A safety culture improvement measure generation method and system

Abstract

The application discloses a kind of safety culture promotion measure generation method and system, wherein the method includes identifying safety culture evaluation report, carding safety data, fly inspection investigates possible reason, analyzes project safety culture problem, checks safety culture problem, fly inspection confirms safety culture problem, proposes safety problem promotion measure, checks safety culture promotion measure, confirms safety culture promotion measure feasibility, analyzes measure advantages and disadvantages and generates conclusion report.The present application is based on large language model, character recognition, picture recognition, speech recognition, database technology combined with smart phone and high memory workstation and the like hardware, realizes project safety culture promotion measure intelligent generation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of safety culture improvement. More particularly, the present application relates to a safety culture improvement measure generation method and system. BACKGROUND

[0002] The construction industry, as a pillar of the national economy, plays a crucial role in the development of modern society. The safety production of the construction industry is directly related to the life and health of workers. Therefore, taking effective safety measures and standardizing operations to ensure the safety of workers during the construction process is a basic requirement to protect their rights and well-being. The safety production of the construction industry is of great importance to the stability and sustainable development of society.

[0003] Statistical results show that 79% of accidents are caused by unsafe human behavior. Unsafe human behavior includes violating safety regulations, being careless, making mistakes, etc. These behaviors may be caused by factors such as personal attitude, knowledge level, skill proficiency, and psychological state.

[0004] The cultivation of safety culture plays an important role in reducing the occurrence of unsafe human behavior, which is mainly through the following aspects. First, safety culture helps to establish the safety values of workers. Safety values will help employees recognize the importance of safety and regard it as a top priority. By instilling correct safety values, employees will be more conscious of complying with safety management systems. Second, safety training and education help to cultivate employees' safety awareness and safety skills. Training content can include safety regulations and procedures, hazard identification and control, emergency handling, etc. By increasing employees' understanding of safety knowledge and skills, they will be more familiar with safety requirements and better able to deal with potential safety risks, reducing the probability of unsafe behavior. Third, improve the efficiency of communication and participation. Safety culture encourages open and active communication, as well as employee participation. Organizations should provide multiple communication channels for employees to report potential hazards and make suggestions for improvement. At the same time, organizations should encourage employees to participate in the decision-making process and the development of safety regulations. Through this participation mechanism, employees will feel valued and their voices will be heard, making them more actively participate in safety behavior and reduce the occurrence of unsafe behavior. Fourth, establish incentive mechanisms. Safety culture encourages employees' safety behavior through incentive and reward mechanisms. This can include praising and rewarding individual or team safety achievements, establishing safety performance evaluation and reward systems, etc. By giving positive incentives, employees will be more proactive in complying with safety regulations and actively participating in safety activities, reducing the occurrence of unsafe behavior.

[0005] The current design of safety culture improvement measures includes two parts: diagnosing safety culture problems and designing improvement measures. It mainly relies on experts to conduct research and cooperate with project management teams. The cycle is long, and it depends on experts and cannot activate the project team to actively find their own problems and improve them. There are the following problems: first, the subjective and one-sided nature of expert personal experience and knowledge. The application of expert personal experience and knowledge in safety culture management often has subjectivity and one-sidedness. Each expert's experience and opinion is limited by their personal background, experience, and professional field, so other factors and perspectives may be overlooked. This can lead to biased decision-making, which may overlook other effective safety culture improvement measures and strategies. Second, personal experience cannot meet the systematic and comprehensive needs of safety culture improvement. The design of safety culture improvement measures needs to consider multiple factors such as project organizational structure, work content, and safety standards. Relying on expert personal experience and knowledge may not meet the comprehensive and systematic requirements. Third, expert seminars cannot meet the timeliness and effectiveness requirements of safety culture. In order to avoid the personal factors of experts, when it comes to safety culture improvement measures, expert seminars are often used to brainstorm with project senior management and safety experts. However, this method is difficult to obtain improvement measures in a short time, and the measures proposed by experts may face challenges in actual implementation. SUMMARY

[0006] To at least solve the technical problems described in the background section, the present application provides a safety culture improvement measure generation method and system. The present application can realize intelligent generation of project safety culture improvement measures. In view of this, the present application provides solutions in the following aspects.

[0007] The first aspect of the present application provides a safety culture improvement measure generation method, comprising: step 1, identifying a safety culture evaluation report, extracting text data in the report, then identifying text data related to safety culture, and arranging the identification results to form scores of safety culture items and dimensions and storing, the storage structure includes pusher identity information, safety culture evaluation report, safety culture supplementary information, organizational structure and project safety data; step 2, combing safety data, classifying and analyzing the data stored in the above step except the safety culture evaluation report, outputting and storing safety culture related information, action driving information, value orientation information and safety problem information; step 3, fly inspection of possible reasons, low score safety culture items and safety problem materials are sampled and sent to participants of various identities in the form of web forms, the safety problems include two parts of why the dimension of safety culture has low items and why there are safety problems; step 4, analyzing project safety culture problems, the safety culture evaluation report and safety culture supplementary information content are disassembled according to the dimension, the description of each low score item is obtained, then the low score item description, the corresponding safety culture evaluation item, the safety culture related information, the behavior driving information, the value orientation information, the audio to text information and the corresponding safety problem are input into the expert module, and the possible safety culture problems of the dimension are obtained; step 5, checking safety culture problems, determining which identity people should be given the checking work of different low score item situation description and low score item safety problem; step 6, fly inspection of safety culture problems, the low score item situation description and low score item safety problem output in step 4 are pushed to the participants and the pushers in the form of web forms, the participants in step 5 are pushed in priority, the pushers are all pushed, and the time from receiving the form to opening the form and recording the video is recorded; step 7, proposing safety problem improvement measures, inputting the low score item situation description, the low score item problem and the recognition degree score of each level to the description and the problem, obtaining the low score item improvement measures and the expected effect description of implementation measures; step 8, checking safety culture improvement measures, determining which identity people should be given the checking work of different low score item improvement measures; step 9, confirming the feasibility of safety culture improvement measures, pushing the safety culture improvement measures and the expected effect of implementation measures obtained in step 7 to the pushers in the project and the high ranking personnel obtained in step 8, and to the safety culture experts outside the project; the push includes low score item situation description, low score item safety problem, safety culture improvement measures and expected effect of implementation measures; step 10, analyzing the pros and cons of the measures to generate a conclusion report, by inputting the safety culture improvement measures, the expected effect of implementation measures and the reply opinions obtained in step 9, an evaluation report and a use instruction of the improvement measures are generated for the project manager from multiple angles such as cost, efficiency and feasibility.

[0008] The second aspect of the present application provides a safety culture improvement measure generation system, the safety culture improvement measure generation method, comprising: an information interaction module, a safety culture identification unit, a project insight module, a safety culture expert module, a checking module, and a large language model module; the information interaction module is used for interaction through video, audio, text and file respectively, and converts the voice recorded by the user into text; the safety culture identification module is used for analyzing the safety evaluation report uploaded by the safety culture construction promoter; the project insight module is used for classifying and analyzing the user uploaded data, including safety culture analysis, system analysis, measure analysis and safety problem analysis; the safety culture expert module is used for analyzing safety culture problems, proposing intervention measures and analyzing the advantages and disadvantages of the measures; the checking module is used for checking the safety culture problems and safety culture improvement measures generated by the safety culture expert module; and the large language model module is a deep learning model trained using a large amount of text data, which can generate natural language text or understand the meaning of language text.

[0009] In one embodiment, the conversion of the voice recorded by the user into text specifically includes: S1, a noise reduction step based on frequency domain filtering, converting the audio signal to the frequency domain, filtering through a low-pass filter, and then converting the signal back to the time domain to obtain a first noise reduction result; S2, a noise reduction step based on time domain filtering, performing time domain filtering on the first noise reduction result through a preset filter set to obtain a second noise reduction result, the preset filter set including one or a combination of several of median filtering, Kalman filtering, and wavelet noise reduction; S3, performing feature extraction on the second noise reduction result through a Mel frequency cepstral coefficient (MFCC) or a Mel filter bank energy (FBank) method; S4, using a deep learning speech recognition algorithm to perform speech recognition on the feature extraction result; and S5, calling a large language model of the large language model module and inputting the recognition result into the large language model.

[0010] In one embodiment, the safety culture identification module comprises a Word document identification unit for identifying a user-uploaded safety culture assessment report in Word format (.docx), outputting a docx format document highlighting safety culture related content and an xlsx format document including scores of each item of the safety culture, and storing the highlighted text paragraphs in json format in a cloud server, and not processing pictures in the document; a PPT document identification unit for identifying a user-uploaded safety culture assessment report in PPT format (.pptx), storing the text box content as a word document (.docx) format, storing the chart content as an excel table format, and processing using the Word document identification unit; a PDF document identification unit for identifying a user-uploaded safety culture assessment report in PDF format (.pdf), cutting the text according to paragraphs, inputting each paragraph of text into a large language model module for confirmation, and outputting the results to a word file, and processing using the Word document identification unit; and a detection unit for determining whether a user can proceed to the next step, and detecting the completeness of the file and content when the user has uploaded a safety culture assessment report.

[0011] In one embodiment, the project insight module comprises a safety culture analysis unit, a system analysis unit, a measure analysis unit, and a safety problem analysis unit; the safety culture analysis unit is used to filter safety culture related data obtained from the information interaction module, and classify the data by item using the expert large language model unit in the safety culture expert module; the system analysis unit is used to extract the value judgment method encouraged by the organization, i.e., data related to value orientation, from non-safety culture data obtained from the information interaction module, and classify the data by item using the expert large language model unit in the large language model module; the measure analysis unit is used to extract the behavior encouraged by the organization, i.e., data related to behavior drive, from non-safety culture data obtained from the information interaction module, and classify the data by item using the expert large language model unit in the large language model module; and the safety problem analysis unit is used to extract specific safety problems from non-safety culture related data obtained from the information interaction module, and classify the data by item using the expert large language model unit in the large language model module.

[0012] In one embodiment, the safety culture expert module comprises: a safety culture problem analysis unit for analyzing interaction data to obtain the correlation of safety problems and safety problem possible cause materials and safety culture items; a safety problem improvement measure unit for proposing safety culture improvement measures according to safety culture problem feedback, including screening the authenticity and rationality scores of each safety culture problem, and only transmitting single problems, problem corresponding descriptions, scorer identities and behavior driving data with scores of 3 or more into the expert large language model module; a safety culture improvement measure analysis unit for generating an evaluation report and a usage instruction of the improvement measure for project managers from multiple angles such as cost, efficiency and feasibility, based on the safety culture improvement measure, the expected effect of the implementation measure and the reply opinion; and an expert large language model unit, which is obtained by training a pre-trained language model including BERT, GPT and ChatGLM through fine-tuning technology.

[0013] In one embodiment, the checking module comprises: a safety culture problem checking unit for checking the safety culture problems generated by the safety culture expert module, including calling the expert large language model of the safety culture expert module, comparing the similarity of the fly inspection answers and the safety culture problem related materials, taking the results as the object identity of the information transmission module, calculating and sorting the scores of each non-safety job department, and storing the sorting list; and a safety culture improvement measure checking unit for checking the safety culture measures generated by the safety culture expert module, including calling the expert large language model of the safety culture expert module, comparing the correlation of the fly inspection answers and the safety culture improvement measures, weighting and averaging the rationality and authenticity scores of the results, calculating and sorting the scores of each non-safety job department, and storing the sorting list.

[0014] In one embodiment, the large language model module is used to complete the task of text generation: given a number of generated words in front, the next word that can maximize the sequence probability is calculated, and the word is output as the prediction result; then the model adds the predicted word to the given sequence and repeats the above process to continue predicting the next one until the predicted next word is an end symbol or the required length is reached.

[0015] The present application realizes intelligent generation of project safety culture improvement measures based on a large language model, character recognition, picture recognition, speech recognition, database technology, and intelligent mobile phones (including front camera functions) and high memory workstations and other hardware. BRIEF DESCRIPTION OF DRAWINGS

[0016] The above and other objects, features and advantages of the present exemplary embodiments will become more apparent from the following detailed description read in conjunction with the accompanying drawings, in which like reference numerals refer to like elements throughout. The accompanying drawings, which are shown by way of example, illustrate several embodiments of the present application and, where appropriate, refer to identical or corresponding parts of more than one embodiment. In the drawings:

[0017] Figure 1 is a flow chart illustrating a method and system for generating a safety culture improvement measure according to an embodiment of the present application;

[0018] Figure 2 is a storage structure of step 1 according to an embodiment of the present application;

[0019] Figure 3 is a storage structure of step 2 according to an embodiment of the present application;

[0020] Figure 4 is a storage structure of step 3 according to an embodiment of the present application;

[0021] Figure 5 is a storage structure of step 4 according to an embodiment of the present application;

[0022] Figure 6 is a storage structure of step 6 according to an embodiment of the present application;

[0023] Figure 7 is a safety culture dimension and item according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0025] It should be understood that the terms "first", "second", "third", and "fourth" and the like in the claims, the specification and the drawings of the present application are used to distinguish different objects, and are not used to describe a particular sequence. The terms "include" and "contain" used in the specification and claims of the present application indicate the presence of described features, whole, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, whole, steps, operations, elements, components and / or sets thereof.

[0026] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used in this specification and the appended claims, the singular forms "a," "an" and "the" include plural referents unless the context clearly dictates otherwise. It is further to be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items, and that the term "at least one of' as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0027] As used in this specification and claims, the terms "if' and "when" can each be interpreted to mean "upon determination" or "in response to a determination" or "upon detection" or "in response to a detection," depending on the context. Similarly, the phrase "if determined" or "if detected [a described condition or event]" can be interpreted to mean "upon determination" or "in response to a determination" or "upon detection" or "in response to a detection," depending on the context.

[0028] The specific embodiments of the present application will now be described in detail with reference to the following figures.

[0029] Figure 1 A safety culture improvement measure generation method according to an embodiment of the present application is shown. As shown in Figure 1 The safety culture improvement measure generation method in the present application can be described as including steps 1-10 as shown in

[0030] Step 1, identify the safety culture evaluation report, extract the text data in the report, then identify the text data related to safety culture, and organize the identification results to form the scores of each item and dimension of safety culture and store them, the storage structure includes the identity information of the person who promotes, the safety culture evaluation report, the safety culture supplementary information, the organizational structure and the project safety data;

[0031] Step 2, sort out the safety data, classify and analyze the data stored in the previous step except the safety culture evaluation report, output and store safety culture related information, action driving information, value orientation information and safety problem information;

[0032] Step 3, fly inspection of possible reasons, sample and send the low-score safety culture items and safety problem materials to participants of various identities in the form of a web page form, the safety problems include two parts of why the dimension of safety culture has low items and why there are safety problems;

[0033] Step 4, analyze the safety culture problem, decompose the safety culture evaluation report and safety culture supplementary information content according to the dimension to obtain the description of each low-score item, then input the low-score item description, corresponding safety culture evaluation item, safety culture related information, behavior driving information, value orientation information, audio to text information and corresponding safety problem into the expert module to obtain the possible safety culture problem of the dimension;

[0034] Step 5, check the safety culture problem, determine which identity should be responsible for checking the description of different low-score items and low-score item safety problems;

[0035] Step 6, fly inspection confirms the safety culture problem, the low-score item description and low-score item safety problem output in step 4 are pushed to the participants and promoters in the form of web page form, the participants in step 5 are pushed, the promoters are all pushed, and the time from receiving the form to opening the form and recording the video is recorded;

[0036] Step 7, propose safety problem improvement measures, input the low-score item description, low-score item problem and the recognition score of each level to the description and problem to obtain the low-score item improvement measures and the expected effect description of implementation measures;

[0037] Step 8, check the safety culture improvement measures, determine which identity should be responsible for checking the description of different low-score items and low-score item safety problems;

[0038] Step 9, confirm the feasibility of safety culture improvement measures, push the safety culture improvement measures and the expected effect of implementation measures obtained in step 7 to the promoters in the project and the high ranking personnel obtained in step 8, and to the safety culture experts outside the project; The push contains the low-score item description, low-score item safety problem, safety culture improvement measures and the expected effect of implementation measures;

[0039] Step 10, analyze the advantages and disadvantages of the measures to generate a conclusion report, by inputting the safety culture improvement measures, the expected effect of implementation measures and the reply obtained in step 9, an evaluation report and use instruction of the improvement measures are generated for the project manager from the aspects of cost, efficiency and feasibility.

[0040] Specifically, in the present application, the safety culture improvement measure generation method and system involve 3 types of objects, which are safety culture construction promoters, safety culture construction participants and external safety culture experts. The safety culture construction promoter refers to the personnel in charge of safety culture construction in the enterprise project, generally the safety personnel led by the enterprise safety director. The safety culture construction participant refers to the personnel at all levels and in all professions who bear responsibility for safety in the enterprise project. Under the enterprise-wide safety production responsibility system, it generally refers to all employees of the project, which can be divided into two categories: fixed employees and employees with strong mobility. The latter generally refers to workers engaged in specific scene operations. The external safety culture expert refers to professional personnel who provide evaluation and improvement services for enterprise safety culture. Specifically as shown in Figure 1 , the data generated at each link is stored in the cloud server. Except for the safety culture expert module, the large language model used in the module subunit is provided by the large language model module. In order to intuitively display the operation process, the underlying storage and large language model module are not shown in the figure. The process includes 10 steps, and the work content of each step is as follows:

[0041] 1. Identify the safety culture evaluation report: The safety culture evaluation report generally issued by professional external consulting team, format respectively PPT / Word / PDF. In this step, the safety culture construction promoter (hereinafter referred to as the promoter) accesses the information interaction module such as WeChat public number (Dingding, Feishu and other commonly used instant messaging products of other companies Dialog robots can be used), fills in the personal identity information, project name and other basic information according to the pop-up window, and uploads the evaluation report of the external consulting team. The information interaction module transmits the report to the safety culture identification module, which first extracts the text data in the report, then identifies the text data related to safety culture, and highlights it in the original report. Through the information interaction module, it is sent to the promoter again for confirmation. The promoter can send other text information related to the safety culture evaluation report separately, so as to reduce the identification error of the identification module. The identification of the safety culture evaluation report will arrange the scores of each item and dimension of safety culture (see Figure 7 ) and store it. After the uploading of the safety culture evaluation report is completed, this module will push a web form in the information interaction module, requiring the promoter to upload the project organizational structure, safety system files, safety training files (excluding examination questions), safety measures files and all safety logs in the past half year and other project safety materials into the system. All uploaded files are stored in the cloud server, and the file storage name is recorded according to the steps, and the storage structure is as Figure 2 .

[0042] 2. Collate safety data: In this step, the system will classify and analyze the data stored in the previous step in the project insight module, excluding the safety culture assessment report. Since the project data may overlap, i.e., safety training documents may contain some safety system document data, the project insight module algorithm needs to be invoked to classify all text materials, output and store four types of information: safety culture related information, action driving information, value orientation information, and safety problem information, as shown in Figure 3 . Safety culture related information refers to information related to the safety culture items in Figure 7 , behavior driving information refers to encouraging or limiting behavior information, value orientation information refers to encouraging or criticizing non-safety culture related value orientation information, and safety problem information refers to information containing safety problems, such as safety hazards.

[0043] 3. Flyover survey possible causes: In this step, the system first sends a web form to the promoter, which is forwarded to all safety culture construction participants (hereinafter referred to as participants) through IM. The form contains the department name information entered in step 1, and all participants are required to select their own department as the basis for subsequent flyover survey sampling. The information exchange module will sample the low-score safety culture items and safety problem materials to participants of various identities in the form of a web form, and record the time from receiving the form to opening the form and recording the video. After opening the form, automatically turn on the camera and microphone of the device, and if video or audio data cannot be obtained, automatically close the form. The flyover personnel are required to stop the work they are doing and focus on answering the questions after opening the form. The questions include why the safety culture dimension has low items and why there are safety problems. The flyover personnel are required to answer in voice, and the personnel can see the results of voice input, and if they are not satisfied, they can edit it in text. Store data as Figure 4 .

[0044] 4. Analyze project safety culture problems: In this step, the system will invoke the safety culture expert module. First, the safety culture assessment report and safety culture supplement information content in step 1 are disassembled according to the dimension (see Figure 7 ), and the low-score item description is obtained. Then, the low-score item description, corresponding safety culture assessment item (see Table 1), safety culture related information, behavior driving information, value orientation information in step 2, audio to text information in step 3, and corresponding safety problem are input into the expert module to obtain the possible safety culture problems of this dimension. Store data as Figure 5 .

[0045] Table 1 Safety Culture Assessment Items

[0046]

[0047]

[0048]

[0049] 5、Review safety culture issues: In this step, the system determines which identities should be assigned to review the low-scoring item situation descriptions and low-scoring item safety issues through the review module. The system compares the descriptions and issues output in step 4 with the audio-to-text information obtained in step 3, sorts them according to content relevance, and obtains the content relevance sorting of non-safety personnel participants. The sorting list is stored.

[0050] 6、Fly inspection confirms safety culture issues: In this step, the system pushes the low-scoring item situation descriptions and low-scoring item safety issues output in step 4 to the participants and promoters in the form of a web form through the information interaction module. The top-ranked participants in step 5 are given priority, and all promoters are given priority. The time from receiving the form to opening the form and recording the video is recorded. After the form is opened, the camera and microphone of the device are automatically turned on. If video or audio data cannot be obtained, the form is automatically closed. The person being inspected is required to stop the work being done and focus on answering the questions after opening the form. The questions include a low-scoring item situation description and a low-scoring item safety issue. The person being inspected is required to select the score. The data is stored as follows: Figure 6 .

[0051] 7、Propose safety issue improvement measures: In this step, the system inputs the low-scoring item situation descriptions, low-scoring item issues, and the recognition scores of each level for the descriptions and issues into the safety culture expert module to obtain the expected effect description of the low-scoring item improvement measures and implementation measures.

[0052] 8、Review safety culture improvement measures: In this step, the system determines which identities should be assigned to review the low-scoring item improvement measures through the review module. First, all promoters are the reviewers of the improvement measures. Then, the system compares the audio-to-text information obtained in step 3 with the improvement measures obtained in step 7, sorts them according to content relevance and the scores in step 6, and obtains the content relevance sorting of non-safety personnel participants. The sorting list is stored.

[0053] 9、Confirm the feasibility of safety culture improvement measures: In this step, the system pushes the safety culture improvement measures obtained in step 7 and the expected effect of the implementation measures to the promoters within the project and the high-ranking personnel obtained in step 8, as well as to the safety culture experts outside the project through the information interaction module. The push includes low-scoring item situation descriptions, low-scoring item safety issues, safety culture improvement measures, and the expected effect of the implementation measures. The pushed person is required to read it within 1 week, reply in writing, and fill out the form.

[0054] 10. Analyzing the pros and cons of the measures to generate a conclusion report: In this step, the system generates an evaluation report and usage instructions for the project manager to improve the measures by inputting the safety culture improvement measures, the expected effects of implementing the measures, and the feedback obtained in step 9 from multiple perspectives such as cost, efficiency, and feasibility.

[0055] According to the safety culture improvement measure generation method described in Tables 1-7, the second aspect of the present application also provides a safety culture improvement measure generation system, which includes an information interaction module, a safety culture identification unit, a project insight module, a safety culture expert module, a checking module, and a large language model module.

[0056] Information interaction module:

[0057] The information interaction module refers to the module through which various users interact with the system through a mobile terminal instant messaging system, including a video interaction unit, an audio interaction unit, a text interaction unit, a file interaction unit, and an audio-to-text unit. The video interaction unit includes a camera that can be called to open in a networked environment, and the video stream can be directly transmitted to the cloud server. The audio interaction unit includes a microphone that can be called to open in a networked environment, and the audio stream can be directly transmitted to the cloud server. The text interaction unit refers to the conventional interaction method of inputting in Chinese characters or virtual keyboard pinyin, and the mobile terminal can send text online. The picture interaction unit refers to the interaction method of the mobile terminal that can upload and send PDF, Word, PPT, and other formats of documents and Png, Jpeg, JPG, and other formats of pictures. The above-mentioned units are present in conventional smart phones, and enterprise employees can realize the functions of the above-mentioned four units through mobile terminal instant messaging software (such as WeChat). When conducting a flyover inspection, the user clicks on the web link in the instant messaging software to automatically open the camera and microphone, and the video stream and audio stream are stored in the cloud. The audio-to-text unit refers to a unit that converts the user's flyover inspection recorded voice into text, including four steps of noise reduction processing, feature extraction, speech recognition, and speech checking, converts it into text and stores it in json format, and the storage content also includes filling in the user's identity information.

[0058] Due to the complexity of the construction site, there may be noise, so it is necessary to first perform noise reduction on the audio to eliminate noise interference in the audio signal, to improve the quality and intelligibility of the audio. First, a noise reduction algorithm based on frequency domain filtering is needed: this algorithm converts the audio signal to the frequency domain, then filters out the noise frequency components through a filter, and finally converts the signal back to the time domain. Common frequency domain filters include band-stop filters, band-pass filters, etc. This algorithm is effective for Solid The noise effect of fixed frequency is better (complying with the noise of fixed equipment in some construction sites), but it is less effective for nonlinear and time-varying noise. The basic principle of this method can be described by the following symbols and formulas:

[0059] Description:

[0060] Suppose the input audio signal is audiomsg=x(n), its Discrete Fourier Transform (DFT) is X(k), and the corresponding noise signal is n(n), its DFT is N(k). The observed noise-polluted audio signal can be expressed as:

[0061] y(n)=x(n)+n(n)

[0062] The DFT of y(n) is Y(k)=X(k)+N(k). Since the power spectrum of the noise is low, the noise can be removed by using a low-pass filter in the frequency domain. Let the frequency response of the filter be H(k), then the output of the filter is:

[0063] Z(k)=H(k)Y(k)

[0064] The IDFT of Z(k) gives the noise-reduced audio signal z(n), i.e.

[0065]

[0066] where N is the number of sampling points of the audio signal. The final noise-reduced audio signal z(n) can be used for subsequent audio processing.

[0067] Secondly, a time-domain filter-based noise reduction algorithm is needed: this algorithm directly filters the audio signal in the time domain, and removes noise by designing different filters. Common time-domain filtering algorithms include median filtering, Kalman filtering, and wavelet noise reduction. This algorithm can effectively remove nonlinear and time-varying noise (consistent with some intermittent noise in construction sites), but it is less effective for fixed-frequency noise.

[0068] After completing the noise reduction, in order to reduce the computational power pressure, the present application selects the Mel Frequency Cepstral Coefficients (MFCCs) or Mel Filter Bank Energy (FBank) method to complete feature extraction. First, the audio signal is divided into short time frames (frame size 30 milliseconds, step 15 milliseconds), then the Fast Fourier Transform (a method for efficiently implementing the Discrete Fourier Transform DFT and its inverse transform) is performed on each frame, and then the power spectrum of the FFT is applied to a set of Mel filters, and the log of the filter bank energy is taken. The MFCCs method also needs to convert the log energy through Discrete Cosine Transform (DCT) to obtain a set of coefficients.

[0069] The conversion from audio to text is called automatic speech recognition (ASR). The present application selects a deep learning-based speech recognition algorithm: deep learning models including deep neural networks (DNNs), recurrent neural networks (RNNs), convolutional neural networks (CNNs), time recurrent neural networks (LSTMs), gated recurrent unit neural networks (GRUs), Transformer models, BERT models, etc. to model the relationship between audio signals and text, thereby realizing speech recognition.

[0070] In the last step, the unit will call the large language model of the large language model module, pass the recognition result into the large language model, and prompt the word as follows: "The input is a paragraph of speech recognition. Confirm whether there is a recognition error in this paragraph. Score from 1 to 5. 1 is completely correct, 5 is completely wrong. The input greater than 3 will be automatically corrected. The number of input and output results is required to be the same. The input less than 4 and the corrected input will be output in json format." Store the json format file, name the time identity item / safety culture question / safety question / measures, for example, 2013-12-05 technician safety communication.json.

[0071] Safety culture recognition unit:

[0072] The safety culture recognition module refers to the module used in the present application to analyze the safety evaluation report uploaded by the pusher, including a Word document recognition unit, a PPT document recognition unit, a PDF document recognition unit, and a detection unit.

[0073] (1) Word document recognition unit: This unit is used to identify the user uploaded Word format (.docx) safety culture evaluation report, output the highlighted safety culture related content docx format document and the xlsx format document including the safety culture score of each item, and store the highlighted text paragraph in json format in the cloud server. The first step is to open the evaluation report word based on the python-docx library, traverse all the paragraphs of the report, and store each paragraph as a json format document based on the json library, with the naming rule of document name + p1, for example, the first paragraph of the safety culture evaluation document is stored as "safety culture evaluation_p1.json". The second step is to call the expert large language model unit in the safety culture expert module, and input each json file obtained in the first step into the large language model respectively, with the prompt as follows: "Evaluate the correlation coefficient between the input text and safety culture, score from 1-5, 5 for very relevant; Evaluate the correlation coefficient between the input text and safety culture sub-dimension (Appendix 1), score from 1-5, 5 for very relevant; Output the highest value of the two evaluations". The corresponding json document with an output result greater than 2 is stored in a separate folder, named as the file name + safety culture + paragraph number, for example, "W safety culture evaluation_safety culture p1.json", W means the input source is word, and the documents starting with W in the following text all have this meaning. The third step is to copy the user uploaded word document, and the corresponding text of the json document filtered in the previous step is highlighted in yellow in the copied document, which is implemented by using the python-docx library Document, qn, parse_xml, RGBColor, etc. Finally, it is stored as "W safety culture evaluation S1.docx", W means the input source is word, and S1 means the first step. The fourth step is to identify the excel chart embedded in the word document based on base64, zipfile, os, io, etc. By opening the word file as a zip file, then retrieving and extracting the OLE object to get the excel chart, the file name is "report score 1.xlsx", and the naming of multiple charts is similar. Then read the excel content through the openpyxl library and the BytesIO class of the io library, and store the data as a json format, then input it into the expert large language model unit in the safety culture expert module to determine the safety culture dimension and item (Appendix 1) score, with the prompt as follows: "The input content is a json file content that may contain safety culture dimension and item scores, based on the input content, output the safety culture dimension and item scores, and judge whether all dimension scores have data." The output score is stored in the excel file, and the file name is "W safety culture score.xlsx".Fifth, send W safety culture evaluation S1.docx and W safety culture score.xlsx to the user through the information interaction module, and ask the user to confirm whether the report highlights and excel document content are comprehensive. If the user returns the document again, a new folder named "user has confirmed safety evaluation report" is created in the server, and the last version of the returned word and excel format documents with the same name is stored.

[0074] (2) PPT document recognition unit: this unit is used to identify the safety culture evaluation report in PPT format (.pptx) uploaded by the user. The first step is to apply the python-pptx library to traverse all the text box contents and chart contents in the safety culture evaluation report in.pptx format, store the text box contents as a word document (.docx) format, and store the chart contents as an excel. The second step is to input the word document into the Word document recognition unit and the fifth step of the Word document recognition unit, store the files "P safety culture evaluation S1.docx", "P safety culture evaluation_safety culture p1.json", "P safety culture score.xlsx", and P indicates that the input source is PPT. Similarly, the information interaction module needs to be called, the user needs to be confirmed, and the last version of the returned word and excel format documents with the same name needs to be stored.

[0075] (3) PDF document recognition unit: this unit is used to identify the safety culture evaluation report in PDF format (.pdf) uploaded by the user. Use the pdfplumber and python-docx libraries to cut the text in the pdf file according to the paragraph (\n symbol), input each paragraph of text into the large language model module, and prompt the word prompt "The input text is extracted from the pdf, there may be space and carriage return errors, confirm whether there are extra spaces in each paragraph of text, and whether the segmentation between paragraphs is correct, only output the original text after adjusting the space and carriage return, output according to the paragraph", output the text to the word file. Then call the Word document recognition unit to identify the safety culture related text, and finally store the "F safety culture evaluation S1.docx" document, F indicates that the input source is PDF. Similarly, the information interaction module needs to be called, the user needs to be confirmed, and the last version of the returned word format document with the same name needs to be stored.

[0076] (4) Detection unit: This unit is used to determine whether the user can proceed to the next step. When the user has uploaded the safety culture evaluation report, they can input "enter the next step" through the information interaction module. At this time, this unit is opened. First, it determines whether the "W safety culture evaluation S1.docx", "W safety culture evaluation_ safety culture p1.json", "W safety culture score.xlsx", "P safety culture evaluation S1.docx", "P safety culture evaluation_ safety culture p1.json", "P safety culture score.xlsx", and "F safety culture evaluation S1.docx" documents exist. If the W / P / F three safety culture evaluation S1.docx does not exist or the W / P two safety culture score.xlsx does not exist, the user is prompted to upload the safety culture evaluation report again. If the files all exist, the unit detects whether the scores of each item in the W / P two safety culture score.xlsx exist. If the score of a certain dimension does not exist or the score of a certain dimension in the W / P two documents is different, the user is prompted to confirm the score of the certain dimension and upload the final version. The complete safety culture score.xlsx is obtained finally. The scores of all items are sorted in size order, and the items with scores less than 4 or scores in the last 20% are stored separately as "low score safety culture items.xlsx". Finally, the user is required to fill in the names of all departments in the project in the webpage and upload the project safety culture supplementary information and other safety system files, safety training files, safety measures files, and all safety logs in the last half year. The formats include Word, PPT, and PDF. In this unit, only the text of all uploaded data is extracted and stored in a json file with the same file name. The folder name is safety culture supplementary information and project safety data.

[0077] Project insight module:

[0078] The project insight module refers to the classification and analysis of user-uploaded data in the present application, including a safety culture analysis unit, a system analysis unit, a measure analysis unit, and a safety problem analysis unit.

[0079] (1) Safety culture analysis unit: This unit is used to filter and classify the safety culture-related data obtained in step 1. The safety culture json file obtained in step 1 is transmitted to the expert large language model unit in the safety culture expert module through the json library. The prompt word is: "classify the input text content in the safety culture item and score it from 1 to 5. 1 means completely unable to be classified in this concept category, and 5 means completely able to be classified in this category. Output the category name with a score of 3 or above." The output result is stored as a new json file with the file name increased by "_classification", such as "P safety culture evaluation_ safety culture p1_ classification.json".

[0080] (2) System analysis unit: This unit is used to extract the value judgment mode encouraged by the organization from the non-safety culture data obtained in step 1, that is, the data related to value orientation. The non-safety culture json file obtained in step 1 is transmitted into the large language model module in the large language model through the json library one by one, and the value content in table 2 (each item corresponds to a json file) is transmitted at the same time, and the prompt word prompt is: "Judge whether the input text will make employees have the tendency of value content X item, score 1-5, 1 for no tendency at all, 5 for complete tendency, output the complete content of X item tendency and the content related to this item in the input text in json format if the score is 3 or above." The output result is stored as a new json file, the file name is added with "_value orientation" after the input file name, for example "P safety training_value orientation.json", that is, the value orientation information stored in step 2.

[0081] Table 2: List of job value orientation

[0082]

[0083]

[0084] (3) Measure analysis unit: This unit is used to extract the behavior encouraged by the organization from the non-safety culture data obtained in step 1, that is, the data related to behavior drive. The non-safety culture json file obtained in step 1 is first transmitted into the large language model module in the large language model through the json library one by one, and the prompt word prompt is: "Judge the behavior of employees contained in the input text, output the behavior name in json format, and press enter after each name." The output result is stored as a new json file, the file name is added with "_behavior" after the input file name, for example "P safety training_behavior.json". Then the non-safety culture json file obtained in step 1 and the behavior json file just obtained are transmitted into the large language model module in the large language model, and the prompt word prompt is: "Judge whether the behavior in the input text is encouraged, score 1-5, 1 for completely prohibited, 5 for completely encouraged. Output the behavior with a score of 3 or above in json format, and the output content includes the behavior name, the behavior encouraged score and the judgment basis, press enter after the judgment basis output." The output result is stored as a new json file, the file name is added with "_behavior drive" after the input file name, for example "P safety training_behavior drive.json", that is, the behavior drive information stored in step 2.

[0085] (4) Safety problem analysis unit: This unit is used to refine specific safety problems from the non-safety culture related data obtained in step 1. The non-safety culture related json file obtained in step 1 is transmitted into the large language model module in the large language model library one by one, and the prompt word is: "first judge whether the input information is safety training materials, if it is, end, if not, judge what construction safety problems exist in the input information, output the number of safety problems identified in json format first, then output the judgment result, the result includes the safety problems identified and the judgment basis text, after outputting the judgment basis, enter." The output result is stored as a new json file, the file name is added "_ safety problem" after the input file name, for example "P safety log_ safety problem.json", which is the behavior driving information stored in step 2.

[0086] Safety culture expert module:

[0087] The safety culture expert module refers to the module used in the present application to analyze safety culture problems, propose intervention measures and analyze the pros and cons of the measures. It includes safety culture problem analysis unit, safety culture improvement measure generation unit, safety culture improvement measure analysis unit and expert large language model unit.

[0088] (1) Safety culture problem analysis unit: this unit is used to analyze the safety problems existing in the project from the data obtained in steps 2 and 3. First, analyze the relevance of safety problems and safety problem possible cause materials and safety culture items obtained in steps 2 and 3, pass the data into the expert large language model, the prompt is "analyze the relevance of the input content and each safety culture item, and score 1-5, 1 for irrelevant, 5 for relevant, output safety culture items with a score of 3 or above and the input content to json", store the json name as safety problem 1_safety culture item 1.json, for example "high fall hidden danger not timely handling_safety communication.json". Then take the intersection of safety culture low item materials and the safety problem and safety culture item relevance classification just obtained, and the value orientation file and behavior driving file together, pass into the expert large language model, for example, if the safety communication item score is low, then "P safety culture assessment_safety culture p1_classification.json" (contains safety communication) and "high fall hidden danger not timely handling_safety communication.json" and all value orientation files and behavior driving files are passed into the expert large language model. The prompt is "pass in XXX safety culture item related materials, the order is safety culture low item corresponding description, low item related safety problem, organizational value orientation and organizational behavior driving material, analyze the possible safety culture problem of the item comprehensively, and score 1-5 according to the logical reasonableness, output the safety culture problem with a score of 3 or above, that is, the basis and reasoning process of the problem, and output to the json file." The output is stored in json format, named XXX item safety culture problem.json.

[0089] (2) Safety culture improvement measure generation unit: this unit is used to propose safety culture improvement measures from the safety culture problem feedback obtained in step 6. First, filter the authenticity and reasonableness scores of each safety culture problem, only pass the single problem, problem corresponding description, scorer identity and all behavior driving data obtained in step 2 into the expert large language model module with a score of 3 or above. The prompt is "combine organizational behavior driving data and XXX identity to generate safety culture improvement measures to solve safety culture problems and describe the effect of the solved safety culture corresponding item." The output is stored in json format, named XXX item safety culture improvement measure 1.json. The number represents that a safety culture item may have multiple safety culture problems.

[0090] (3) Safety culture improvement measure analysis unit: This unit is used to generate an evaluation report and usage instructions for project managers from the safety culture improvement measures obtained in step 7, the expected effects of the implementation measures, and the feedback obtained in step 9, from the perspectives of cost, efficiency, feasibility, and other aspects. Enter the safety culture improvement measures, the feedback on the measures, and the value-oriented data obtained in step 2 into the expert large language model, with the prompt "Combine the feedback of experts and the value-oriented data of the organization, analyze the safety culture improvement measures and the effects of the safety culture improvement measures from the perspectives of cost, efficiency, feasibility, and impact on the existing values of the organization." The results are stored as a word (.docx) document with the file name "XXX Safety Culture Improvement Measures and Analysis.docx".

[0091] (4) Expert large language model unit: refers to a professional large language model unit for the construction industry safety culture obtained by fine-tuning a pre-trained language model (such as BERT, GPT, ChatGLM, etc.). Here, fine-tuning techniques include full model fine-tuning, partial fine-tuning, adaptive fine-tuning, feature extraction, multi-task fine-tuning, domain-specific fine-tuning, and other fine-tuning methods. For ChatGLM models, weight fine-tuning can be performed using Freeze, LoRA, and P-tuning methods. In addition, we also need to consider how to obtain a large number of training samples. On the one hand, existing materials can be formatted, and on the other hand, human questions can be collected and corresponding answers can be given.

[0092] Verification module:

[0093] The verification module refers to the module used to verify the safety culture problems and safety culture improvement measures generated by the safety culture expert module in the present invention. It includes two units: safety culture problem verification and safety culture improvement measure verification.

[0094] (1) Safety culture problem verification unit: refers to the unit used to verify the safety culture problems generated by the safety culture expert module. First, call the expert large language model of the safety culture expert module, compare the flyover inspection answers obtained in step 3 and the safety culture problem related materials output in step 4 for similarity, with the prompt "Based on safety culture knowledge, compare the two contents, do they describe the same safety culture problem? 1-5 score, 1 for completely different, 5 for completely the same, output the identity of the flyover inspection answer object with a score of 4 or above." The results are directly used as the identity of the information sending object of the information interaction module, the scores of each non-safety job department are calculated and sorted, and the sorted list is stored.

[0095] (2) Safety culture improvement measure checking unit: refers to a unit for checking the safety culture measures generated by the safety culture expert module. First, the expert large language model of the safety culture expert module is called, and the safety culture improvement measures output in step 7 are compared with the flight inspection answers obtained in step 3 for relevance. The prompt is "Based on safety culture knowledge, compare the two pieces of content. Does the safety culture improvement measure solve the safety culture problem mentioned in the other piece of text? 1-5 points, 1 point for completely irrelevant, 5 points for completely relevant. Output flight inspection answers with scores of 2 points or more. The identity of the object." The result and the rationality and authenticity scores obtained in step 6 are weighted and averaged to calculate the scores of each non-safety department and sort them, and the sorted list is stored.

[0096] Large language model module:

[0097] A large language model refers to a deep learning model trained using a large amount of text data, which can generate natural language text or understand the meaning of language text. Previous language models often focus on a certain type of natural language task, such as text classification, translation, and question answering. However, researchers have found that increasing the size of the model can exhibit stronger natural language processing capabilities (such as learning through context). GPT (Generative Pre-Trained Transformer) is a typical LLM that exhibits strong emergent capabilities. ChatGPT is based on LLM and changes to a dialogue form, using an LLM to complete multiple tasks by giving prompt instructions during the interaction. In this module, the large language model can return text output results according to the given text input instructions. Currently, there are many open-source LLMs, such as T0, ChatGLM, and Alpaca.

[0098] A language model can be treated as a black box that accepts token strings as input (where tokens can be Chinese characters or English words, etc.) and outputs a probability indicating the probability that the token string is a normal human sentence (or fragment). Mathematically, it is formalized as follows:

[0099] Given a token sequence (u1, u2,..., un), the language model outputs a probability p(u1, u2,..., un) representing the probability that these tokens form a sentence (or fragment) in order. The following formula expresses the above language model by expanding the probability into a conditional probability form:

[0100] p(u1, u2,..., un) = p(u1) ∏ p(ui | u1, u2,..., ui-1)

[0101] The language model described above can complete the task of text generation: given the first several generated words, calculate the next word that can maximize the sequence probability, and output this word as the prediction result; then the model will add the predicted word to the given sequence, and repeat the above process to continue predicting the next one until the predicted next word is the end symbol or the required length is reached.

[0102] The safety culture is a combination of safety beliefs and values built on people, things and objects in a project, and reflects the views and habits of all people in the collective on life and safety. The promotion of safety culture needs to mobilize as many people at all levels of the project as possible to participate, and to jointly find and improve the weak links in the cultural level. On the one hand, the present application can help the project to better base on the project organization structure when diagnosing problems and designing improvement measures, and let more comprehensive personnel at all levels actively participate in the improvement work. On the other hand, it can automatically analyze safety culture related problems based on internal and external personnel and existing written materials of the project, and generate high-quality improvement measures. The generated improvement measures will be checked by the present application, and evaluation reports and usage instructions of the improvement measures will be generated for project managers from the aspects of cost, efficiency, feasibility and the like, to help the improvement measures to better land. The present application activates the safety culture improvement initiative participation consciousness of the project personnel from bottom to top, provides effective improvement measures for the project from top to bottom, and activates the initiative safety culture improvement consciousness of the key node personnel. The improvement measure design method based on the large language model can make full use of a large amount of data for learning, and generate more scientific and efficient safety culture improvement scheme, avoiding the limitations of relying on subjective experience of experts. And the system can customize safety culture improvement scheme for users according to the characteristics and needs of specific projects, and improve the pertinence and effectiveness of actual implementation. Compared with the traditional method, the present application has higher efficiency, scientificity and comprehensiveness, and can better meet the needs of safety culture improvement in the construction industry, and improve the project safety management level.

[0103] Although the present application has been illustrated and described with multiple embodiments, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Numerous modifications, changes and substitutions can occur to those skilled in the art without departing from the spirit and scope of the present application. It should be understood that various alternatives to the embodiments of the present application described herein can be employed in practicing the present application. The appended claims are intended to cover all such alternatives as would be included within the scope of the present application.

Claims

1. A safety culture improvement measure generation method characterized by comprising: Comprise: Step 1, identify the safety culture evaluation report, extract the text data in the report, then identify the text data related to safety culture, and organize the identification results to form the scores of each item and dimension of safety culture and store, the storage structure includes the identity information of the person who promotes, the safety culture evaluation report, the safety culture supplementary information, the organizational structure and the project safety data; Step 2, sort out the safety data, classify and analyze the data stored in the previous step except the safety culture evaluation report, output and store safety culture related information, action driven information, value oriented information and safety problem information; Step 3, fly inspection of possible reasons, sample sending of low score safety culture items and safety problem materials to participants of various identities in the form of web forms, the safety problems include why the dimension of safety culture has low score items and why there are safety problems; Step 4, analyze the safety culture problems of the project, disassemble the safety culture evaluation report and safety culture supplementary information content according to the dimension, get the description of each low score item, then input the description of low score item, corresponding safety culture evaluation item, safety culture related information, behavior driven information, value oriented information, audio to text information and corresponding safety problem into safety culture expert module, get the possible safety culture problems of this dimension; Step 5, check the safety culture problems, determine which identity people should be responsible for checking different low score item descriptions and low score item safety problems; Call the expert large language model of the safety culture expert module, compare the fly inspection answers and safety culture problem related materials for similarity, take the result as the object identity of the information sent by the information interaction module, calculate the scores of each non-safety department and sort, store the sorting list; Step 6, fly inspection of safety culture problems, push the low score item description and low score item safety problem output in step 4 to the participants and promoters in the form of web forms, the participants with high ranking in step 5 are pushed, the promoters are all pushed, and the time from receiving the form to opening the form is recorded; Step 7, propose safety culture improvement measures, input the low score item description, low score item problem and the recognition score of each level to the description and problem into the safety culture expert module, get the low score item improvement measures and the expected effect description of implementation measures; Call the expert large language model of the safety culture expert module, compare the fly inspection answers and safety culture improvement measures for relevance, take the result as the object identity of the information sent by the information interaction module, calculate the scores of each non-safety department and sort, store the sorting list; Step 8, check the safety culture improvement measures, determine which identity people should be responsible for checking different low score item improvement measures; Step 9, confirm the feasibility of safety culture improvement measures, push the safety culture improvement measures and the expected effect of implementation measures obtained in step 7 to the promoters in the project and the high ranking personnel obtained in step 8, and to the safety culture experts outside the project; The low score item description, low score item safety problem, safety culture improvement measure and the expected effect of implementation measure are included in the push. Step 10, analyze the pros and cons of the measures to generate a conclusion report, input the safety culture improvement measures, expected effect of the implementation measures and the reply opinions obtained in step 9 into the safety culture expert module, so that it generates an evaluation report and usage instructions for the project manager from the perspectives of cost, efficiency and feasibility.

2. A safety culture improvement measure generation system that operates a safety culture improvement measure generation method according to claim 1, characterized by Comprise: Information interaction module, safety culture identification module, project insight module, safety culture expert module, checking module, large language model module; The information interaction module is used for interaction through video, audio, text and file respectively, and converts the voice recorded by the user into text; The safety culture identification module is used for analyzing the safety evaluation report uploaded by the safety culture construction promoter; The project insight module is used for classifying and analyzing user-uploaded materials, including safety culture analysis, system analysis, measure analysis and safety problem analysis; The safety culture expert module is used for analyzing safety culture problems, proposing intervention measures and analyzing the pros and cons of the measures; The checking module is used for checking the safety culture problems and safety culture improvement measures generated by the safety culture expert module; The large language model module is a deep learning model trained using a large amount of text data to generate natural language text or understand the meaning of language text.

3. The safety culture improvement measure generation system according to claim 2, characterized by, The voice recorded by the user is converted into text, which specifically includes: S1, frequency domain filtering denoising step, convert the audio signal to frequency domain, filter through low-pass filter, and convert the audio signal back to time domain to obtain the first denoising result; S2, time domain filtering denoising step, time domain filtering the first denoising result through a preset filter set to obtain a second denoising result, the preset filter set includes one or a combination of several of median filtering, Kalman filtering and wavelet denoising; S3, feature extraction of the second denoising result by Mel frequency cepstral coefficient (MFCC) or Mel filter bank energy (FBank) method; S4, speech recognition of the feature extraction result by using deep learning speech recognition algorithm; S5, calling the large language model of the large language model module, and inputting the recognition result into the large language model.

4. The safety culture improvement measure generation system according to claim 2, characterized by, The safety culture identification module comprises: Word document identification unit, for identifying the user-uploaded Word format (.docx) safety culture evaluation report, outputting the docx format document with highlighted safety culture related content and the xlsx format document including the scores of each item of safety culture, and storing the highlighted text paragraphs in json format in the cloud server, and not processing the pictures in the document; PPT document identification unit, for identifying the user-uploaded PPT format (.pptx) safety culture evaluation report, storing the text box content as a word document (.docx) format, storing the chart content as an excel table format, and processing it by using the word document identification unit; The PDF document recognition unit is used for recognizing a safety culture evaluation report in PDF format (.pdf) uploaded by a user, cutting the text in the report according to paragraphs, inputting each paragraph of text into a large language model module for confirmation, and outputting a result to a word file, wherein the word document recognition unit is used for processing; The detection unit is used for judging whether a user can enter the next step, and detecting the integrity of a file and content when the user uploads a safety culture evaluation report.

5. The safety culture improvement measure generation system of claim 2, wherein, The project insight module includes a safety culture analysis unit, a system analysis unit, a measure analysis unit, and a safety problem analysis unit; The safety culture analysis unit is used for screening safety culture related data obtained from the information interaction module, and classifying the data according to items by using an expert large language model unit in a safety culture expert module; The system analysis unit is used for extracting a value judgment mode encouraged by an organization, i.e., data related to value orientation, from non-safety culture data obtained from the information interaction module, and classifying the data according to items by using the expert large language model unit in the large language model module; The measure analysis unit is used for extracting a behavior encouraged by an organization, i.e., data related to behavior driving, from non-safety culture data obtained from the information interaction module, and classifying the data according to items by using the expert large language model unit in the large language model module; The safety problem analysis unit is used for extracting specific safety problems from non-safety culture related data obtained from the information interaction module, and classifying the data according to items by using the expert large language model unit in the large language model module.

6. The safety culture improvement measure generation system according to claim 2, wherein, The safety culture expert module includes; The safety culture problem analysis unit is used for analyzing interaction data to obtain the correlation between safety problems, possible causes of safety problems, and safety culture items; The safety problem improvement measure unit is used for providing safety culture improvement measures according to feedback on safety culture problems, including screening the authenticity and reasonableness scores of each safety culture problem, and only transmitting single problems, problem corresponding descriptions, scorer identities, and behavior driving data with scores of 3 or more into the expert large language model module; The safety culture improvement measure analysis unit is used for generating an evaluation report and a usage instruction for improvement measures for a project manager from multiple angles of cost, efficiency, and feasibility, based on safety culture improvement measures, expected effects of implementation measures, and reply opinions; The expert large language model unit is a building industry safety culture professional large language model unit obtained by training a pre-trained language model, including BERT, GPT, and ChatGLM, through Fine-tuning technology.

7. The safety culture improvement measure generation system of claim 2, wherein, The checking module includes: The safety culture problem checking unit is used for checking safety culture problems generated by the safety culture expert module, including calling the expert large language model of the safety culture expert module, comparing the similarity between the answers and safety culture problem related materials, taking the result as the object identity of the information sent by the information interaction module, calculating and sorting the scores of each non-safety department, and storing the sorted list. Safety culture improvement measure checking, used to check the safety culture measures generated by the safety culture expert module, including calling the expert large language model of the safety culture expert module, flying to check the answers and the correlation of the safety culture improvement measures, reasonably weighting the scores of the rationality and the authenticity, calculating the scores of each non-safety job department and sorting, and storing the sorting list.

8. The safety culture improvement measure generation system of claim 2, wherein, The large language model module is used for completing the task of text generation: Given a number of generated words in the preceding, the next word that can maximize the sequence probability is calculated, and the word is output as the prediction result; then the model adds the predicted word to the given sequence and repeats the above process to continue predicting the next one until the predicted next word is the end symbol or the required length is reached.

Images