Method and apparatus, device, medium for storing building project knowledge

Abstract

The application is applied to the technical field of building project knowledge digitization, and discloses a method and device, equipment and medium for storing building project knowledge. The method comprises the following steps: after the start of a building project, obtaining project raw data to be sorted every day; generating a first question and answer matching pair corresponding to the project raw data; at intervals of a first time period, correcting the knowledge of the first question and answer matching pair generated in the first time period to obtain a second question and answer matching pair; at intervals of a second time period, screening the causal credibility of the second question and answer matching pair generated in the second time period to obtain a third question and answer matching pair; and after the completion of the building project, generating structured knowledge data corresponding to the building project based on all the third question and answer matching pairs of the building project. In this way, through knowledge correction and causal credibility screening, the quality and reliability of the finally generated structured knowledge data are improved.

Description

Technical Field

[0001] This invention relates to the field of digital knowledge technology for construction projects, and specifically to a method, apparatus, equipment, and medium for storing knowledge of construction projects. Background Technology

[0002] Current construction projects are characterized by long cycles, numerous participants, complex on-site conditions, and diverse data types. The entire project lifecycle generates massive amounts of raw data, including drawings, construction logs, change orders, hazard rectification records, acceptance documents, and technical briefings. The industry commonly uses document folders, simple spreadsheets, or scattered online document libraries to store this engineering data. This storage method only enables the archiving and retention of relevant project documents and is insufficient for forming a reusable, structured engineering knowledge system.

[0003] In the digital transformation of the construction industry, the application of large language models has become a trend. Existing large architectural models generally adopt a passive mode of "fine-tuning of a general model + feeding into a static knowledge base + training with manual annotation." While this mode can achieve basic specification retrieval and text generation, manual annotation is costly and of inconsistent quality. Furthermore, existing large architectural models only establish statistical mappings based on the frequency and distribution patterns of words and sentences in massive amounts of text. This means they can only discover the correlation between two events, but not the causal relationship, which easily leads to false associations and causal inversions. Consequently, the quality and reliability of the final knowledge generated are poor.

[0004] Therefore, in order to overcome the above-mentioned technical problems, the present invention provides a method, apparatus, equipment, and medium for storing knowledge of building projects. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to improve the quality of the knowledge formed. The purpose is to provide a method, apparatus, equipment, and medium for storing knowledge of construction projects, so as to improve the quality of the knowledge formed.

[0006] This invention is achieved through the following technical solution:

[0007] In a first aspect, a method for storing knowledge about a construction project is provided, the method comprising: acquiring raw project data to be processed daily after the construction project is initiated; generating a first question-and-answer matching pair corresponding to the raw project data; performing knowledge correction on the first question-and-answer matching pair generated within the first time period at a first time interval to obtain a second question-and-answer matching pair; performing causal credibility filtering on the second question-and-answer matching pair generated within the second time period at a second time interval to obtain a third question-and-answer matching pair; and generating structured knowledge data corresponding to the construction project based on all the third question-and-answer matching pairs of the construction project after the construction project is completed.

[0008] In some embodiments, generating the first question-and-answer matching pair corresponding to the original project data includes: extracting question text and answer text corresponding to the question text from the original project data; and generating the first question-and-answer matching pair based on the question text and the answer text.

[0009] In some embodiments, the first question-and-answer matching pair is stored in a preset transient memory database; the transient memory database is cleared every first time interval; the step of correcting the knowledge of the first question-and-answer matching pair generated within the first time interval to obtain the second question-and-answer matching pair includes: inputting all the first question-and-answer matching pairs stored in the transient memory database into a preset building hybrid expert model to obtain the description text of the first question-and-answer matching pair; using a preset building engineering causal directed acyclic graph to determine whether there is a conflict between the description texts; if there is a conflict between the description texts, correcting the first question-and-answer matching pair corresponding to the conflicting description text to obtain the second question-and-answer matching pair; and / or, if there is no conflict between the description texts, determining the first question-and-answer matching pair corresponding to the non-conflicting description text as the second question-and-answer matching pair.

[0010] In some embodiments, the building hybrid expert model includes a building backbone network, a gating routing module, and several pre-defined sub-expert networks corresponding to different building engineering specialties. The building backbone network is used to obtain the subdivided domain feature vectors corresponding to the first question-and-answer matching pairs and to obtain the descriptive text corresponding to the first question-and-answer matching pairs with the descriptive text feature vectors. The gating routing module calculates the activation probability of each sub-expert network based on the subdivided domain feature vectors and selects several sub-expert networks to be activated based on the activation probabilities. The sub-expert networks to be activated are used to receive the subdivided domain feature vectors and output the professional domain description vectors corresponding to the building engineering specialties. Based on the activation probabilities output by the gating routing module, the subdivided domain feature vectors output by all the sub-expert networks to be activated are weighted and fused to obtain the descriptive text feature vectors.

[0011] In some embodiments, the method for storing architectural project knowledge further includes: fine-tuning the architectural hybrid expert model at third time intervals using all third question-and-answer matching pairs.

[0012] In some embodiments, the second question-and-answer matching pairs are stored in a preset short-term memory database; the short-term memory database is cleared every second time interval; the step of filtering the second question-and-answer matching pairs generated within the second time interval based on causal credibility to obtain a third question-and-answer matching pair includes: for all second question-and-answer matching pairs stored in the short-term memory database, determining the parameters corresponding to each layer node in a preset multi-layer Bayesian network based on the second question-and-answer matching pair, and obtaining the Bayesian parameter sequence corresponding to the second question-and-answer matching pair; the multi-layer Bayesian network includes: a hazard type layer, a risk factor layer, a treatment measure layer, an acceptance standard layer, and a treatment effect layer; calculating the causal strength and causal closure degree corresponding to the second question-and-answer matching pair based on the Bayesian parameter sequence corresponding to the second question-and-answer matching pair and a preset historical parameter sequence; determining the score corresponding to the second question-and-answer matching pair based on the causal strength and the causal closure degree; filtering out scores greater than or equal to a preset score threshold; and determining the second question-and-answer matching pair corresponding to the filtered score as the third question-and-answer matching pair.

[0013] In some embodiments, the third question-and-answer matching pair is stored in a preset long-term memory database; the structured knowledge data is stored in a preset full lifecycle database; and the long-term memory database is cleared after the structured knowledge data is generated.

[0014] Secondly, an apparatus for storing knowledge of construction projects, the apparatus comprising: a first generation module configured to acquire raw project data to be processed daily after the construction project is initiated; generate a first question-and-answer matching pair corresponding to the raw project data; a correction module configured to perform knowledge correction on the first question-and-answer matching pairs generated within the first time period at intervals of a first time period to obtain a second question-and-answer matching pair; a filtering module configured to perform causal credibility filtering on the second question-and-answer matching pairs generated within the second time period at intervals of a second time period to obtain a third question-and-answer matching pair; and a second generation module configured to generate structured knowledge data corresponding to the construction project based on all third question-and-answer matching pairs of the construction project after the construction project is completed.

[0015] Thirdly, an electronic device includes a processor and a memory storing program instructions, the processor being configured to execute the aforementioned method for storing construction project knowledge when the program instructions are executed.

[0016] Fourthly, a storage medium stores program instructions that, when executed, perform the aforementioned method for storing knowledge of building projects.

[0017] Compared with existing technologies, this invention acquires raw project data daily after the construction project begins; generates a first question-and-answer matching pair corresponding to the raw project data; then, at a first time interval, performs knowledge correction on the first question-and-answer matching pairs generated within the first time interval to obtain a second question-and-answer matching pair; at a second time interval, performs causal credibility screening on the second question-and-answer matching pairs generated within the second time interval to obtain a third question-and-answer matching pair; and finally, after the construction project is completed, generates structured knowledge data corresponding to the construction project based on all the third question-and-answer matching pairs. In this way, this solution sequentially performs knowledge correction and causal credibility screening on the first question-and-answer matching pairs corresponding to the raw project data, and generates structured knowledge data corresponding to the construction project based on the corrected and screened third question-and-answer matching pairs. Compared with existing technologies, knowledge correction ensures the accuracy of the second question-and-answer matching pair, and causal credibility screening ensures that the causal credibility of the third question-and-answer matching pair meets the standard, thereby improving the accuracy and causal reliability of the final generated structured knowledge data, and thus enhancing the quality, reliability, and portability of the final generated structured knowledge data. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0019] Figure 1 This is a flowchart illustrating a method for storing knowledge of building projects provided in an embodiment of this disclosure;

[0020] Figure 2 This is an example diagram of a causal directed acyclic graph for a building project provided in an embodiment of this disclosure;

[0021] Figure 3 This is a schematic diagram of the structure of a building hybrid expert model provided in an embodiment of this disclosure;

[0022] Figure 4 This is a schematic diagram of a device for storing construction project knowledge provided in an embodiment of this disclosure;

[0023] Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0025] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0026] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0027] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0028] In this application, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0029] It should be noted that the disc motor to be monitored has multiple partitions, and each partition is equipped with a permanent magnet and a stator winding.

[0030] Please see Figure 1 , Figure 1 This is a schematic flowchart illustrating a method for storing knowledge about a construction project, as shown in an exemplary embodiment of this application. The method includes:

[0031] Step S101: After the construction project is launched, acquire the raw project data to be processed every day; generate the first question-and-answer matching pair corresponding to the raw project data.

[0032] Step S102: After a first time interval, perform knowledge correction on the first question-and-answer matching pair generated within the first time interval to obtain the second question-and-answer matching pair.

[0033] Step S103: After a second time interval, perform causal credibility filtering on the second question-and-answer matching pairs generated within the second time interval to obtain the third question-and-answer matching pairs.

[0034] Step S104: After the construction project is completed, generate structured knowledge data corresponding to the construction project based on all third question-answer matching pairs of the construction project.

[0035] In this embodiment, after the construction project starts, the original project data to be processed is acquired daily. A first question-and-answer matching pair corresponding to the original project data is generated. Then, at a first time interval, the first question-and-answer matching pairs generated within the first time interval are corrected to obtain a second question-and-answer matching pair. At a second time interval, the second question-and-answer matching pairs generated within the second time interval are filtered for causal reliability to obtain a third question-and-answer matching pair. Finally, after the construction project is completed, structured knowledge data corresponding to the construction project is generated based on all the third question-and-answer matching pairs. In this way, this solution sequentially corrects the knowledge and filters the causal reliability of the first question-and-answer matching pairs corresponding to the original project data, and generates structured knowledge data corresponding to the construction project based on the corrected and filtered third question-and-answer matching pairs. Compared with existing technologies, knowledge correction ensures the accuracy of the second question-and-answer matching pair, and causal reliability filtering ensures that the causal reliability of the third question-and-answer matching pair meets the standard. This improves the accuracy and causal reliability of the final generated structured knowledge data, thereby enhancing the quality, reliability, and portability of the final generated structured knowledge data.

[0036] Furthermore, the system acquires raw project data daily, including by connecting to the enterprise project management system, smart construction site platform, and OA (Office Automation) system corresponding to the construction project through pre-defined standardized interfaces.

[0037] It should be noted that the original project data may include business data generated during the construction project, such as construction logs, inspection records, hazard rectification orders, construction plans, technical briefings, change orders, cost ledgers, and completion acceptance documents.

[0038] Furthermore, generating the first question-and-answer matching pair corresponding to the original project data includes: extracting the question text and the corresponding answer text from the original project data; and generating the first question-and-answer matching pair based on the question text and the answer text.

[0039] In this way, the question text and the corresponding answer text are extracted directly from the original project data; the first question-answer matching pair is generated based on the question text and answer text, without the need for manual annotation of the original project data, which is efficient and accurate.

[0040] Furthermore, the question text and the corresponding answer text are extracted from the original project data, including: using a building hybrid expert model to read the original project data in order to extract the question text and the corresponding answer text from the original project data.

[0041] It should be noted that the architecture hybrid expert model is a MoE (Mixture of Experts) large language model. It can extract the original question text from the project's original data; then, it matches the extracted original answer text with the original question text in the original project data. The original question text and its corresponding original answer text are then deduplicated, aligned, denoised, and anonymized to obtain the question text and its corresponding answer text.

[0042] The original question text is the raw textual expression of the question in the project's original data. The original answer text is the raw textual expression of the methods used to address the question and the results after those methods.

[0043] Deduplication involves removing duplicate descriptions of the same question. Alignment represents the format mapping between the original question text and its corresponding original answer text; for example, unifying time and colleague formats, address formats, and entity naming. Noise reduction removes irrelevant content such as template slogans, garbled characters, meaningless punctuation, handwritten typos, blank notes, and repeated separators from the original text, ensuring that the denoised text retains only the question and answer content. Sensitization blurs or removes sensitive entities from the original text; sensitive entities can include sensitive personal information such as names and phone numbers, or commercially sensitive content such as contract numbers, profits, and base prices.

[0044] It should be noted that generating the first question-and-answer matching pair based on the question text and answer text includes: generating an initial question-and-answer pair that includes the question text and the corresponding answer text; adding a question ID (identity) tag, a question specialty tag, a data source tag, and a generation time tag to the initial question-and-answer pair to obtain the first question-and-answer matching pair.

[0045] The issue ID label describes the issue ID corresponding to the issue text. The issue ID corresponds one-to-one with the issue text and can be an automatically generated number.

[0046] The question-specific note describes the relevant field of expertise for the question text.

[0047] The data source label indicates the origin of the problem text, specifying which file within the business data it comes from. This file could be a construction log, inspection record, hazard rectification order, construction plan, technical briefing, change order, cost ledger, or final acceptance documentation, etc.

[0048] The generation timestamp indicates when the first question-and-answer matching pair was generated.

[0049] The first question-and-answer matching pair includes the question text and the corresponding answer text.

[0050] The problem text includes location text, main body text, and phenomenon text. Location text indicates the location where the problem occurs; main body text describes the subject of the problem; and phenomenon text describes the phenomenon of the problem.

[0051] The answer text may include specific handling methods and / or handling results.

[0052] For example: The problem text is "There is a crack in the east wall of Room 201, 3rd Floor, XX Building". The location text is "east side of Room 201, 3rd Floor, XX Building"; the main text is "wall"; and the phenomenon text is "there is a crack".

[0053] It should be noted that the first question-and-answer matching pair is stored in a preset instantaneous memory database; the instantaneous memory database is cleared at the first time interval.

[0054] In the instantaneous memory database, the first question-answer matching pair is stored in JSON (JavaScript Object Notation) format.

[0055] In some embodiments, the transient memory database serves as a high-speed temporary cache area. Leveraging the high read / write performance of memory, it provides millisecond-level data access support for subsequent knowledge correction tasks. The transient memory database acts as a buffer layer between the first question-answer matching pair and model learning.

[0056] It should be noted that the first time period lasts for 24 hours. During knowledge correction, after reading all the first question-and-answer matching pairs, the database is automatically cleared, releasing memory resources in the transient memory database to provide storage space for the first question-and-answer matching pairs generated in the next first time period.

[0057] Furthermore, the first question-and-answer matching pairs generated within the first time period are corrected to obtain the second question-and-answer matching pairs. This includes: inputting all the first question-and-answer matching pairs stored in the transient memory database into a preset building hybrid expert model to obtain the description text of the first question-and-answer matching pairs; using a preset causal directed acyclic graph of building engineering to determine whether there is a conflict between the description texts; if there is a conflict between the description texts, correcting the first question-and-answer matching pairs corresponding to the conflicting description texts to obtain the second question-and-answer matching pairs; and / or, if there is no conflict between the description texts, determining the first question-and-answer matching pairs corresponding to the non-conflicting description texts as the second question-and-answer matching pairs.

[0058] In this way, by inputting all the first question-and-answer matching pairs stored in the transient memory database into the preset building hybrid expert model, the description text of the first question-and-answer matching pair is obtained. Then, the preset building engineering causal directed acyclic graph is used to determine whether there is a conflict between the description texts, so as to correct the first question-and-answer matching pairs corresponding to the conflicting description texts and obtain the second question-and-answer matching pairs. The first question-and-answer matching pairs corresponding to the non-conflicting description texts are determined as the second question-and-answer matching pairs, thereby improving the accuracy and reliability of the second question-and-answer matching pairs.

[0059] It should be noted that the root node of the causal directed acyclic graph of construction projects is the name of the causal directed acyclic graph of construction projects, for example, it can be "causal relationship of construction projects".

[0060] It should be noted that the first-level nodes of the causal directed acyclic graph of construction projects include one or more of the following: project ID node, project stage node, business type node, risk and hazard node, disposal plan node, and execution effect node.

[0061] The value of the project ID node is the identity ID corresponding to the building project. It should be noted that each building project can be associated with a separate causal directed acyclic graph (DAG).

[0062] The value of a project phase node represents the phase in which the construction project is located. Specifically, the phase includes: the in-process phase, the suspension phase, the resumption phase, and the completion phase.

[0063] In some embodiments, the function for determining the stage of a construction project is: .in, The stage at which the construction project P is located; The current date; The update date is the last record in the original data of the building project P; the completion tag is used to indicate the completion and clearing of the building project P; in some embodiments, if the building project P is completed, its corresponding completion tag is 1, indicating completion; if the building project P is not completed, its corresponding completion tag is 0, indicating incompleteness.

[0064] In some embodiments, incremental learning can be performed daily to process the raw project data for that day while the construction project is in progress.

[0065] When a construction project is in a standstill, daily incremental learning can be paused, and historical data can be retained for use as the third question-and-answer matching pairs generated.

[0066] When a construction project is in the resumption phase, it is necessary to catch up on the original project data from the period of suspension.

[0067] When a construction project is in the completion phase, the entire lifecycle knowledge base generation process is automatically triggered to generate structured knowledge data.

[0068] It should be noted that the value of the business type node represents the business type corresponding to the construction project P. For example: housing construction business, municipal business, highway transportation business, etc.

[0069] Risk and hazard nodes, response plan nodes, and implementation effectiveness nodes all have secondary nodes, which are further subdivisions of primary nodes. For secondary nodes under risk and hazard nodes, tertiary nodes can be set to further refine the secondary nodes.

[0070] In a causal directed acyclic graph (DAG) for building engineering, leaf nodes can be connected by causal edges to represent causal relationships between them. The direction of a causal edge represents the causal direction between the two nodes it connects.

[0071] Specifically, the leaf nodes under the "Risk / Hazard" node are causal nodes, representing the problem; the leaf nodes under the "Disposal Plan" node are disposal process nodes, representing the disposal method; and the leaf nodes under the "Execution Effectiveness" node are result nodes, representing the result after disposal. Causal nodes are connected to disposal process nodes via causal edges; disposal process nodes are connected to result nodes via causal edges.

[0072] In some embodiments, please refer to Figure 2 , Figure 2 This is an example diagram of a causal directed acyclic graph for a building project. (Example:) Figure 2 As shown, the root node of the causal directed acyclic graph of construction projects represents the causal relationships within the construction project. Red nodes are first-level nodes, specifically including: Project ID node, Project Stage node, Risk / Hazard node, Remediation Plan node, and Implementation Results node. Blue nodes are second-level nodes. The second-level nodes under the Risk / Hazard node include: Structural Hazard node, Mechanical and Electrical Hazard node, and Safety Hazard node. The second-level nodes under the Remediation Plan node include: Remediation Plan 1 node and Remediation Plan 2 node, representing specific remediation methods. The second-level nodes under the Implementation Results node include: Completely Resolved node, Basically Resolved node, and Unresolved node, representing the results of the remediation. The Safety Hazard node also has third-level nodes: Safety Hazard 1 node and Safety Hazard 2 node, representing specific safety hazards, which can include both the phenomenon and the cause of the safety hazard.

[0073] Among them, the safety hazard node 1 and the disposal plan node 1 are connected by a causal edge ( Figure 2The nodes are connected by dashed lines; the node for handling solution 1 and the node for complete resolution are connected by causal edges; this causal chain represents that the problem represented by node safety hazard 1 was handled according to the handling method represented by node handling solution 1, and was ultimately completely resolved. The node for safety hazard 2 and the node for handling solution 2 are connected by causal edges; the node for handling solution 2 and the node for basic resolution are connected by causal edges; this causal chain represents that the problem represented by node safety hazard 2 was handled according to the handling method represented by node handling solution 2, and was ultimately basically resolved.

[0074] Furthermore, a pre-defined causal directed acyclic graph (DAG) for architectural engineering is used to determine whether conflicts exist between descriptive texts. This includes: for the descriptive text corresponding to the same first question-and-answer matching pair, the descriptive text is broken down according to a pre-defined format to construct a question-and-answer causal chain; the validity of the question-and-answer causal chain is verified; the question-and-answer causal chain is stored in the causal directed acyclic graph for architectural engineering; the cause nodes of the question-and-answer causal chains are extracted from the causal directed acyclic graph for architectural engineering; and the existence of conflicts is determined based on the content of the cause nodes, which is then judged by the architectural hybrid expert model. If the content of the cause node is conflicting, then the descriptive text is determined to be conflicting; if the content of the cause node is not conflicting, then the descriptive texts are determined not to be conflicting. In this way, the causal relationship of the descriptive text of the first question-and-answer matching pair can be extracted through the causal directed acyclic graph for architectural engineering, which facilitates the identification of conflicts between causal relationships, thereby reducing the causal contradictions in the corrected second question-and-answer matching pair.

[0075] It should be noted that the preset large language model is the MOE large language model. The professional domain description text can include problem phenomenon text, problem cause text, handling method text, and handling result text. Specifically, the problem phenomenon text is generated from the problem text, and it can be a combination of the main text and the phenomenon text; the problem cause text can be generated from the problem text, or it can be generated by the architectural hybrid expert model based on the problem text; the handling method text and the handling result text can be generated from the answer text.

[0076] It should be noted that the processing result text indicates whether the processing was successful or unsuccessful.

[0077] Furthermore, for the description text corresponding to the same first question-and-answer matching pair, the description text is decomposed according to a preset format to construct a question-and-answer causal chain, including: extracting the problem phenomenon text and problem cause text from any description text; combining the problem phenomenon text and all problem cause texts to obtain the content of the question-and-answer cause node in the question-and-answer causal chain; determining the handling method text in the description text as the content of the handling process node in the question-and-answer causal chain; and determining the handling result text in the description text as the content of the result node in the question-and-answer causal chain.

[0078] In this way, since the professional domain description text corresponding to the same first question-and-answer match describes the handling methods and results of the same phenomenon, the content of the stored result node and the handling process node can be directly determined based on the relevant descriptions in any description text, while the content about the cause in the cause node is determined by the architectural hybrid expert model. This allows a question-and-answer causal chain to fully reflect the completion and handling of the corresponding problem for the first question-and-answer match.

[0079] It should be noted that the legality verification of the question-and-answer causal chain is to verify the causal flow of the question-and-answer causal chain. If no causal error is found in the causal flow, the legality verification is determined to pass; otherwise, the legality verification is determined to fail.

[0080] Specifically, the causal flow direction is the direction of the causal edge. If the causal flow direction is not "cause node flows to process node and process node flows to result node", a causal error is determined to have occurred; otherwise, no causal error is determined to have occurred.

[0081] It should be noted that each question-and-answer causal chain is mounted and stored in the causal directed acyclic graph of the construction project. Specifically, the parent node is determined based on the content of the cause node in the question-and-answer causal chain. Then, the cause node in the question-and-answer causal chain is used as the leaf node of the parent node, the result node in the question-and-answer causal chain is used as the leaf node of the disposal process node, and the result node is used as the leaf node under the execution effect node, all mounted in the causal directed acyclic graph of the construction project. It should be noted that the parent node is a second-level node under the risk / hazard node, which can be set arbitrarily or determined by understanding the content of the cause node through a large language model.

[0082] Furthermore, determining whether a conflict exists based on the content of the cause node includes: extracting several problem cause texts from the content of the cause node. If the causes represented by the problem cause texts are different, a conflict is determined to exist in the content of the cause node. Otherwise, a conflict is determined to exist in the content of the cause node.

[0083] In some embodiments, a cause node contains the following: "Phenomenon: Exterior wall leakage; Cause: Exterior wall leakage is due to improper treatment of construction joints, and exterior wall leakage is due to insufficient compaction of concrete during pouring." The extracted cause text includes both "Exterior wall leakage is due to improper treatment of construction joints" and "Exterior wall leakage is due to insufficient compaction of concrete during pouring." Since the two causes are not identical, a conflict is identified.

[0084] It should be noted that correcting the first question-and-answer matching pair corresponding to the conflicting description vectors to obtain the second question-and-answer matching pair triggers the manual review process, which corrects the first question-and-answer matching pair.

[0085] It should be noted that "correction" refers to manually adding relevant reasons to the question text in the first question-answer matching pair. It should also be noted that manual correction can be made based on the content of the question reason text.

[0086] Furthermore, the building hybrid expert model includes a building backbone network, a gating routing module, and several pre-defined sub-expert networks corresponding to different building engineering specialties. The building backbone network is used to obtain the subdivided domain feature vectors corresponding to the first question-answer matching pair and to obtain the descriptive text corresponding to the first question-answer matching based on the descriptive text feature vectors. The gating routing module calculates the activation probability of each sub-expert network based on the subdivided domain feature vectors and selects several sub-expert networks to be activated based on the activation probabilities. The sub-expert networks to be activated are used to receive the feature vectors and output the professional domain description vectors corresponding to the building engineering specialty. Based on the activation probabilities output by the gating routing module, the subdivided domain feature vectors output by all sub-expert networks to be activated are weighted and fused to obtain the descriptive text feature vectors.

[0087] In this way, by employing a hierarchical architecture of a backbone gating network and multi-disciplinary sub-expert networks, the architecture backbone network obtains the sub-domain feature vectors and activation probabilities corresponding to the first question-answer matching pair. Within the sub-expert networks, it determines the sub-expert networks to be activated. These sub-expert networks then determine their corresponding architectural engineering specialty description vectors based on the sub-domain feature vectors, making the model's output more specialized. Furthermore, by activating only a subset of sub-expert networks using activation probabilities, compared to using all sub-expert networks to obtain the specialty description vectors, the memory requirements are lower and the efficiency is higher.

[0088] It should be noted that the activation probability of the sub-expert network to be activated is higher than that of other inactive networks. The number of sub-expert networks to be activated is a preset number.

[0089] In some embodiments, please refer to Figure 3 , Figure 3 This is a structural schematic diagram of a hybrid expert model for architecture.

[0090] like Figure 3 As shown, the building hybrid expert model 300 includes a building backbone network 301, several pre-defined sub-expert networks 302 corresponding to different building engineering specialties, and a gated routing module 303. Specifically, the building backbone network 301 includes multiple linear transformation layers and decoding layers.

[0091] Each linear transformation layer is used to perform a linear transformation on the first question-answer matching pair to obtain the subdivided domain feature vector corresponding to the first question-answer matching pair.

[0092] The decoding layer is used to perform deep feature extraction on the unified descriptive text feature vector to generate the descriptive text corresponding to the first question-answer match of the descriptive text feature vector.

[0093] The gated routing module 303 includes a normalization layer, which normalizes the feature vectors of the subdivided domain to obtain the activation probabilities of each sub-expert network; it then identifies the sub-expert networks corresponding to the n highest activation probabilities as the sub-expert networks to be activated, and sends the subdivided domain feature vectors to each sub-expert network to be activated. Here, n is the number of activations.

[0094] In this embodiment, n can be 2.

[0095] The activation sub-expert network is used to determine the professional domain description vector of the architectural engineering specialty corresponding to the sub-expert network to be activated based on the feature vector. Then, based on the activation probability output by the gating routing module, the sub-domain feature vectors output by all the sub-expert networks to be activated are weighted and fused to obtain the descriptive text feature vector.

[0096] At this time, the other sub-expert networks are in a dormant state.

[0097] In some embodiments, n is 2, and the gated routing module 303 calculates... This yields a unified feature vector describing the text. Among these features, To provide a unified description of text feature vectors; This is the index number corresponding to the first sub-expert network to be activated; This represents the activation probability of the first sub-expert network to be activated. The feature vector of the sub-domain corresponding to the first question-and-answer matching pair; This is the domain description vector corresponding to the first expert network to be activated. This is the index number corresponding to the second sub-expert network to be activated; This represents the activation probability of the second sub-expert network to be activated. This is the domain description vector corresponding to the second expert network to be activated.

[0098] It should be noted that, It is a unified feature vector obtained after weighted fusion by multiple experts. This unified feature vector comprehensively represents the professional domain information corresponding to the current structured first question-and-answer matching pair, eliminating the domain bias of a single expert's output.

[0099] Pki is the weighted weight obtained by normalizing the original probabilities of the two activation experts twice.

[0100] and These are floating-point feature vectors with identical dimensions obtained after encoding by the activation expert network through the embedding layer. The vectors have a uniform length and are used for... , Perform element-wise weighted operations.

[0101] It should be noted that the second question-and-answer matching pair is stored in a preset short-term memory database; the short-term memory database is cleared every second time interval.

[0102] Furthermore, the second question-and-answer matching pairs generated within the second time period are subjected to causal credibility screening to obtain the third question-and-answer matching pairs. This includes: for all second question-and-answer matching pairs stored in the short-term memory database, determining the parameters corresponding to each layer node in a preset multi-layer Bayesian network based on the second question-and-answer matching pairs, and obtaining the Bayesian parameter sequence corresponding to the second question-and-answer matching pairs; the multi-layer Bayesian network includes: hazard type layer, risk factor layer, treatment measure layer, acceptance standard layer, and treatment effect layer; calculating the causal strength and causal closure degree corresponding to the second question-and-answer matching pairs based on the Bayesian parameter sequence corresponding to the second question-and-answer matching pairs and the preset historical parameter sequence; determining the score corresponding to the second question-and-answer matching pairs based on the causal strength and causal closure degree; filtering out scores greater than or equal to a preset score threshold; and determining the second question-and-answer matching pairs corresponding to the filtered scores as the third question-and-answer matching pairs.

[0103] In this way, using a multi-layer Bayesian network as the quantitative foundation, the causal strength and causal closure of the second question-and-answer matching pair after professional correction are obtained. Based on the dual indicators of causal strength and causal closure, the causal credibility of the second question-and-answer matching pair after professional correction is quantitatively scored. Then, the filtering is completed based on the score, so that the causal logic of the selected third question-and-answer matching pair is complete and credible. This achieves a second quality purification of the question-and-answer matching pair from "professional compliance" to "causal credibility", and further improves the reliability of the third question-and-answer matching pair.

[0104] It should be noted that the second time period lasts for 7 days.

[0105] It should be noted that the nodes in the hazard type layer are discrete nodes, and their values ​​are hazard types in building engineering such as structural hazards, electromechanical hazards, and safety hazards. They correspond to the secondary nodes under the risk hazard nodes in the causal directed acyclic graph of building engineering.

[0106] The nodes in the risk factor layer are discrete nodes, and their values ​​are core risk factors such as geological conditions, climate conditions, construction technology, and material quality; these values ​​are set manually.

[0107] The nodes in the handling measures layer are discrete nodes, and their values ​​are the descriptions of the handling methods in the answer text of the second question-answer matching pair;

[0108] The nodes in the acceptance criteria layer are discrete nodes, and their values ​​are the acceptance criteria corresponding to the values ​​in the hazard type layer; these can be specified by national and / or industry standards.

[0109] The treatment effect layer consists of discrete nodes, with values ​​representing three possibilities: "completely resolved," "basically resolved," and "unresolved." These correspond to the second-level nodes under the performance effect node in the causal directed acyclic graph of the building project.

[0110] The Bayesian parameter sequence corresponding to the second question-answer matching pair is the sequence of values ​​of the second question-answer matching pair in the multilayer Bayesian network. If there is no value, the value of the element is set to NULL.

[0111] It should be noted that the historical parameter sequence is the Bayesian parameter sequence obtained in the past.

[0112] Furthermore, based on the Bayesian parameter sequence and the preset historical parameter sequence corresponding to the second question-and-answer matching pair, the causal strength and causal closure degree corresponding to the second question-and-answer matching pair are calculated. This includes: determining the number of second question-and-answer matching pairs with the same question text representation and the same treatment plan based on the values ​​of the nodes in the treatment effect layer of the Bayesian parameter sequence and the historical parameter sequence; determining this number as the total number of executed cases; determining the number of valid cases that achieved the treatment based on the treatment results; and determining the actual number of valid nodes corresponding to the second question-and-answer matching pair. The causal strength is obtained by calculating the ratio between the number of valid cases and the total number of executed cases. The causal closure degree is obtained by calculating the ratio between the actual number of valid nodes and the preset theoretical maximum number of nodes.

[0113] It should be noted that the number of valid cases is the number of sequences whose values ​​in the treatment effect layer of the Bayesian parameter sequence and the historical parameter sequence are completely resolved; the total number of executed cases is the total number of sequences in the Bayesian parameter sequence and the historical parameter sequence.

[0114] The actual number of valid nodes is the number of non-empty elements in the Bayesian parameter sequence corresponding to the second question-answer matching pair; the actual number of valid nodes is the total number of all elements in the Bayesian parameter sequence, which is 5.

[0115] Furthermore, the score corresponding to the second question-and-answer pair is determined based on causal strength and causal closure, including: obtaining the score corresponding to the second question-and-answer pair by calculating F=A×(1+0.5×B) / 1.5. Where F is the score corresponding to the second question-and-answer pair; A is the causal strength corresponding to the second question-and-answer pair; and B is the causal closure corresponding to the second question-and-answer pair.

[0116] In some embodiments, the scoring threshold can be 0.5, or it can be set to other values, which are not limited here.

[0117] Furthermore, the third question-and-answer matching pairs are stored in a preset long-term memory database; the structured knowledge data is stored in a preset full-lifecycle database; the long-term memory database is cleared after the structured knowledge data is generated.

[0118] In this way, by storing third-party question-and-answer matching pairs in a long-term memory database and structured knowledge data in a lifecycle database, and by clearing the long-term memory database after generating structured knowledge data, layered data storage, isolated storage, and cyclical reuse are achieved. Simultaneously, the long-term memory database acts as a cache layer between the third-party question-and-answer matching pairs and the structured knowledge data, and is cleared immediately after the structured data is generated and archived, reducing the redundant storage caused by the repeated accumulation of third-party question-and-answer matching pairs.

[0119] It should be noted that since the structured knowledge data is generated after completion, the long-term memory database is cleared after completion.

[0120] This solution simulates the human brain's memory mechanism, employing a hybrid storage architecture that combines instantaneous memory databases, short-term memory databases, long-term memory databases, and full life-cycle databases to achieve efficient storage, dynamic transfer, and precise retrieval of different types of knowledge.

[0121] In some embodiments, please refer to Table 1, which is a parameter example table for the hybrid storage architecture of this solution.

[0122] Table 1

[0123]

[0124] As shown in Table 1, the hybrid storage architecture has multiple memory levels, namely, transient memory database, short-term memory database, long-term memory database and full lifecycle database.

[0125] The transient memory database is stored in DDR5 high-speed memory, with a single DDR5 high-speed memory cell having a capacity of 64GB. The total number of cells can be set according to actual needs. The transient memory database retains data for 24 hours. It uses a hash index and has a read / write speed of 45GB / s.

[0126] The short-term memory database is stored as a file; its capacity is configured to be 10TB; the data retention period is 7 days; its indexing method is file path search; and its query rate is ≤50ms / time.

[0127] The long-term memory database is stored as a file; its capacity is configured to be 100TB; the data in the long-term memory database will be retained until completion; its indexing method is file path search; and its query rate is ≤200ms / time.

[0128] The full lifecycle database is stored in two formats: structured file storage and RAG (Retrieval-augmented Generation) vector library storage. The database has a capacity of 100TB and is used to store structured knowledge data. The structured knowledge data stored in the full lifecycle database is permanently retained; its indexing methods include vector indexing and file path searching; and its query speed is ≤200ms / query.

[0129] It should be noted that the completion of a construction project is determined as follows: obtain the completion label corresponding to the construction project; if the completion label indicates completion, the construction project is determined to be completed.

[0130] Furthermore, based on all third-party question-and-answer matching pairs for the construction project, structured knowledge data corresponding to the construction project is generated. This includes: reading and merging the original project data corresponding to the construction project to obtain a summary text; summarizing the summary text using a pre-set large language model to obtain a structured report; the structured report includes blocks on project overview, schedule management, quality and safety, cost control, lessons learned, and key causal knowledge. This structured report is then identified as structured knowledge data.

[0131] It should be noted that the full lifecycle database includes a vector database and a file database. When storing structured knowledge data in the full lifecycle database, a vector can be generated for each block of the structured knowledge data and stored in the vector database; at the same time, the structured knowledge data is stored in the file database in standard Markdown (lightweight markup language) format.

[0132] In this way, data from the entire lifecycle of the project is aggregated, a project management cognitive framework text is built, vectorized storage is completed, and project experience is standardized and reused across projects.

[0133] In some embodiments, when construction personnel discover a problem, they can query the corresponding handling method through a full lifecycle database. After handling, a hazard rectification order corresponding to the problem is issued. This hazard rectification order can then serve as raw data for a new project. Based on the method described in this solution, its corresponding third-party question-and-answer matching pair is uploaded, and then further structured knowledge data is generated.

[0134] It should be noted that this solution can utilize a large language model to generate project debriefing reports. This involves using a large language model to analyze the previous day's raw project data and generate a project debriefing report. The project debriefing report focuses on four dimensions: progress execution, quality issues, safety hazards, and key management points. It outlines the project's operational status, prominent problems, risky aspects, and subsequent management priorities for the day. The project debriefing report can be specifically sent to designated project management personnel.

[0135] Furthermore, the method for storing architectural project knowledge also includes: fine-tuning the architectural hybrid expert model by using all third question-and-answer matching pairs at third time intervals.

[0136] In this way, by using a fixed third time period as the iteration cycle, and using high-quality third question-and-answer matching pairs that have undergone professional correction, dual screening of causal credibility, and storage in long-term memory, the architectural hybrid expert model is periodically fine-tuned, forming a closed-loop self-optimization mechanism from data quality control to model iteration. This can unify the reasoning logic and judgment criteria of all sub-experts, so as to facilitate the accurate generation of the next professional domain description vector.

[0137] In some embodiments, the duration of the third time period can be 15 days, or it can be set to other durations, without limitation.

[0138] It should be noted that the building hybrid expert model is fine-tuned by using all the third-party question-answering pairs. In other words, all the third-party question-answering pairs are used to incrementally train the building hybrid expert model in order to fine-tune the building hybrid expert model.

[0139] It should be noted that during incremental training, the loss function is Loss = 0.7 × Lj + 0.3 × Lk. Here, Loss is the loss value of the loss function; Lj is the cross-entropy loss, used to ensure the output conforms to engineering specifications; and Lk is the KL divergence loss, used to constrain the stability of the inference distribution within sub-experts and prevent the forgetting of historical abilities during incremental learning. In this way, by minimizing the loss, the hybrid expert model for construction is fine-tuned, enabling the extraction of cross-domain common business rules, unifying the inference logic and judgment criteria of all sub-experts, correcting cognitive biases generated by long-term incremental learning, and synchronously updating the fine-tuned universal rules to the long-term memory database, thus completing the iterative optimization of the knowledge system.

[0140] This solution constructs a multi-path parallel autonomous learning architecture using a hybrid expert model of architecture, forms a four-level hierarchical distributed causal memory mechanism through a transient memory database, a short-term memory database, a long-term memory database, and a full lifecycle database, incorporates a MoE dynamic expert module, incremental unlabeled training, and introduces causal inference and project immersion learning strategies. Compared with existing large-scale model solutions in the construction industry, it achieves the following significant technical effects:

[0141] 1. Significantly reduces implementation costs and eliminates reliance on manual annotation. Question-answer matching pairs are generated based on raw business data and aggregated into structured knowledge data after completion, eliminating the need for manual compilation and annotation of engineering documents. Simultaneously, the MoE sparse dynamic hybrid expert model for buildings is adopted, activating only the two most probable sub-expert networks for computation, reducing GPU requirements from high-end H100 to RTX4090. Compared to the traditional solution's total cost of approximately 12 million RMB for annotation of millions of data points, this solution only requires system deployment and computing power costs of approximately 50,000 RMB, representing a cost reduction of 99.6% and achieving low-cost deployment.

[0142] 2. Significantly improved knowledge update efficiency and effective accumulation of tacit experience. By replacing the traditional batch full-scale fine-tuning mode with a mechanism of timed knowledge correction and monthly incremental training, the knowledge update delay is ≤24 hours, which is more than 30 times faster than the traditional 7-15 day update cycle; at the same time, immersive learning of the original project data can automatically extract tacit experience such as on-site problem handling and project management skills, and store it in a distributed memory pool, solving the problem that tacit experience cannot be accumulated in traditional solutions.

[0143] 3. The model exhibits balanced professional capabilities, significantly improving inference accuracy. MoE enables autonomous learning through multi-path parallel processing, breaking down architectural engineering knowledge into sub-expert networks across multiple professional domains. This facilitates specialized learning across different domains and avoids the imbalances in capabilities caused by single-threaded learning. Furthermore, by combining knowledge correction and causal credibility screening, it avoids "spurious associations" arising from correlational inference in traditional models, greatly enhancing inference accuracy.

[0144] 4. High portability of knowledge assets, mitigating data silos and privacy concerns. A four-level hierarchical distributed causal memory pool, comprised of an instantaneous memory database, a short-term memory database, a long-term memory database, and a full lifecycle database, is used for storage. Standardized storage formats are employed, supporting integration with any large model base, database, and engineering business system. The accumulated knowledge assets can be directly exported and transferred to other enterprises or projects, resolving the issues of deep data binding and poor portability inherent in traditional models. Furthermore, original data does not need to be shared; only standardized knowledge assets are accumulated, balancing data privacy protection with industry knowledge reuse needs.

[0145] 6. The model's growth rate far exceeds that of manual labor, significantly improving business support efficiency. This solution implements an adaptive learning mechanism for the entire lifecycle of construction projects. The large language model participates immersively in the entire lifecycle of a single project and participates in multiple projects in batches and in parallel, enabling the accumulation of project experience equivalent to more than 20 years of experience by senior engineering experts within 1-2 years. Data processing efficiency is greatly improved, with the structured processing speed of various professional issues being 50 times faster than manual processing, the efficiency of construction plan compilation and review increasing by 167%, and the training cycle for new personnel shortened by 73%, comprehensively empowering all business scenarios in the construction industry.

[0146] Combination Figure 4 As shown, this disclosure provides an apparatus 400 for storing knowledge of building projects. The apparatus includes a first generation module 401, a correction module 402, a filtering module 403, and a second generation module 404.

[0147] The first generation module 401 is configured to acquire the raw project data to be processed daily after the construction project is started; and generate the first question-and-answer matching pair corresponding to the raw project data.

[0148] The correction module 402 is configured to perform knowledge correction on the first question-and-answer matching pair generated within the first time period at intervals of the first time period to obtain the second question-and-answer matching pair.

[0149] The filtering module 403 is configured to perform causal credibility filtering on the second question-and-answer matching pairs generated within the second time period at intervals of the second time period to obtain the third question-and-answer matching pairs.

[0150] The second generation module 404 is configured to generate structured knowledge data corresponding to the building project based on all third question-answer matching pairs of the building project after the building project is completed.

[0151] The apparatus for storing construction project knowledge provided in this disclosure acquires raw project data to be processed daily after the construction project begins; generates a first question-and-answer matching pair corresponding to the raw project data; then, at a first time interval, performs knowledge correction on the first question-and-answer matching pairs generated within the first time interval to obtain a second question-and-answer matching pair; at a second time interval, performs causal credibility filtering on the second question-and-answer matching pairs generated within the second time interval to obtain a third question-and-answer matching pair; and finally, after the construction project is completed, generates structured knowledge data corresponding to the construction project based on all the third question-and-answer matching pairs. In this way, this solution sequentially performs knowledge correction and causal credibility filtering on the first question-and-answer matching pairs corresponding to the raw project data, and generates structured knowledge data corresponding to the construction project based on the corrected and filtered third question-and-answer matching pairs. Compared with existing technologies, knowledge correction ensures the accuracy of the second question-and-answer matching pair, and causal credibility filtering ensures that the causal credibility of the third question-and-answer matching pair meets the standard, thereby improving the accuracy and causal reliability of the finally generated structured knowledge data, and thus enhancing the quality and reliability of the finally generated structured knowledge data.

[0152] Furthermore, the first generation module is configured to generate the first question-and-answer matching pair corresponding to the original project data in the following manner: extracting the question text and the answer text corresponding to the question text from the original project data; and generating the first question-and-answer matching pair based on the question text and the answer text.

[0153] Furthermore, the first question-and-answer matching pairs are stored in a preset transient memory database; the transient memory database is cleared every first time interval; the correction module is configured to correct the knowledge of the first question-and-answer matching pairs generated within the first time interval in the following manner to obtain the second question-and-answer matching pairs: input all the first question-and-answer matching pairs stored in the transient memory database into the building hybrid expert model to obtain the descriptive text corresponding to the first question-and-answer matching; use a preset causal directed acyclic graph of building engineering to determine whether there is a conflict between the descriptive texts; if there is a conflict between the descriptive texts, correct the first question-and-answer matching pairs corresponding to the conflicting descriptive texts to obtain the second question-and-answer matching pairs; and / or, if there is no conflict between the descriptive texts, correct the first question-and-answer matching pairs corresponding to the conflicting descriptive texts. The first question-and-answer matching pair corresponding to conflicting descriptive texts is determined as the second question-and-answer matching pair. All first question-and-answer matching pairs stored in the transient memory database are input into a preset architectural hybrid expert model to obtain several professional domain descriptive texts corresponding to the first question-and-answer matching pairs. A preset causal directed acyclic graph of architectural engineering is used to determine whether there are conflicts between the professional domain descriptive texts. If there are conflicts between the professional domain descriptive texts, the first question-and-answer matching pairs corresponding to the conflicting professional domain descriptive texts are corrected to obtain the second question-and-answer matching pairs. And / or, if there are no conflicts between the professional domain descriptive texts, the first question-and-answer matching pairs corresponding to the non-conflicting professional domain descriptive texts are determined as the second question-and-answer matching pairs.

[0154] Furthermore, the building hybrid expert model includes a building backbone network, a gating routing module, and several pre-defined sub-expert networks corresponding to different building engineering specialties. The building backbone network is used to obtain the sub-domain feature vectors corresponding to the first question-answer matching pair and to obtain the descriptive text corresponding to the first question-answer matching based on the descriptive text feature vectors. The gating routing module calculates the activation probability of each sub-expert network based on the sub-domain feature vectors and selects several sub-expert networks to be activated based on the activation probabilities. The sub-expert networks to be activated receive the sub-domain feature vectors and output the corresponding professional domain description vectors of the building engineering specialty. The sub-domain feature vectors output by all the sub-expert networks to be activated are weighted and fused based on the activation probabilities output by the gating routing module to obtain the descriptive text feature vectors.

[0155] Furthermore, the device also includes a fine-tuning module configured to fine-tune the building hybrid expert model at third time intervals using all third question-and-answer matching pairs.

[0156] Furthermore, the second question-and-answer matching pairs are stored in a preset short-term memory database; the short-term memory database is cleared every second time interval; the filtering module is configured to perform causal credibility filtering on the second question-and-answer matching pairs generated within the second time interval in the following manner to obtain the third question-and-answer matching pairs: for all second question-and-answer matching pairs stored in the short-term memory database, the parameters corresponding to each layer node in the preset multi-layer Bayesian network are determined based on the second question-and-answer matching pairs, and the Bayesian parameter sequence corresponding to the second question-and-answer matching pairs is obtained; the multi-layer Bayesian network includes: hazard type layer, risk factor layer, treatment measure layer, acceptance standard layer, and treatment effect layer; the causal strength and causal closure degree corresponding to the second question-and-answer matching pairs are calculated based on the Bayesian parameter sequence corresponding to the second question-and-answer matching pairs and the preset historical parameter sequence; the score corresponding to the second question-and-answer matching pairs is determined based on the causal strength and causal closure degree; scores greater than or equal to the preset score threshold are filtered out; and the second question-and-answer matching pairs corresponding to the filtered scores are determined as the third question-and-answer matching pairs.

[0157] Furthermore, the third question-and-answer matching pairs are stored in a preset long-term memory database; the structured knowledge data is stored in a preset full-lifecycle database; the long-term memory database is cleared after the structured knowledge data is generated.

[0158] Combination Figure 5 As shown, this embodiment of the disclosure provides an electronic device 500, including a processor 501 and a memory 502. Optionally, the device may further include a communication interface 503 and a bus 504. The processor 501, communication interface 503, and memory 502 can communicate with each other via the bus 504. The communication interface 503 can be used for information transmission. The processor 501 can call logical instructions in the memory 502 to execute the above-mentioned... Figure 1 The method shown is for storing knowledge about building projects.

[0159] Furthermore, the logic instructions in the aforementioned memory 502 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0160] The memory 502, as a storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in this embodiment. The processor 501 executes functional applications and data processing by running the program instructions / modules stored in the memory 502, thus achieving the above-described functionality. Figure 1 The method shown is for storing knowledge about building projects.

[0161] The memory 502 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 502 may include high-speed random access memory and may also include non-volatile memory.

[0162] This disclosure provides a storage medium storing computer-executable instructions configured to perform the method described above for storing knowledge of building projects.

[0163] The aforementioned storage media can be either transient computer-readable storage media or non-transitory computer-readable storage media. Non-transitory storage media include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, and can also be transient storage media.

[0164] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0165] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0166] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than those disclosed in the description; sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. Each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A method for storing knowledge about construction projects, characterized in that, include: After the construction project is launched, the raw project data to be processed is obtained daily; Generate the first question-and-answer matching pair corresponding to the original data of the project; After a first time interval, the first question-and-answer matching pair generated within the first time interval is corrected to obtain the second question-and-answer matching pair. After a second time interval, the second question-and-answer matching pairs generated within the second time interval are subjected to causal credibility filtering to obtain the third question-and-answer matching pairs. After the construction project is completed, structured knowledge data corresponding to the construction project is generated based on all third question-answer matching pairs of the construction project.

2. The method according to claim 1, characterized in that, The generation of the first question-and-answer matching pair corresponding to the original data of the project includes: Extract the question text and the corresponding answer text from the original data of the project; The first question-and-answer matching pair is generated based on the question text and the answer text.

3. The method according to claim 1, characterized in that, The first question-and-answer matching pair is stored in a preset instantaneous memory database; the instantaneous memory database is cleared every first time interval; the knowledge correction of the first question-and-answer matching pair generated within the first time interval to obtain the second question-and-answer matching pair includes: Input all the first question-and-answer matching pairs stored in the instantaneous memory database into the building hybrid expert model to obtain the descriptive text corresponding to the first question-and-answer matching; The existence of conflicts between the descriptive texts is determined using a pre-defined causal directed acyclic graph of construction projects. In the event of a conflict between the description texts, the first question-and-answer matching pair corresponding to the conflicting description text is corrected to obtain the second question-and-answer matching pair; and / or, in the event of no conflict between the description texts, the first question-and-answer matching pair corresponding to the non-conflicting description text is determined as the second question-and-answer matching pair.

4. The method according to claim 3, characterized in that, The building hybrid expert model includes a building backbone network, a gated routing module, and several pre-defined sub-expert networks corresponding to different building engineering specialties. The building backbone network is used to obtain the subdivided domain feature vector corresponding to the first question-answer matching pair; and to obtain the descriptive text corresponding to the first question-answer matching pair corresponding to the descriptive text feature vector. The gated routing module calculates the activation probability of each sub-expert network based on the feature vector of the subdivided domain, and selects several sub-expert networks to be activated based on the activation probabilities. The sub-expert network to be activated is used to receive the subdivided domain feature vector and output the corresponding professional domain description vector of the architectural engineering profession; based on the activation probability output by the gated routing module, the subdivided domain feature vectors output by all the sub-expert networks to be activated are weighted and fused to obtain the description text feature vector.

5. The method according to claim 4, characterized in that, Also includes: At the third time interval, the building hybrid expert model is fine-tuned using all the third question-and-answer matching pairs.

6. The method according to claim 1, characterized in that, The second question-and-answer matching pair is stored in a preset short-term memory database; the short-term memory database is cleared every second time interval; the causal reliability screening of the second question-and-answer matching pairs generated within the second time interval to obtain the third question-and-answer matching pair includes: For all second question-and-answer matching pairs stored in the short-term memory database, the parameters corresponding to each layer node are determined in a preset multi-layer Bayesian network based on the second question-and-answer matching pairs to obtain the Bayesian parameter sequence corresponding to the second question-and-answer matching pairs; the multi-layer Bayesian network includes: hazard type layer, risk factor layer, treatment measure layer, acceptance standard layer, and treatment effect layer; Calculate the causal strength and causal closure of the second question-answer matching pair based on the Bayesian parameter sequence and the preset historical parameter sequence. The score corresponding to the second question-answer matching pair is determined based on the causal strength and the causal closure. Filter out ratings that are greater than or equal to a preset rating threshold; The second question-and-answer matching pair corresponding to the selected score is determined as the third question-and-answer matching pair.

7. The method according to claim 1, characterized in that, The third question-and-answer matching pair is stored in a preset long-term memory database; the structured knowledge data is stored in a preset full lifecycle database; the long-term memory database is cleared after the structured knowledge data is generated.

8. A device for storing knowledge about construction projects, characterized in that, include: The first generation module is configured to acquire the raw project data to be processed daily after the construction project starts. Generate the first question-and-answer matching pair corresponding to the original data of the project; The correction module is configured to perform knowledge correction on the first question-and-answer matching pair generated within the first time period at intervals to obtain the second question-and-answer matching pair. The filtering module is configured to perform causal credibility filtering on the second question-and-answer matching pairs generated within the second time period at intervals to obtain the third question-and-answer matching pairs. The second generation module is configured to generate structured knowledge data corresponding to the building project based on all third question-answer matching pairs of the building project after the building project is completed.

9. An electronic device comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to, when executing the program instructions, perform the method for storing building project knowledge as described in any one of claims 1 to 7.

10. A storage medium storing program instructions, characterized in that, When the program instructions are executed, they perform the method for storing building project knowledge as described in any one of claims 1 to 7.

Images