the above classification and its corresponding mapping values ​​are examples. In practical applications, these parameters can be preset and adjusted based on the project type, specification requirements, and historical experience using an engineering knowledge base.

[0175] As another approach, the estimated percentage can be directly used as a characteristic value affecting the scope of the project. This characteristic value ranges from 0% to 100%, corresponding to the normalized interval [0, 1]. When the percentage cannot be obtained precisely, the aforementioned graded quantification method is used.

[0176] The characteristic value that affects the scope of the project ranges from 0 to 1, with a larger value indicating a wider affected area.

[0177] The feature values ​​calculated for the three dimensions above, each between 0 and 1, are weighted and summed according to preset weights to obtain a comprehensive quantitative score of external environmental factors.

[0178] In one illustrative embodiment, the three dimensions are weighted equally, each accounting for one-third. Therefore, the comprehensive quantitative score is equal to (intensity feature value + duration feature value + scope of influence feature value) / 3.

[0179] In other embodiments, the weights can be dynamically adjusted based on the type of project or the characteristics of environmental factors. For example, for instantaneous heavy rainfall, the weight of the intensity can be appropriately increased; for persistent low temperatures, the weight of the duration can be appropriately increased. The specific weight values ​​for each dimension are stored in an engineering domain knowledge base and can be configured by the user according to the actual situation.

[0180] The comprehensive quantitative score ranges from 0 to 1, with a higher score indicating a more severe adverse impact of external environmental factors on project quality.

[0181] (2) Matching the corresponding correction coefficient

[0182] The comprehensive quantitative score of external environmental factors is compared with a preset first threshold and second threshold, and a corresponding gradient correction coefficient is matched according to the threshold range in which the score falls. The first threshold, the second threshold, and the gradient correction coefficients at each level are all collaboratively calibrated using an expert knowledge base in the field of engineering quality in conjunction with current engineering construction standards.

[0183] The matching rules are as follows:

[0184] When the overall quantitative score is below the first threshold, the first gradient correction coefficient is applied. This coefficient corresponds to the case of weak influence and is usually set to a baseline value, such as 1.0, meaning that no additional correction is made for external environmental factors.

[0185] When the comprehensive quantitative score falls between the first and second thresholds, a second gradient correction coefficient is applied. This coefficient corresponds to the moderate impact scenario, and its value is greater than the first gradient correction coefficient and less than the third gradient correction coefficient; for example, it can be set to a preset value between 1.2 and 1.5.

[0186] When the comprehensive quantitative score is higher than or equal to the second threshold, a third gradient correction coefficient is applied. This coefficient corresponds to the case of strong influence and its value is greater than the second gradient correction coefficient; for example, it can be set to a preset value between 1.5 and 2.0.

[0187] It should be noted that the specific values ​​mentioned above, such as 1.0, 1.2 to 1.5, and 1.5 to 2.0, are merely illustrative examples and do not constitute a limitation on the scope of protection. In practical applications, the specific values ​​of the first threshold, the second threshold, and each gradient correction coefficient are determined based on factors such as the type of project, the characteristics of environmental factors, and regional climate conditions, using an engineering knowledge base combined with current engineering construction standards. For example, the classification provisions regarding the impact of environmental conditions on construction quality in the "Unified Standard for Acceptance of Construction Quality" can be referenced, and the specific values ​​can be determined using the Delphi method or the analytic hierarchy process, taking into account the scoring results of multiple engineering quality appraisal experts. The calibrated parameters are used to quantify the amplifying or amplifying effect of natural conditions (such as extreme weather, unpredictable geological changes, etc.) on quality defects.

[0188] (iii) Global Coupling Correction Coefficient

[0189] The global coupling correction coefficient is used to correct the weighted result of each causal event when there is a coupling reinforcement effect between two or more causal events, so as to reflect the nonlinear enhancement effect generated by the superposition of multiple factors.

[0190] The activation condition for the global coupling correction coefficient is automatically determined based on the topological characteristics of the causal relationship structure, requiring no manual intervention. The specific determination rules are as follows:

[0191] Rule 1 (Multiple Causes, Multiple Effects): When there are two or more causal event nodes in the causal relationship structure that point to the same entity state deviation node (i.e., multiple causal relationship edges converge to the same quality defect node), the coupling reinforcement effect is determined to be satisfied.

[0192] Rule 2 (Temporal Dependency Common Cause): When there are two or more causal event nodes that have temporal dependencies (i.e., directly or indirectly connected through temporal relationship edges), and these causal event nodes all have causal paths with the same entity state deviation node, the coupling reinforcement effect is determined to be satisfied.

[0193] When any of the above rules are met, the global coupling correction coefficient is enabled.

[0194] The global coupling correction coefficient C is typically greater than 1, indicating that the coupling reinforcement effect amplifies the overall effect when each triggering event acts independently. Its specific value is determined based on the mechanism of multiple risk superposition and the experience of experts in the field of engineering quality. For example, it can be determined by the ratio of the measured value of the severity of quality defects when multiple factors act simultaneously to the theoretical superposition value of a single factor in historical cases. When specific calibration data is lacking, C=1.2 can be preset as the default value, and subsequently updated dynamically through an expert consultation mechanism.

[0195] In the quantitative attribution calculation process, if the coupling reinforcement effect is satisfied, the weighted scores of each relevant triggering event are multiplied by the global coupling correction coefficient C to obtain the final comprehensive score, wherein the weighted scores have been adjusted for the environment; if the effect is not satisfied, then C=1, that is, no correction is performed.

[0196] Furthermore, the S500 specifically includes:

[0197] S501, extract the feature values ​​of each triggering event on multiple preset quantitative dimensions from the structured causal feature set.

[0198] In the structured causal feature set, each triggering event corresponds to a set of quantifiable numerical features, which have already been extracted in S400. This step selects the corresponding feature values ​​from this set according to a preset quantification dimension system. In one illustrative embodiment, the quantification dimension includes the following four items:

[0199] Parameter deviation is used to characterize the degree of deviation of the parameters involved in the causative event from engineering specifications or design thresholds;

[0200] State impact magnitude is used to characterize the magnitude of the impact of the entity state deviation caused by the triggering event on the overall performance of the project.

[0201] Causal path dependence is used to characterize the strength of the direct association between the triggering event and the deviation of the entity's state along the causal path;

[0202] The process sequence constraint degree is used to characterize the criticality of the triggering event in the construction process sequence and the degree of constraint and influence on other processes.

[0203] The feature values ​​of the above four dimensions have all been calculated in S400, and can be read directly in this step.

[0204] S502, calculate the inherent impact value of each triggering event based on the characteristic values ​​of each triggering event in multiple preset quantitative dimensions.

[0205] The inherent impact value is used to quantify the basic influence of a triggering event, determined by its own attributes, without considering external environmental factors and coupling effects. In an illustrative embodiment, the inherent impact value is calculated by taking the arithmetic mean of the characteristic values ​​of the four dimensions mentioned above. For example, if the parameter deviation of a triggering event is 0.8, the state influence amplitude is 0.6, the causal path dependence is 0.7, and the process sequence constraint is 0.5, then its inherent impact value is the sum of the four items divided by 4, i.e., (0.8 + 0.6 + 0.7 + 0.5) / 4 = 0.65. In other embodiments, different weighting coefficients can be set according to the importance of each dimension for weighted averaging, and the weighting coefficients are preset by the engineering domain knowledge base.

[0206] S503, multiply the inherent impact value of each triggering event by the corresponding topological weight parameter to obtain the initial weighted score.

[0207] The topological weight parameters have been determined in S400 based on the topological information of the causal relationship structure. Multiplying the inherent influence value by the topological weight parameters yields the initial weighted score of the triggering event. This initial weighted score reflects the combined result of the structural importance of the triggering event within the causal relationship structure and its inherent influence.

[0208] S600 quantifies and corrects the initial weighted scores of each triggering event based on external environmental factors to obtain the final score of each triggering event.

[0209] Specifically, the quantization correction includes the following sub-steps:

[0210] S601, calculate the comprehensive quantitative score of each external environmental factor based on the characteristic values ​​of each external environmental factor in multiple preset environmental dimensions.

[0211] Calculate the comprehensive quantitative score of each external environmental factor according to the method described in S400.

[0212] S602, match gradient correction coefficients and perform environmental correction on the initial weighted scores of each triggering event.

[0213] Based on the comprehensive quantitative score obtained in S601, the corresponding gradient correction coefficient is determined according to the matching rules described in S400. The initial weighted score of each triggering event is multiplied by this gradient correction coefficient to obtain the preliminary corrected score. It should be noted that this gradient correction coefficient is applied uniformly to all triggering events; that is, under the same external environmental conditions, the environmental amplification or attenuation ratio of each triggering event is the same.

[0214] S603, when there is a coupling reinforcement effect between two or more causal events, the preliminary correction score is multiplied by the global coupling correction coefficient to obtain the final score of each causal event.

[0215] According to the coupling reinforcement effect determination rules described in S400, determine whether there is a coupling reinforcement effect between two or more causal events:

[0216] When it is determined that there is a coupling reinforcement effect between two or more causal events, the preliminary correction score of each causal event obtained in S602 is multiplied by the global coupling correction coefficient to obtain the final score of each causal event.

[0217] When it is determined that there is no coupling reinforcement effect between two or more triggering events, the global coupling correction coefficient is set to 1, and the preliminary correction score of each triggering event is the final score.

[0218] After the above environmental and coupling corrections, each triggering event corresponds to a final score, which comprehensively reflects the intrinsic influence of the triggering event, the amplification effect of the external environment, and the nonlinear effect of multiple factors superimposed.

[0219] S700 outputs the percentage of causal effect values ​​for each project participant and external environmental factor based on the final scores of each inducing event.

[0220] Specifically, it includes the following sub-steps:

[0221] S701, calculate the percentage of causal effect values ​​for each inducing event and each external environmental factor.

[0222] First, the final score for each external environmental factor is calculated. The final score for each external environmental factor is equal to its comprehensive quantitative score multiplied by its corresponding environmental weight parameter, which can be preset by the engineering domain knowledge base. For the triggering event, its final score has already been calculated in S600.

[0223] Then, the final scores of the triggering events and external environmental factors are globally normalized: the final scores of all triggering events are added to the final scores of all external environmental factors to obtain a total score. For each triggering event, its final score is divided by this total score to obtain the causal effect percentage of that triggering event; for each external environmental factor, its final score is divided by this total score to obtain the causal effect percentage of that external environmental factor. The sum of all causal effect percentages is 1.

[0224] S702, sums up the results according to the affiliation relationship and outputs the results.

[0225] Based on the attribution relationship between the project participants and the causal events, the causal effect value proportion of each causal event is accumulated and then added to the corresponding project participant to obtain the causal effect value proportion of each project participant. The attribution relationship is defined in the causal relationship structure.

[0226] Finally, the output includes the percentage of causal effects for each project participant and the percentage of causal effects for each external environmental factor. This output can be presented in list, chart, or structured data format for quantitative attribution analysis of the causal forces affecting project quality.

[0227] Furthermore, after outputting the causal effect percentages of each project participant and external environmental factor, the process also includes a dynamic iterative update of the weight parameters, as detailed below:

[0228] (a) Detection and error reporting of abnormal weight combinations

[0229] An online expert feedback calibration mechanism is established. This mechanism continuously monitors the weight parameter combinations in the quantification parameter system, including topological weight parameters, environmental weight parameters, gradient correction coefficients, and global coupling correction coefficients. When a significant deviation is detected between the current weight combination and the typical weight distribution pattern stored in the engineering domain knowledge base, or when the weight combination causes an abnormal distribution in the proportion of causal effect values ​​in the output, the weight combination is judged as an abnormal weight combination, and an error is triggered. A significant deviation refers to the difference between the current weight parameter and the typical weight value stored in the engineering domain knowledge base exceeding a preset allowable range, such as an Euclidean distance greater than 0.3, or any weight parameter exceeding a preset reasonable range. An abnormal distribution refers to, for example, the proportion of a certain responsible entity exceeding a preset reasonable upper limit, such as greater than 0.8, or the variance of the proportions of each entity exceeding a preset threshold, such as greater than 0.25.

[0230] (ii) Obtaining feedback from online experts

[0231] The identified abnormal weight combinations and their corresponding engineering quality case information, including basic case characteristics and multimodal data statistical summaries, are sent to a pre-defined online expert terminal. This online expert terminal is a remote terminal connected to the computer device on which the method described in this embodiment runs, and is used by registered engineering quality appraisal experts or engineering law experts. Upon receiving the abnormal information, the experts, based on their professional knowledge and experience, verify and analyze the weight parameters and provide feedback, including adjusted weight values ​​or adjustment suggestions. The system waits for the expert feedback, and a timeout mechanism, such as 48 hours, can be set; after the timeout, the default processing rules are automatically applied.

[0232] (III) Dynamic updating of weight parameters and iteration of knowledge base

[0233] Based on expert feedback, the relevant weight parameters are updated. The updated weight parameters are then verified and stored in the engineering domain knowledge base, replacing the original parameter values. Simultaneously, the abnormal weight combination, expert feedback, and updated parameters are stored as a new experience record in the knowledge base for comparison and reference in subsequent anomaly detection and for optimizing weight calibration rules. Through this process, the weight calibration rules are dynamically and iteratively updated, enabling the quantitative parameter system to continuously optimize as expert experience accumulates.

[0234] In summary, this invention addresses the problems in existing engineering quality causal force analysis, such as fragmented multimodal information, the inability of statistical correlation to distinguish between causation and correlation, and the reliance on subjective expert weighting for weight parameters. It proposes a quantitative attribution architecture that integrates multimodal data-driven approaches with a causal knowledge graph. Its design logic includes: unifying heterogeneous data into computable feature representations through multimodal fusion preprocessing; constructing a causal knowledge graph with "participating subject—causal event—entity state deviation" as elements, making implicit causal relationships explicit as a traversable topological structure; using the graph's topological features, including node positions, edge density, and path length, as the basis for calculating causal event weights, thus reducing the reliance on subjective expert weighting for weight generation; and incorporating external environmental factors as independent nodes into the quantitative system, quantifying the interference of confounding factors through gradient correction coefficients, and characterizing the nonlinear effects of multiple superimposed causal factors through coupling correction coefficients. Therefore, this invention forms a technical path from data input to causal modeling, weighted self-driving, confounding elimination, and normalized attribution, extending the analysis of causal forces in engineering quality from correlation speculation to causal quantification, and providing an implementable technical solution for multimodal data utilization, causal relationship differentiation, and subjective weighting substitution.

[0235] To further improve the accuracy and adaptability of causal force quantification, this embodiment also includes the following steps:

[0236] S800 uses graph attention networks to adaptively optimize topological weight parameters and global coupling correction coefficients.

[0237] Specifically, the steps include:

[0238] S801, calculate the attention coefficient between the causal event nodes.

[0239] The structured causal feature vectors of each triggering event node extracted by S400 are used as input to the graph attention network. Based on the causal knowledge graph, the multi-head attention mechanism of the graph attention network is used to calculate the attention coefficient for each pair of triggering event nodes with a connection. This coefficient quantifies the influence strength of the source node triggering event on the target node triggering event. The specific calculation method of the attention coefficient adopts the well-known multi-head attention mechanism in graph attention networks, and its implementation method is well known to those skilled in the art and will not be elaborated here. The attention coefficients between all calculated node pairs constitute an attention coefficient matrix. This matrix is ​​automatically generated by the graph attention network during the forward propagation process based on the input graph structure and node features, and is not given externally in advance. The rows and columns of the matrix correspond to the triggering event nodes, and each element in the matrix represents the attention weight between the corresponding node pairs.

[0240] S802 generates a deep embedded representation of the node.

[0241] For each triggering event node, the feature vectors of all nodes in its neighborhood are weighted and aggregated according to the attention coefficient, and then propagated through multi-layer graph attention to obtain a deep embedding representation of the node. This embedding representation integrates the node's original features and the interaction information passed from neighboring nodes through attention weights.

[0242] S803 generates adaptive topology weights.

[0243] The deep embedding representations of each triggering event node are input into the fully connected layer and mapped to adaptive topological weight parameters. These adaptive topological weight parameters can replace the preset topological weights calculated based on graph topology in S500, or they can be weighted and fused with preset weights, with the fusion coefficients preset by the engineering domain knowledge base.

[0244] S804 generates dynamic global coupling correction coefficients.

[0245] The attention coefficient matrix obtained from S801 is input into the global coupling effect estimation sub-network. This sub-network extracts the coupling features of the entire causal knowledge graph through global pooling operations such as average pooling or max pooling, and outputs a dynamic global coupling correction coefficient. This coefficient is used to replace or weight and fuse the preset global coupling correction coefficient in S600, thereby dynamically adjusting the multi-factor superposition effect according to the actual coupling mode of the input data. The above graph attention network needs to be pre-trained. The training dataset uses historical engineering quality cases, each case containing multimodal engineering data and the corresponding ground truth of responsibility proportions, such as the proportion of causal force confirmed by judicial judgment or expert consensus. The training loss function uses mean squared error, that is, the mean squared error between the network's predicted causal force proportion vector and the true responsibility proportion vector. The training process aims to minimize this loss and iteratively update the network parameters.

[0246] After training, during the inference phase, the network can automatically generate adaptive topological weights and dynamic global coupling correction coefficients based on the input multimodal engineering data, significantly reducing the reliance on manually calibrated weights.

[0247] In this embodiment, a graph attention network is introduced to automatically learn the attention weights between causal event nodes from a causal knowledge graph. Based on this, adaptive topological weights and dynamic global coupling correction coefficients are generated, reducing reliance on manually calibrated parameters. This allows the weights to adapt to different case data, more accurately capturing multi-causal nonlinear superposition effects. Furthermore, end-to-end training using historical cases can continuously optimize attribution accuracy, further improving the objectivity, adaptability, and predictive accuracy of quantitative analysis of engineering quality causal forces. This adaptive optimization mechanism works in conjunction with an online expert feedback calibration mechanism: the S800, as a routine automatic weight generation module, is executed first during the inference phase; when the generated weight parameters trigger anomaly judgment conditions, the online expert feedback calibration mechanism is activated for manual verification analysis. The expert feedback results can be used for subsequent iterative training of the graph attention network, thus forming a dynamic parameter optimization closed loop with automatic learning as the primary method and expert verification as a secondary method.

[0248] Second Embodiment

[0249] This embodiment provides a quantitative analysis system for engineering quality causal factors, used in the method of the first embodiment, such as... Figure 2 As shown, it includes:

[0250] The multimodal data acquisition module 1 is used to acquire multimodal engineering data, which includes text data, image data and structured data; its specific implementation is the same as step S100 in the first embodiment, and will not be repeated here.

[0251] The multimodal preprocessing module 2 is used to classify and preprocess the multimodal engineering data to obtain standardized multimodal feature data; its specific implementation method is the same as step S200 in the first embodiment, and will not be described again here.

[0252] The causal relationship construction module 3 is used to perform semantic parsing on the standardized multimodal feature data based on the engineering domain knowledge base, and to construct a causal relationship structure to characterize the causal relationship between engineering participants, causal events and entity state deviations; its specific implementation method is the same as step S300 of the first embodiment, and will not be repeated here.

[0253] The causal feature extraction module 4 determines the topological weight parameters of each causal event based on the topological structure of the causal relationship information, and extracts the structured causal feature set of each causal event; its specific implementation method is the same as step S400 of the first embodiment, and will not be repeated here.

[0254] The first calculation module 5 is used to perform weighted calculations on each causal event in the structured causal feature set based on a preset quantification parameter system to obtain an initial weighted score for each causal event; wherein, at least some parameters in the quantification parameter system are determined according to the topological information of the causal relationship structure; its specific implementation method is the same as step S500 of the first embodiment, and will not be repeated here.

[0255] The second calculation module 6 is used to quantify and correct the initial weighted scores of each triggering event based on external environmental factors, so as to obtain the final scores of each triggering event; its specific implementation method is the same as step S600 of the first embodiment, and will not be described again here.

[0256] Attribution module 7 is used to output the proportion of causal effect values ​​of each project participant and external environmental factor based on the final scores of each inducing event, as the quantitative analysis result of the causal forces of project quality. Its specific implementation method is the same as step S700 in the first embodiment, and will not be repeated here.

[0257] Optionally, the system may further include an adaptive optimization module for adaptively optimizing the topological weight parameters and global coupling correction coefficients based on the graph attention network, and its specific implementation is the same as step S800 in the first embodiment.

[0258] The above modules are connected in sequence to form a complete data processing chain, which together realizes the objective quantitative attribution of the causes of engineering quality.

[0259] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in this invention.

[0260] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0261] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0262] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A quantitative analysis method for the causes of engineering quality problems, characterized in that, The method includes the following steps: Acquire multimodal engineering data, which includes text data, image data, and structured data; The multimodal engineering data is classified and preprocessed to obtain standardized multimodal feature data; Based on an engineering domain knowledge base, the standardized multimodal feature data is semantically parsed to construct a causal relationship structure that characterizes the causal associations between engineering participants, triggering events, and deviations from entity states; the causal relationship structure is a causal knowledge graph. The node types of the causal knowledge graph include project participant nodes, causal event nodes, and entity state deviation nodes. The edge types of the causal knowledge graph include: participation relationship edges between project participant nodes and causal events, causal relationship edges between causal event nodes and entity state deviation nodes, and temporal relationship edges between causal event nodes. Based on the topological information of the causal relationship structure, extract the set of structured causal features for each triggering event; Based on a preset quantification parameter system, each triggering event in the structured causal feature set is weighted and calculated to obtain an initial weighted score for each triggering event; wherein, at least some parameters in the quantification parameter system are determined based on the topological information of the causal relationship structure. The initial weighted scores of each triggering event are quantitatively corrected based on external environmental factors to obtain the final scores of each triggering event. Based on the final scores of each inducing event, the causal effect percentages of each project participant and external environmental factor are output as the quantitative analysis results of the causal forces affecting project quality. The preset quantization parameter system includes: The topological weight parameters configured for each triggering event are determined based on the position of the triggering event node in the causal path and the number of associated edges in the causal relationship structure. The environmental weight parameters and gradient correction coefficients configured for external environmental factors are matched according to the preset threshold range in which the comprehensive quantitative score of the external environmental factors falls. The global coupling correction coefficient is used to correct the weighted result of each triggering event when there is a coupling reinforcement effect between two or more triggering events. Based on a preset quantification parameter system, each triggering event in the structured causal feature set is weighted and calculated to obtain an initial weighted score for each triggering event, specifically including: Extract feature values ​​of each triggering event across multiple preset quantization dimensions from the structured causal feature set; The inherent impact value of each triggering event is calculated based on the characteristic values ​​of each triggering event across multiple preset quantitative dimensions. The initial weighted score is obtained by multiplying the inherent impact value of each triggering event by the corresponding topological weight parameter. The initial weighted scores of each triggering event are quantitatively adjusted based on external environmental factors to obtain the final score for each triggering event, specifically including: The comprehensive quantitative score of each external environmental factor is calculated based on the characteristic values ​​of each factor across multiple preset environmental dimensions. Based on the comprehensive quantitative score of external environmental factors, the corresponding gradient correction coefficient is matched, and the initial weighted score of each triggering event is multiplied by the gradient correction coefficient to obtain the preliminary correction score. When there is a coupling reinforcement effect between two or more triggering events, the preliminary correction score is multiplied by the global coupling correction coefficient to obtain the final score of each triggering event.

2. The method according to claim 1, characterized in that, The multimodal engineering data is classified and preprocessed to obtain standardized multimodal feature data, including: The text data is subjected to text normalization processing, and word segmentation and semantic fragment segmentation are performed using a large language model to obtain the first processed data; The image data is processed by extracting defect area features and engineering construction node information using a multimodal visual language model, and then converted into structured text descriptions to obtain the second processed data. The structured data is numerically normalized and missing values ​​are filled in. A large language model is used to extract the semantic description of the table to obtain the third-processed data. The first processed data, the second processed data, and the third processed data are fused using multimodal features to obtain the standardized multimodal feature data.

3. The method according to claim 1, characterized in that, Based on the final scores of each inducing event, the causal effect percentages of each project participant and external environmental factor are output, specifically including: The final scores of each triggering event and each external environmental factor are globally normalized, and the causal effect value ratio of each triggering event and each external environmental factor is calculated separately. The final score of each external environmental factor is obtained by multiplying the comprehensive quantitative score of each external environmental factor by the corresponding environmental weight parameter. Based on the attribution relationship between the project participants and the causal events, the causal effect value of each causal event is accumulated and attributed to the corresponding project participant, and the causal effect value of each project participant and the causal effect value of each external environmental factor are output.

4. The method according to claim 1, characterized in that, After outputting the causal effect percentages of each project participant and external environmental factor, the following is also included: Establish an online expert feedback and calibration mechanism to report abnormal weight combinations and send them to the online expert terminal to obtain feedback results; Based on the feedback results, the topology weight parameters, environment weight parameters, gradient correction coefficients, and global coupling correction coefficients are updated, and the updated parameters are stored in the engineering domain knowledge base to achieve dynamic iterative optimization of the weight parameters.

5. The method according to claim 1, characterized in that, Also includes: The topology weight parameters and the global coupling correction coefficients are adaptively optimized based on a graph attention network. The adaptive optimization includes: Using the structured causal feature set as input to the graph attention network, the attention coefficients between each causal event node are calculated to generate an attention coefficient matrix; For each triggering event node, the feature vectors of all nodes in the neighborhood of the triggering event node are weighted and aggregated according to the attention coefficient, and a deep embedding representation of the triggering event node is obtained through multi-layer graph attention propagation. The deep embedding representation of the input fully connected layer is mapped to adaptive topological weight parameters, which are used to replace or weighted fuse the topological weight parameters. The attention coefficient matrix is ​​input into the global coupling effect estimation subnetwork, and global coupling features are extracted through global pooling operation. Dynamic global coupling correction coefficients are output, which are used to replace or weighted fuse the global coupling correction coefficients.

6. A quantitative analysis system for engineering quality causal factors, characterized in that, include: A multimodal data acquisition module is used to acquire multimodal engineering data, which includes text data, image data, and structured data. A multimodal preprocessing module is used to perform classification preprocessing on the multimodal engineering data to obtain standardized multimodal feature data; The causal relationship construction module is used to perform semantic parsing on the standardized multimodal feature data based on the engineering domain knowledge base, and construct a causal relationship structure to characterize the causal association between engineering participants, triggering events and entity state deviations; the causal relationship structure is a causal knowledge graph. The node types of the causal knowledge graph include project participant nodes, causal event nodes, and entity state deviation nodes. The edge types of the causal knowledge graph include: participation relationship edges between project participant nodes and causal events, causal relationship edges between causal event nodes and entity state deviation nodes, and temporal relationship edges between causal event nodes. The causal feature extraction module extracts a set of structured causal features for each triggering event based on the topological information of the causal relationship structure. The first calculation module is used to perform weighted calculations on each causal event in the structured causal feature set based on a preset quantification parameter system to obtain an initial weighted score for each causal event; wherein, at least some parameters in the quantification parameter system are determined based on the topological information of the causal relationship structure. The second calculation module is used to quantify and correct the initial weighted scores of each triggering event based on external environmental factors, so as to obtain the final score of each triggering event. The attribution module is used to output the proportion of causal effect values ​​of each project participant and external environmental factor based on the final score of each triggering event, as the quantitative analysis result of the cause force of project quality. The preset quantization parameter system includes: The topological weight parameters configured for each triggering event are determined based on the position of the triggering event node in the causal path and the number of associated edges in the causal relationship structure. The environmental weight parameters and gradient correction coefficients configured for external environmental factors are matched according to the preset threshold range in which the comprehensive quantitative score of the external environmental factors falls. The global coupling correction coefficient is used to correct the weighted result of each triggering event when there is a coupling reinforcement effect between two or more triggering events. The first calculation module is specifically used for: Extract feature values ​​of each triggering event across multiple preset quantization dimensions from the structured causal feature set; The inherent impact value of each triggering event is calculated based on the characteristic values ​​of each triggering event across multiple preset quantitative dimensions. The initial weighted score is obtained by multiplying the inherent impact value of each triggering event by the corresponding topological weight parameter. The second calculation module is specifically used for: The comprehensive quantitative score of each external environmental factor is calculated based on the characteristic values ​​of each factor across multiple preset environmental dimensions. Based on the comprehensive quantitative score of external environmental factors, the corresponding gradient correction coefficient is matched, and the initial weighted score of each triggering event is multiplied by the gradient correction coefficient to obtain the preliminary correction score. When there is a coupling reinforcement effect between two or more triggering events, the preliminary correction score is multiplied by the global coupling correction coefficient to obtain the final score of each triggering event.

Images