An engineering safety inspection method, system equipment, and medium based on knowledge base and multimodal large model.

Abstract

This invention relates to the field of engineering safety management technology, specifically to an engineering safety inspection method, system equipment, and medium based on a knowledge base and a multimodal large model. First, based on the set inspection environment, key safety categories, and matching regulatory clauses retrieved from the knowledge base, structured prompts are generated. Then, based on collected on-site photos and the generated structured prompts, the multimodal large model is input to identify potential hazards. Finally, work orders are generated and dispatched based on the hazard information, and typical cases are obtained based on the recorded rectification progress and added to the safety knowledge base, achieving accurate and standardized matching.

Description

Technical Field

[0001] This invention relates to the field of engineering safety management technology, and more specifically, to an engineering safety inspection method, system equipment, and medium based on a knowledge base and a multimodal large model. Background Technology

[0002] Safety management at construction sites is a crucial aspect of ensuring project quality and the safety of personnel. Traditional safety inspections rely on manual patrols, which suffer from low efficiency, significant human error, and incomplete hazard mitigation. While intelligent safety management systems have developed rapidly in recent years, they still have significant technical shortcomings, as detailed below: For example, Chinese invention patents with publication number "CN121390138B" titled "A Construction Site Safety Management Method and System Based on Enhanced Retrieval Generation" and publication number "CN120495954A" titled "A Construction Site Safety Monitoring Method and System Based on Large Model and RAG" use a visual language large model to identify potential hazards and combine it with RAG technology to retrieve specifications from a knowledge base, automatically generating safety logs or responses. However, such systems directly analyze images using a general model or only retrieve specifications in the post-processing stage, lacking a clear understanding of the current inspection scenario and safety concerns. This leads to vague retrieval and reasoning directions, making it difficult to accurately match detailed specification clauses for specific scenarios, resulting in weak relevance and enforceability of the generated content.

[0003] For example, the Chinese invention patent with publication number "CN121303862A" and titled "Intelligent Analysis and Closed-Loop Management System for Multimodal Inspection Data of Construction Sites" collects data through a camera array and combines it with a preset rule base and machine learning model for analysis and evaluation. However, this type of model is trained for specific scenarios and requires the deployment of fixed camera arrays, sensor networks, and other hardware, with parameters adapted to fixed scenarios. Once the scenario or safety focus changes, redeployment and retraining are required, resulting in high costs, long cycles, and poor scenario portability. Summary of the Invention

[0004] This invention addresses the problems of low efficiency, large human error, and incomplete hazard handling in existing engineering safety inspections that rely on manual patrols. It proposes an engineering safety inspection method, system, equipment, and medium based on a knowledge base and a multimodal large-scale model. First, based on the set inspection environment, key safety categories, and matching regulatory clauses retrieved from the knowledge base, structured prompts are generated. Then, collected on-site photos and the generated structured prompts are input into the constructed multimodal large-scale model to identify hazard information. Finally, work orders are generated and dispatched based on the hazard information, and typical cases are obtained based on the recorded rectification progress and added to the safety knowledge base, achieving accurate and standardized matching.

[0005] The specific implementation details of this invention are as follows: An engineering safety inspection method based on a knowledge base and a multimodal large model includes the following steps: Step S1: Generate structured prompts by setting the inspection environment, key security categories, and retrieving matching regulatory clauses from the knowledge security base; Step S2: Input the collected on-site photos and generated structured prompts into the constructed multimodal large model to identify potential hazards; Step S3: Generate and dispatch work orders based on hazard information, obtain typical cases based on the recorded rectification progress, and add them to the safety knowledge base.

[0006] To better realize the present invention, step S1 further includes the following steps: Step S11: Send the set scenario and collected on-site data to the security knowledge base; the scenario includes space type, security category, and inspection focus; the on-site data includes on-site photos and on-site videos; Step S12: Retrieve matching specification clauses from the security knowledge base and compose structured prompts based on the set parameters.

[0007] To better realize the present invention, step S2 further includes the following steps: Step S21: Obtain on-site data and scenario specifications from the safety knowledge base, and construct a quadruple pair with the set natural language questions and standard hazard descriptions to obtain quadruple training data; Step S22: Use the open-source visual language large model as the base model to build a multimodal large model, and call the lightweight fine-tuning method according to the training data of the quadruples and the constructed loss function to obtain the optimized multimodal large model; Step S23: Input the collected field data and generated structured prompts into the optimized multimodal large model to obtain hazard information.

[0008] To better implement the present invention, the specific operation of step S21 is as follows: obtain on-site data and scenario specifications from the safety knowledge base, and construct a quadruple with the set natural language questions and standard hazard descriptions to obtain quadruple training data; wherein, the standard hazard description is a standard hazard description marked by safety experts based on historical inspection conclusions.

[0009] To better realize the present invention, step S22 further includes the following steps: Step S221: Use the open-source visual language large model as the base model for the multimodal large model; Step S222: Construct a loss function based on the vocabulary size of the multimodal large model, the one-hot encoding of the real word labels, and the prediction probability of the multimodal large model; Step S223: Minimize the loss by backpropagation based on the constructed loss function, and obtain the optimized multimodal large model based on the training data of the quadruples.

[0010] To better realize the present invention, step S2 further includes: Step S24: Evaluate the trained multimodal large model based on the acquired historical inspection data; evaluation indicators include the accuracy of hazard identification, the correctness of matching standard clauses, and the standardization of output format.

[0011] To better realize the present invention, the specific operation of step S3 is further as follows: Step S31: Parse the hazard information into a unified format and send it to the work order system; Step S32: Generate a work order based on the parsed hazard information and dispatch it according to the work order category; Step S33: Obtain typical cases based on the recorded rectification progress and add them to the safety knowledge base.

[0012] Based on the aforementioned engineering safety inspection method based on a knowledge base and a multimodal large model, and to better realize this invention, a further engineering safety inspection system based on a knowledge base and a multimodal large model is proposed to execute the aforementioned engineering safety inspection method based on a knowledge base and a multimodal large model; including a retrieval and generation unit, a hazard identification unit, and a work order dispatch unit; The retrieval generation unit is used to generate structured prompts based on the set inspection environment, key security categories, and matching regulatory clauses retrieved from the knowledge security base. The hazard identification unit is used to identify hazard information by inputting the collected on-site photos and generated structured prompts into the constructed multimodal large model; The work order dispatch unit is used to generate and dispatch work orders based on hazard information, and to obtain typical cases based on the recorded rectification progress and add them to the safety knowledge base.

[0013] Based on the aforementioned engineering security inspection method based on a knowledge base and a multimodal large model, to better realize this invention, an electronic device is further proposed, including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, the aforementioned engineering security inspection method based on a knowledge base and a multimodal large model is implemented.

[0014] Based on the aforementioned engineering security inspection method based on a knowledge base and a multimodal large model, and to better realize this invention, a computer-readable storage medium is further proposed, wherein computer instructions are stored on the computer-readable storage medium; when the computer instructions are executed on the aforementioned electronic device, the aforementioned engineering security inspection method based on a knowledge base and a multimodal large model is realized.

[0015] The present invention has the following beneficial effects: (1) This invention uses preset space types and security categories as strong constraints for RAG retrieval, making the retrieval results highly focused on the relevant specifications of the current scenario, thus avoiding irrelevant content caused by generalized retrieval. The retrieved specifications are injected into the large model as prompt words, so that the large model has clear clause basis when recognizing, and the output rectification suggestions directly cite the specification number, which greatly improves the executability.

[0016] (2) By constructing quadruples of training data and loss function, this invention forces the model to establish a mapping relationship between the characteristics of hidden dangers and the requirements of the standard during the model training stage. When the model is actually reasoning, it can check the constraint framework one by one like a safety expert who is proficient in the standard, which significantly improves the accuracy and logical rigor of the description of hidden dangers.

[0017] (3) The large model of this invention does not memorize any specifications; all specifications are stored in a knowledge base and retrieved on demand using RAG technology. When it is necessary to adapt to a new scenario or a new specification, the knowledge base maintainer only needs to add a specification entry to the knowledge base, and the RAG retrieval unit can automatically retrieve the specification and include it in the prompt words the next time. The large model can identify potential hazards based on the new specifications without any adjustments or retraining. This greatly reduces the deployment cost and cycle of the system in new scenarios.

[0018] (4) This invention directly drives work order management with the recognition results. Typical cases that pass the acceptance are manually confirmed and then entered into the database, becoming part of the RAG retrieval corpus, so that the prompts for subsequent checks can include historical experience. As the knowledge base becomes richer with the use of the system, the accuracy and coverage of RAG retrieval continue to improve, forming a positive feedback loop.

[0019] (5) This invention utilizes professional knowledge to set up the inspection environment and focus points, while the system is responsible for efficient retrieval of standards and large-scale image recognition. Humans and machines perform their respective duties and complement each other's strengths. Compared with a fully automated system, this invention is more flexible and controllable; compared with purely manual inspection, this solution significantly improves efficiency and standardization. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart of the engineering safety inspection method based on a knowledge base and a multimodal large model provided by the present invention.

[0021] Figure 2 This is a schematic diagram of the overall process provided for an embodiment of the present invention. Detailed Implementation

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments, and therefore should not be regarded as a limitation on the scope of protection. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set up," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] Example 1: This embodiment proposes an engineering safety inspection method based on a knowledge base and a multimodal large model, which specifically includes the following steps: Step S1: Generate structured prompts by setting the inspection environment, key security categories, and retrieving matching regulatory clauses from the knowledge security base; Step S1 specifically includes the following steps: Step S11: Send the set scenario and collected on-site data to the security knowledge base; the scenario includes space type, security category, and inspection focus; the on-site data includes on-site photos and on-site videos; Step S12: Retrieve matching specification clauses from the security knowledge base and compose structured prompts based on the set parameters.

[0025] Step S2: Input the collected on-site photos and generated structured prompts into the constructed multimodal large model to identify potential hazards; Step S2 specifically includes the following steps: Step S21: Obtain on-site data and scenario specifications from the safety knowledge base, and construct a quadruple pair with the set natural language questions and standard hazard descriptions to obtain quadruple training data; The specific operation of step S21 is as follows: obtain on-site data and scenario specifications from the safety knowledge base, and construct a four-tuple with the set natural language questions and standard hazard descriptions to obtain four-tuple training data; wherein, the standard hazard descriptions are standard hazard descriptions marked by safety experts based on historical inspection conclusions.

[0026] Step S22: Use the open-source visual language large model as the base model to build a multimodal large model, and call the lightweight fine-tuning method based on the training data of the quadruples and the constructed loss function to obtain the optimized multimodal large model.

[0027] Step S22 specifically includes the following steps: Step S221: Use the open-source visual language large model as the base model for the multimodal large model; Step S222: Construct a loss function based on the vocabulary size of the multimodal large model, the one-hot encoding of the real word labels, and the prediction probability of the multimodal large model; Step S223: Minimize the loss by backpropagation based on the constructed loss function, and obtain the optimized multimodal large model based on the training data of the quadruples.

[0028] Step S23: Input the collected field data and generated structured prompts into the optimized multimodal large model to obtain hazard information; Step S24: Evaluate the trained multimodal large model based on the acquired historical inspection data; evaluation indicators include the accuracy of hazard identification, the correctness of matching standard clauses, and the standardization of output format.

[0029] Step S3: Generate and dispatch work orders based on hazard information, obtain typical cases based on the recorded rectification progress, and add them to the safety knowledge base.

[0030] The specific operation of step S3 is as follows: Step S31: Parse the hazard information into a unified format and send it to the work order system; Step S32: Generate a work order based on the parsed hazard information and dispatch it according to the work order category; Step S33: Obtain typical cases based on the recorded rectification progress and add them to the safety knowledge base.

[0031] Working Principle: This embodiment utilizes a mechanism of first retrieving standards and then guiding identification, allowing the large model to conduct inspections with a benchmark, improving the targeting and professionalism of identification, and achieving scenario-driven accurate identification; it supports flexible data collection on mobile devices, realizing closed-loop handling of the entire process of hazard work orders, and constructing a flexible closed-loop system; through work order closed-loop management and typical case storage mechanisms, rectification experience is systematically added to the knowledge base, realizing continuous accumulation of knowledge assets, and solving the problems of non-closed-loop hazard handling and difficulty in knowledge accumulation; by embedding a standard reading algorithm in the hazard identification unit, the model algorithm is forced to use specific standard clauses as the reasoning constraint framework, realizing accurate identification from drawing pictures to using standards as the benchmark.

[0032] Example 2: This embodiment is based on the above embodiment 1, such as... Figure 1 , Figure 2 As shown, a specific embodiment will be described in detail.

[0033] like Figure 1 As shown, a detailed explanation is provided, taking the mobile app, security knowledge base, multimodal large model, and work order system as a whole.

[0034] The mobile app module includes a scene setting unit, a multi-type data collection unit, and a hazard viewing and feedback unit; Inspection Scenario Setting Unit: Used to set the space type and safety category before inspection, and optionally add parameters such as inspection focus and process stage; Multiple data acquisition units: Supports data acquisition in four ways: taking photos, recording videos, listening to audio, and reading text. Hazard Inspection and Feedback Unit: Used to display identification results and referenced standard clauses, and supports rectification personnel in receiving work orders and uploading rectification data.

[0035] The security knowledge base module includes a knowledge acquisition and storage unit and a scenario-driven RAG retrieval system. Knowledge acquisition and storage unit: used to store security specifications, regulations, typical cases, etc., organized by multi-dimensional tags such as space type, security category, and keywords; the construction process is: text segmentation → embedding vectorization → storage in a vector database (such as Milvus). Scene-driven RAG search: It receives mobile scene parameters and adopts a hybrid mode of keyword filtering and vector semantic search to accurately retrieve standard terms and generate structured prompts; optional search can be enhanced by focusing on key keywords to ensure that important content is not missed.

[0036] For example, when inspectors set the space type to "elevator shaft" and the safety category to "horizontal protection" in the app, the RAG search unit matches the relevant standard clause from the knowledge base: "4.2.3 Before elevator construction, a safety net should be installed every two floors and no more than 10 meters in the elevator shaft. Isolation and protection facilities should be installed above the construction floor in the elevator shaft." The system combines the original text of this standard with the scene tag to generate the following structured prompt: [Scenario Constraints] Space type: Elevator shaft Safety category: Horizontal protection Key areas of inspection: safety netting, isolation and protection facilities [Standard Clauses] Number: 4.2.3 Content: Before elevator construction, a safety net should be installed every two floors and no more than 10 meters in the elevator shaft. Isolation and protection facilities should be installed above the construction floor in the elevator shaft.

[0037] [Identification Task] Based on the above guidelines, please determine whether the uploaded image violates any regulations. If it complies with the regulations, output "Compliant"; if there are potential safety hazards, output "Safety Hazard Found: Detailed Description, and Specify the Violating Regulations".

[0038] The multimodal large model module includes a hazard identification unit and an identification result generation unit; The hazard identification unit receives structured prompts (containing regulatory clauses) and on-site images / videos, uses a built-in regulatory detailing algorithm for reasoning, and outputs a hazard determination, violation clause, hazard description, and rectification suggestions. The training process of this model is as follows: Data Preparation: Existing on-site photos and corresponding inspection problem records were extracted from the company's historical manual inspection database. Each training data point was constructed as a four-tuple: Scenario Specification – On-site Image – Question – Answer. The scenario specification was provided by the standard clauses in the safety knowledge base; the on-site image was a photo taken during historical inspections; the question was a natural language question guiding the model to compare against the specification (e.g., "Please determine whether there is a safety hazard in the image based on the standard clauses"); the answer was a standard hazard description marked by safety experts based on historical inspection conclusions (including the violating object, violating clause, violation description, and rectification suggestions). All data sources were historical data accumulated from the company's actual engineering safety inspections, ensuring the authenticity and diversity of the training samples.

[0039] Model Training: An open-source visual language model (such as Qwen-VL) is selected as the base model, and lightweight fine-tuning methods (LoRA or P-Tuning v2) are used for training. The model input is "scene specifications + on-site image + question", and the output is the predicted answer. During training, the standard cross-entropy loss function is used. For an answer sequence of length T, the loss function is defined as:

[0040] Where V is the size of the model vocabulary, The one-hot encoding of the real word label at position t (only 1 at the correct word index). Predict the probability that the word i is at position t for the model.

[0041] During training, this loss is minimized through backpropagation, enabling the model to learn to generate accurate hazard descriptions based on specification clauses and images. Through a large amount of training data consisting of four tuples, the model is forced to learn the reasoning logic of "first parsing specification clauses, then identifying the corresponding objects and states from images, and finally outputting structured judgments."

[0042] Model Evaluation: After training, the model's performance is evaluated using a reserved test set (historical inspection data not used in training). Evaluation metrics include: hazard identification accuracy, correct matching rate of regulatory clauses, and output format conformity. The evaluation verifies whether the model has achieved the expected "regulatory-based" identification effect. Only models that pass the evaluation can be deployed online for actual engineering safety inspections.

[0043] Identification Result Generation Unit: This unit parses the structured data (hazard judgment, violation clauses, hazard description, and rectification suggestions) output by the hazard identification unit into a unified format and sends it to the work order system.

[0044] The work order system module includes an automatic work order generation unit, a work order dispatch unit, a rectification process tracking unit, an audit closure loop, and a typical case accumulation unit; Automatic work order generation unit: Generates work orders based on the identification results, clearly citing the violation clause number and original text; Work order dispatch unit: Dispatches work orders to the corresponding responsible persons according to work order type and duty roster; Rectification process tracking unit: Records rectification progress in real time; Review and typical case accumulation unit: After the review and acceptance are passed, the work order is closed; the system prompts to summarize the cases and add them to the database, which is then manually confirmed to enrich the RAG search corpus.

[0045] like Figure 2 As shown, the complete process includes: 1. Inspectors set up scenarios in the app, namely space type and security category; 2. Inspectors should take photos / videos of the scene and submit them along with the photos / videos. 3. The system uses scenario tags and RAG technology to retrieve specification clauses and generate structured prompts; 4. The system inputs the prompt words and images / videos into the hidden danger identification unit of the multimodal large model. The standard reading algorithm inside this unit follows the logic formed in the training phase above, uses the standard clauses in the prompt words as a constraint framework, performs item-by-item comparison with the standard as the yardstick, and outputs the violation clauses and rectification suggestions. 5. The system automatically generates work orders, dispatches, rectification requests, and approvals; 6. Cases that pass the review are manually confirmed and added to the database, becoming new corpus for subsequent RAG searches.

[0046] The other parts of this embodiment are the same as those in Embodiment 1 above, so they will not be described again.

[0047] Example 3: Based on any one of Embodiments 1-2 above, this embodiment proposes an engineering safety inspection system based on a knowledge base and a multimodal large model, used to execute the above-mentioned engineering safety inspection method based on a knowledge base and a multimodal large model; including a retrieval and generation unit, a hazard identification unit, and a work order dispatch unit; The retrieval generation unit is used to generate structured prompts based on the set inspection environment, key security categories, and matching regulatory clauses retrieved from the knowledge security base. The hazard identification unit is used to identify hazard information by inputting the collected on-site photos and generated structured prompts into the constructed multimodal large model; The work order dispatch unit is used to generate and dispatch work orders based on hazard information, and to obtain typical cases based on the recorded rectification progress and add them to the safety knowledge base.

[0048] This embodiment also proposes an electronic device, including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements the above-described engineering security inspection method based on a knowledge base and a multimodal large model.

[0049] This embodiment also proposes a computer-readable storage medium storing computer instructions; when the computer instructions are executed on the aforementioned electronic device, the aforementioned engineering safety inspection method based on a knowledge base and a multimodal large model is implemented.

[0050] The other parts of this embodiment are the same as any one of the above embodiments 1-2, so they will not be described again.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. An engineering safety inspection method based on a knowledge base and a multimodal large model, characterized in that, Specifically, the following steps are included: Step S1: Generate structured prompts based on the set inspection environment, key security categories, and matching regulatory clauses retrieved from the knowledge security base; Step S2: Input the collected on-site photos and generated structured prompts into the constructed multimodal large model to identify potential hazards; Step S3: Generate and dispatch work orders based on hazard information, and obtain typical cases based on the recorded rectification progress and add them to the safety knowledge base.

2. The engineering safety inspection method based on a knowledge base and a multimodal large model according to claim 1, characterized in that, Step S1 specifically includes the following steps: Step S11: Send the set scenario and collected on-site data to the security knowledge base; the scenario includes space type, security category, and inspection focus; the on-site data includes on-site photos and on-site videos; Step S12: Obtain the matching specification clauses from the security knowledge base and generate structured prompt words based on the set parameters.

3. The engineering safety inspection method based on a knowledge base and a multimodal large model according to claim 1, characterized in that, Step S2 specifically includes the following steps: Step S21: Obtain on-site data and scenario specifications from the safety knowledge base, and construct a quadruple pair with the set natural language questions and standard hazard descriptions to obtain quadruple training data; Step S22: Use the open-source visual language large model as the base model to build a multimodal large model, and call the lightweight fine-tuning method according to the training data of the quadruples and the constructed loss function to obtain the optimized multimodal large model; Step S23: Input the collected field data and generated structured prompts into the optimized multimodal large model to obtain hazard information.

4. The engineering safety inspection method based on a knowledge base and a multimodal large model according to claim 3, characterized in that, The specific operation of step S21 is as follows: obtain on-site data and scenario specifications from the safety knowledge base, and construct a four-tuple with the set natural language questions and standard hazard descriptions to obtain four-tuple training data; wherein, the standard hazard descriptions are standard hazard descriptions marked by safety experts based on historical inspection conclusions.

5. The engineering safety inspection method based on a knowledge base and a multimodal large model according to claim 3, characterized in that, Step S22 specifically includes the following steps: Step S221: Use the open-source visual language large model as the base model for the multimodal large model; Step S222: Construct a loss function based on the vocabulary size of the multimodal large model, the one-hot encoding of the real word labels, and the prediction probability of the multimodal large model; Step S223: Minimize the loss by backpropagation based on the constructed loss function, and obtain the optimized multimodal large model based on the training data of the quadruples.

6. The engineering safety inspection method based on a knowledge base and a multimodal large model according to claim 3, characterized in that, Step S2 further includes: Step S24: Evaluate the trained multimodal large model based on the acquired historical inspection data; evaluation indicators include the accuracy of hazard identification, the correctness of matching standard clauses, and the standardization of output format.

7. The engineering safety inspection method based on a knowledge base and a multimodal large model according to claim 1, characterized in that, The specific operation of step S3 is as follows: Step S31: Parse the hazard information into a unified format and send it to the work order system; Step S32: Generate a work order based on the parsed hazard information and dispatch it according to the work order category; Step S33: Obtain typical cases based on the recorded rectification progress and add them to the safety knowledge base.

8. An engineering safety inspection system based on a knowledge base and a multimodal large model, used to execute the engineering safety inspection method based on a knowledge base and a multimodal large model as described in claim 1; characterized in that, It includes a retrieval and generation unit, a hazard identification unit, and a work order dispatch unit; The retrieval generation unit is used to generate structured prompts based on the set inspection environment, key security categories, and matching regulatory clauses retrieved from the knowledge security base. The hazard identification unit is used to identify hazard information by inputting the collected on-site photos and generated structured prompts into the constructed multimodal large model; The work order dispatch unit is used to generate and dispatch work orders based on hazard information, and to obtain typical cases based on the recorded rectification progress and add them to the safety knowledge base.

9. An electronic device, characterized in that, It includes a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements an engineering safety inspection method based on a knowledge base and a multimodal large model as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions; when the computer instructions are executed on the electronic device as described in claim 9, they implement the engineering safety inspection method based on a knowledge base and a multimodal large model as described in any one of claims 1-7.

Images