High-altitude falling construction accident risk prediction method and device based on affair map

Abstract

The invention discloses a high-altitude falling construction accident risk prediction method and device based on a affair atlas, and the method comprises the steps: carrying out the event element extraction of a data entry through a BERT pre-training model after fine adjustment, and converting a front text paragraph and a rear text paragraph of an event relation into a structured event based on the event element; generalization and co-reference event resolution are carried out on the structured events, then the transition probability is calculated, and the event atlas is generated. And for a to-be-predicted construction site, representing an event of the construction site as a real triple, matching a corresponding event chain from the affair map, calculating a risk value of the event chain, and then generating risk reasoning and decision suggestions of high-altitude falling accidents through a large language model. According to the method, on the basis of accident report real data and safety specification documents, a scientific basis is provided for high-altitude falling accident risk assessment by combining a affair graph and a large language model technology, and intelligent support is provided for field safety management decision.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of construction safety management, and particularly relates to a high-altitude falling construction accident risk prediction method and device based on a matter graph. BACKGROUND

[0002] The construction industry plays a significant role in new urbanization construction. However, while producing at a high speed, the construction industry also faces problems such as frequent accidents and high risks. High-altitude falling accidents often cause serious personnel casualties and economic losses, and safety management is crucial and cannot be ignored.

[0003] The rapid development of artificial intelligence technology provides many new solutions for construction safety management. Among them, the matter graph and large language model technology have been applied to the risk management of construction accidents due to their strong visualization ability, good interactive experience, and high degree of intelligence. For example: extracting event knowledge from construction safety accident reports to build a matter graph that is information centralized and knowledge structured; based on a large language model, interpreting accident reports, safety specifications and other text content to provide help for safety decision-making. However, the above technologies still have the following limitations:

[0004] (1) The construction of the matter graph is relatively tedious, such as using a manual construction method, which requires a large amount of labor and time costs.

[0005] (2) Existing matter graphs only consider the time sequence, cause and effect of events, and cannot handle complex scenarios such as multiple causes and one effect, and event turning points; at the same time, most existing matter graphs use subject-predicate-object triples to represent events, and cannot analyze the impact of different event elements on event development and accident consequences.

[0006] (3) Conventional large language models are difficult to analyze strong professional tasks in depth, and their training data do not cover the construction safety field, lacking deep understanding of professional knowledge such as safety specifications, cause theory, and accident mechanism, resulting in less in-depth and accurate analysis in the specific scenario of high-altitude falling accidents.

[0007] (4) Large language models lack strict rule constraint mechanisms and face hallucination problems, which cannot guarantee the accuracy and reliability of the generated content. In tasks involving personnel safety and accident prevention, it is difficult to directly handle important decision-making tasks that require high reliability, and professional knowledge and rules are also needed to limit and constrain the generated content. SUMMARY

[0008] The purpose of the present application is to provide a high-altitude falling construction accident risk prediction method and device based on a matter logic graph. An automatic method is used to construct a high-altitude falling matter logic graph, which is used as an external knowledge base of a large language model, thereby enhancing the content generation and risk reasoning ability of the large language model, and providing an innovative technical path to solve the above challenges.

[0009] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0010] A high-altitude falling construction accident risk prediction method based on a matter logic graph, comprising:

[0011] Based on the event relationship and its corresponding keyword group and regular template, data entries including event relationship and context paragraphs before and after the event are extracted from high-altitude falling accident reports;

[0012] The fine-tuned BERT pre-training model is used to extract event elements from the data entries, and the context paragraphs before and after the event relationship are converted into structured events based on event elements, and a CSV file in the format of "structured event-event relationship-structured event" is obtained;

[0013] The structured events in the CSV file are event generalized to obtain text embedding vectors after generalization, and then the cosine similarity between different text embedding vectors is calculated. Events with a similarity greater than a threshold value are extracted as co-reference events, and the co-reference events are resolved to obtain a list of generalized structured events;

[0014] The occurrence frequency of each event in the generalized event list is counted, and the transition probability of the event relationship is calculated;

[0015] Different colors of nodes are used to represent events and event elements, and different colors of edges are used to represent event relationships. The transition probability of the event relationship is labeled on the edge to generate a matter logic graph;

[0016] For a construction site to be predicted, the events of the construction site are represented as real triples, the event triples closest to the real events of the construction site are retrieved from the matter logic graph, the event chain where they are located is matched in the matter logic graph, and the risk value of the event chain is calculated;

[0017] According to the risk value, the event chain is filtered, the construction background, the real event triple of the construction site, the filtered event chain and its risk value are spliced into prompt words, and the risk reasoning and decision-making suggestions for the occurrence of a high-altitude falling accident are generated by a large language model.

[0018] As a preferred embodiment, the event elements include subject, predicate, object, time, location, falling height, and personnel casualties.

[0019] As a preferred embodiment, the event generalization of the structured events in the CSV file comprises:

[0020] The structured events in the CSV file are textually embedded, and only the subject, predicate and object three event elements of the structured events are retained in the embedding process.

[0021] Preferably, the co-reference event resolution comprises:

[0022] All co-reference events pointing to the same event form a co-reference event set, and for all events in the co-reference event set, their numbers are updated to the minimum value of the numbers in the set, and their text representations are updated to the text representation of the event with the minimum number in the set. Finally, all co-reference events pointing to the same event only retain one event record with uniform number and text representation.

[0023] Preferably, the high-altitude falling construction accident risk prediction method based on the matter-logic graph further comprises:

[0024] The text embedding vector of each event in the filtered event chain is obtained, and the corresponding clause in the construction safety specification is matched through a large language model to obtain a matched safety specification clause.

[0025] Preferably, the risk reasoning and decision suggestion for the occurrence of the high-altitude falling accident generated by the large language model comprises:

[0026] The construction background, the real event triplets of the construction site, the filtered event chain and its risk value, and the matched safety specification clause are spliced into a prompt word, and the risk reasoning and decision suggestion for the occurrence of the high-altitude falling accident generated by the large language model.

[0027] The application also proposes a high-altitude falling construction accident risk prediction device based on a matter-logic graph, comprising a processor and a memory storing a plurality of computer instructions, which are executed by the processor to implement the steps of the above method.

[0028] The high-altitude falling construction accident risk prediction method and device based on the matter-logic graph proposed in the application enhance the professional analysis capability of the large language model in the field of construction safety, and on the other hand, constrain the generated content to comply with the standard, so that the generation capability and reasoning capability of the large language model are both enhanced, and finally the risk reasoning and decision support for the high-altitude falling accident are realized. The application is based on real accident report data and safety specification documents, combined with advanced matter-logic graph and large language model technology, to provide scientific basis for high-altitude falling accident risk assessment and intelligent support for on-site safety management decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The flowchart of the high-altitude falling construction accident risk prediction method based on the matter-logic graph of the application.

[0030] Figure 2 A reasoning graph diagram generated for an embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0032] One embodiment of the present application, as shown in Figure 1 provides a high-altitude falling construction accident risk prediction method based on a reasoning graph, comprising:

[0033] Step S1, based on the event relationship and its corresponding keyword group and regular template, extracting data entries including event relationship and context paragraphs from high-altitude falling accident reports.

[0034] This embodiment first collects construction accident reports and construction safety specifications, sorts out the data entries related to high-altitude falling accidents, obtains accident reports and safety specifications related to high-altitude falling accidents, and stores them as CSV files. The content of the accident report includes the complete course of the accident, treatment measures, etc. The safety specification is the current effective version and is related to the high-altitude falling construction accident. Subsequent accident reports will be used as data sources for the construction of a multi-dimensional reasoning graph, and safety specifications will be used as external knowledge for the generation of large language model retrieval enhancement.

[0035] Generally, in order to establish a reasoning graph, event elements and event relationships need to be defined. This embodiment analyzes the event definitions of existing research, combines expert experience to analyze the event information needs of high-altitude falling accident risk reasoning, and considers the semantic characteristics of construction safety accident reports, introduces the event definition of multi-dimensional reasoning graph, which specifically includes two key contents of event elements and event relationships.

[0036] Among them, 7 types of event elements of high-altitude falling accidents are defined, and events are structured and represented using event elements: ① Subject: as the subject of people or things, is the object of event description; ② Predicate: describes the nature or state of the subject, each event has and only one predicate, which is the determining component of event attributes; ③ Object: the receiver of the subject's action; ④ Time: the occurrence time of the event; ⑤ Place: the occurrence place of the event; ⑥ Falling height: the vertical height of the fall; ⑦ Personnel casualties: the number and degree of personnel injuries involved in the falling accident. Specific examples of 7 types of event elements are given:

[0037] Table

[0038]

[0039] It should be noted that the definition of the event element in this embodiment is based on the high-altitude falling construction accident, and for the prediction of the risk of other similar accidents, the corresponding event elements are increased / decreased, and details are not repeated here.

[0040] Then, multi-dimensional event relationships are also defined. In the development process of the high-altitude falling accident, the relationship between events is divided into 5 categories: ① follow-up relationship: event 1 occurs before event 2; ② cause-and-effect relationship: event 1 leads to the occurrence of event 2; ③ multiple cause and effect relationship: coupling of multiple events, for example, event 1 and event 2, together lead to the occurrence of event 3; ④ turning point relationship: event 2 represents the opposite meaning of event 1; ⑤ co-reference relationship: event 1 and event 2 represent the same event. Specific examples of the 5 types of event relationships are given:

[0041] Table

[0042]

[0043] This embodiment realizes the recognition and extraction of multi-dimensional event relationships through keyword groups and regular templates, as shown in Table 3. Specifically, the natural language processing technology is used to traverse the high-altitude falling accident report text obtained in step S1 to identify the positions of the 28 keyword groups in the accident report, and then match the corresponding regular expression templates, and on this basis, the text segments before and after the keyword groups are extracted to form a unified structure containing event relationships: text segment, event relationship, text segment.

[0044] Table

[0045]

[0046] Step S2, using the fine-tuned BERT pre-training model to extract event elements from the data entries, converting the text segments before and after the event relationship into structured events based on event elements, and obtaining a CSV file in the format of “structured event-event relationship-structured event”.

[0047] The BERT (Bidirectional Encoder Representations from Transformers) architecture is a deep learning model based on a Transformer encoder, mainly used for natural language processing tasks. In this embodiment, a BERT pre-training model (BERT-base-NER) is used to perform the named entity recognition task. The BERT-base-NER pre-training model is fine-tuned using a fine-tuning dataset to obtain a fine-tuned BERT pre-training model (Modified-BERT-NER), which can extract various event elements in Table 1 above.

[0048] In this embodiment, part of the data entries obtained from step S1 are randomly selected, and the 7 types of event elements are labeled and converted into a fine-tuning event element dataset in BIO format.

[0049] During fine-tuning, a semi-supervised learning training method is used, a cross-entropy loss function and an AdamW optimizer are specified, and the BERT-base-NER model is iteratively trained on the fine-tuning event element dataset. The score performance of the model in each iteration round is recorded, and if the score of this round is higher than that of the last round, the model parameters are updated until the model converges to complete the fine-tuning training, thereby creating a Modified-BERT-NER model that can extract the aforementioned 7 types of event elements from high-altitude fall accident report texts.

[0050] The Modified-BERT-NER model of this embodiment can automatically identify 7 types of event elements from the text paragraphs of the accident report, convert the text paragraphs into structured events in the form of [subject; predicate; object; time; location; fall height; personnel casualties], and output as a CSV file in the format of "structured event, event relationship, structured event". The CSV file is a pure text file used to store tabular data.

[0051] In step S3, the structured events in the CSV file are event generalized to obtain generalized text embedding vectors, and then the cosine similarity between different text embedding vectors is calculated. Events with a similarity greater than a threshold value are extracted as co-reference events, and co-reference event resolution is performed to obtain a list of generalized structured events.

[0052] The embodiment performs text embedding on the structured events in the CSV file to obtain corresponding text embedding vectors, thereby converting the natural language text into vectorized representation in a high-dimensional space. The embedding process can be performed using a text embedding vector model, such as text-embedding-v4. Only the subject, predicate, and object of the structured event are retained in the embedding process, as shown in equation (1), to achieve event generalization, so that the generalization ability of the subsequent established knowledge graph is stronger.

[0053]

[0054] In the formula, represents the text embedding vector of the event . respectively represent the subject, predicate, and object of the event .

[0055] The cosine similarity between the text embedding vectors of different events is calculated according to equation (2), and a similarity threshold is set. Events with a similarity greater than the similarity threshold are extracted as co-reference events. All co-reference events pointing to the same event form a co-reference event set. For all events in the co-reference event set, their numbers are updated to the minimum number in the set, and their text representations are updated to the text representation of the event with the minimum number in the set. It should be noted that the text representation here is the natural language description text of the structured event in the CSV file, which can be represented by , that is, the structured event represented by the subject, predicate, and object. Through this way, the co-reference events are resolved, and finally all co-reference events pointing to the same event only retain one event record with a unified number and text representation, obtaining the generalized structured event list. In the co-reference event resolution process, the event relationship between the co-reference events is also resolved.

[0056]

[0057] In the formula, represents the cosine similarity between event vectors . represents the two-norm of the event vector .

[0058] Step S4, count the occurrence frequency of each event in the generalized event list, and calculate the transition probability of the event relationship.

[0059] On the basis of the above work, the occurrence frequency of each event in the generalized event list is counted, and the transition probability of the event relationship is calculated according to equation (3) according to the Bayesian conditional probability theory.

[0060]

[0061] wherein, denotes the known event occurs the probability of occurrence; denotes the event and the event the probability of simultaneous occurrence; denotes the event the probability of occurrence; denotes the event in the event list after the event generalization and co-reference event resolution the frequency of simultaneous occurrence with the event ; denotes the event in the event list after the event generalization and co-reference event resolution the frequency of occurrence.

[0062] When the event and the event jointly cause the event to occur, according to the Bayes multi-condition probability theory, the transfer probability of the double cause-single effect relationship is converted into the transfer probability of the single cause-single effect relationship according to formula (4). Similarly, the transfer probabilities of other multi-cause-single effect event relationships are also similarly converted into the transfer probabilities of single cause-single effect event relationships.

[0063]

[0064] wherein, denotes the known event and the event occurs the probability of occurrence; denotes the event , the event and the event the probability of simultaneous occurrence; The meaning of

[0065] After completing the above steps S3 and S4, two files are output in CSV format: ① the structured event list after the event generalization and co-reference event resolution; and ② the event relationship triplets containing the transfer probability.

[0066] Step S5, different colors of nodes are used to represent events and event elements, and different colors of edges are used to represent event relationships, and the transfer probability of the event relationship is marked on the edge, to generate a fact graph.

[0067] The two CSV files are imported into the Neo4j database through the Cypher LOAD statement, and then the events and event elements are represented by nodes of different colors, and the event relationships are represented by edges of different colors, and the transition probability of the event relationship is marked on the edge. Finally, the multi-dimensional fact graph of the high-altitude falling accident is unfolded. The generated fact graph is shown in Figure 2 It should be noted that different colors are not shown in the graph.

[0068] Step S6, for the construction site to be predicted, the events of the construction site are represented as real triples, the event triples closest to the real events of the construction site are retrieved from the fact graph, the event chain where they are located is matched in the fact graph, and the risk value of the event chain is calculated.

[0069] Specifically, according to the real scene, the events of the construction site are represented as real triple structures: “real event 1, event relationship, real event 2”, that is . Then, the text vectorization algorithm of the large language model (such as text-embedding-v4) is used to obtain the text vector of the real event . Then, first calculate the text similarity between the real event and each event in the generalized event list according to formula (5), and then calculate the comprehensive similarity between the real event triple and the event triple in the fact graph according to formula (6).

[0070]

[0071]

[0072] In the formula, represents the text similarity between the event and each event in the generalized event list; represents the text vector of the event ; represents the text vector of each event in the generalized event list; represents the comprehensive similarity between the real event triple of the construction site and the event triple in the fact graph; respectively represent the similarity weights of ; respectively represent the similarity of calculated according to formula (5) and each event, event relationship in the fact graph.

[0073] The comprehensive similarity the maximum value of the risk value of the event chain, which corresponds to the event triple closest to the real event in the construction site retrieved from the event graph.

[0074] Then, according to the obtained construction site real event mapped into the event triple in the event graph, the event chain where the triple is located in the event graph is matched, the personnel casualty event element in the event chain is identified, the medical project required for personnel casualty is retrieved, and the cost of the medical project is taken as the fall accident loss value, and the risk value of the above event chain is calculated according to formula (7).

[0075]

[0076] In the formula, the risk value of the event chain; the transition probability of all event relationships in the event chain; the accident loss value of the event chain.

[0077] Finally, the 5 event chains with the maximum risk value are output, and the risk reasoning of the construction site real event leading to the fall accident is completed.

[0078] Step S7, screening the event chain according to the risk value, concatenating the construction background, the construction site real event triple, the screened event chain and its risk value into a prompt word, and generating the risk reasoning and decision suggestion of the fall accident through a large language model.

[0079] In this embodiment, a large language model is used for risk reasoning and decision suggestion, for example, qwen3-max can also be used, and other large language models can also be used, which are not limited here. In specific practice, the construction background, the construction site real event triple, the screened event chain and its risk value are concatenated into a prompt word in order, and input into a large language model to obtain the risk reasoning and decision suggestion of the fall accident.

[0080] Another embodiment of the present application, based on the above embodiment, further comprises:

[0081] Obtaining the text embedding vector of each event in the screened event chain, matching the corresponding clause in the construction safety specification through a large language model, and obtaining the matched safety specification clause.

[0082] Further, the large language model is used to generate the risk reasoning and decision suggestion of the fall accident, comprising:

[0083] The construction background, the real event triplets of the construction site, the filtered event chains, the risk values and the matched safety specification clauses are spliced into prompt words, and a large language model is used to generate risk reasoning and decision suggestions for high-altitude falling accidents.

[0084] Specifically, all events in the obtained 5 event chains are stored as a CSV file, and text embedding operation is performed to obtain text vectors of each event in the CSV file. Subsequently, according to the obtained text vectors, corresponding clauses are matched in the construction safety specifications collected in step (1) by the large language model qwen3-max. In the subsequent risk reasoning analysis content generation of the high-altitude falling accident by the large language model qwen3-max, the above matched safety specification clauses will be used as external knowledge for retrieval enhancement generation of the large language model, and also as constraint rules for the large language model to generate high-altitude falling safety management decision suggestions.

[0085] The 5 event chains and their risk values, and the safety specification clauses matched by the accident chains are used as external retrieval knowledge, and the retrieval enhancement generation technology of the large language model qwen3-max is used to output the risk reasoning analysis results of the high-altitude falling accidents caused by the real events in the construction site and the corresponding high-altitude falling safety management decision suggestions in natural language.

[0086] Specifically, by setting the professional role of the large language model qwen3-max, the accuracy of the large language model in the above content generation work is improved, and the construction background, the real event triplets of the construction site, the 5 event chains and their risk values obtained from the event reasoning graph, and the matched construction safety specification clauses are spliced into context and prompt words, so that the large language model qwen3-max can generate risk reasoning and decision suggestions for high-altitude falling accidents according to the retrieval content with constraints.

[0087] According to the literature review and expert experience, the technical solution of the present application proposes an event definition of a high-altitude falling accident multi-dimensional event reasoning graph, which includes 7 types of multi-dimensional event elements: subject, predicate, object, time, place, falling height, and personnel casualty; and 5 types of multi-dimensional event relationships: continuation, one cause and one effect, multiple causes and one effect, transition, and co-reference. According to the personnel casualty and transfer probability, combined with the medical project cost, the specific loss value and risk value of the high-altitude falling accident are calculated. A comprehensive similarity calculation method of real events and event triplets in the event reasoning graph is designed to map the real events to the event reasoning graph. Finally, the specific risk value of the possible high-altitude falling accident in the construction site can be identified and calculated, and the risk reasoning results with strong readability are output in natural language, together with the high-altitude falling accident prevention measures and management suggestions that meet the construction safety specifications.

[0088] In one embodiment, the application also provides a high-altitude falling construction accident risk prediction device based on a matter logic graph, comprising a processor and a memory storing a plurality of computer instructions, which, when executed by the processor, implement the steps of the above method.

[0089] The specific limitations of the high-altitude falling construction accident risk prediction device based on the matter logic graph can be referred to the limitations of the high-altitude falling construction accident risk prediction method based on the matter logic graph in the above, which will not be repeated here. The above high-altitude falling construction accident risk prediction device based on the matter logic graph can be realized by software, hardware and their combination in whole or in part. It can be embedded in the processor in the computer device in hardware form or independent of the processor in the computer device, or it can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the above corresponding operations.

[0090] The memory and the processor are directly or indirectly electrically connected to realize the transmission or interaction of data. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines. The memory stores a computer program that can run on the processor, and the processor realizes the high-altitude falling construction accident risk prediction method based on the matter logic graph in the embodiment of the application by running the computer program stored in the memory.

[0091] The memory can be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read only memory (PROM), an erasable programmable read only memory (EPROM), an electrically erasable programmable read only memory (EEPROM) and the like. The memory is used to store programs, and the processor executes the programs after receiving execution instructions.

[0092] The processor can be an integrated circuit chip with data processing capability. The processor described above can be a general processor, including a central processing unit (CPU), a network processor (NP) and the like. It can realize or execute the methods, steps and logic block diagrams disclosed in the embodiments of the application. The general processor can be a microprocessor or the processor can also be any conventional processor or the like.

[0093] The above-described embodiments are merely illustrative of several embodiments of the present application, which are described in more detail and in a specific and detailed manner, but should not be construed as limiting the scope of the patent. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A high-altitude falling construction accident risk prediction method based on a matter logic map, characterized by, The high-altitude falling construction accident risk prediction method based on the matter graph comprises: Based on the event relationship, the corresponding keyword group and the regular template, data entries including event relationships and before and after text paragraphs are extracted from high-altitude falling accident reports; The event element extraction is performed on the data entries by using the fine-tuned BERT pre-training model, the before and after text paragraphs of the event relationship are converted into structured events based on event elements, and a CSV file in the format of "structured event-event relationship-structured event" is obtained; Event generalization is performed on the structured events in the CSV file to obtain a text embedding vector after generalization, and then the cosine similarity between different text embedding vectors is calculated, events with a similarity greater than a similarity threshold are extracted as co-reference events, and the co-reference events are resolved to obtain a list of generalized structured events; The occurrence frequency of each event in the list of events after generalization is counted, and the transition probability of the event relationship is calculated; Different colors of nodes are used to represent events and event elements, and different colors of edges are used to represent event relationships, and the transition probability of the event relationship is marked on the edge to generate a matter graph; For a construction site to be predicted, the events of the construction site are represented as real triples, the event triples closest to the real events of the construction site are retrieved from the matter graph, the event chains in which the event triples are located are matched in the matter graph, and the risk values of the event chains are calculated; According to the risk values, the event chains are screened, the construction background, the real event triple of the construction site, the screened event chain and the risk value thereof are spliced into prompt words, and a large language model is used to generate risk reasoning and decision suggestions for the occurrence of a high-altitude falling accident. 2.The high fall construction accident risk prediction method based on the matter logic graph according to claim 1, characterized in that, The event elements include subject, predicate, object, time, place, falling height and personnel casualty. 3.The high fall construction accident risk prediction method based on the matter logic graph according to claim 2, characterized in that, The event generalization of the structured events in the CSV file comprises: Text embedding is performed on the structured events in the CSV file, and only the subject, predicate and object of the structured events are retained during the embedding process. 4.The high fall construction accident risk prediction method based on the matter logic graph according to claim 1, characterized in that, The co-reference event resolution comprises: All co-reference events pointing to the same event form a co-reference event set, for all events in the co-reference event set, the numbers of the events are updated to the minimum number in the set, and the text representations of the events are updated to the text representation of the event with the minimum number in the set, and finally all co-reference events pointing to the same event only retain one event record with a unified number and text representation. 5.The high fall construction accident risk prediction method based on the matter logic graph according to claim 1, wherein, The high-altitude falling construction accident risk prediction method based on the matter graph further comprises: Text embedding vectors of events in the screened event chain are obtained, and a large language model is used to match corresponding clauses in construction safety specifications to obtain matched safety specification clauses. 6.The high fall construction accident risk prediction method based on the matter logic graph according to claim 5, characterized in that, The large language model is used to generate risk reasoning and decision suggestions for the occurrence of a high-altitude falling accident. The computer instructions are executed by the processor to implement the steps of the method in any one of claims 1 to 7.

7. A high-altitude falling construction accident risk prediction device based on a matter logic map, comprising a processor and a memory storing a plurality of computer instructions, characterized in that, ​

Images