Industrial safety health risk prediction method and equipment based on machine learning, and medium

Abstract

The invention relates to an industrial safety health risk prediction method and device based on machine learning, and a medium. The method comprises the following steps: obtaining industrial scene description, generating structured data through a large language model, inputting the structured data into a trained industrial safety health risk prediction model, and outputting a predicted potential risk level. Wherein the industrial safety health risk prediction model comprises a feature extraction module, a feature enhancement module and a feature fusion module, and the feature extraction module is used for extracting text features and attribute features; the feature enhancement module comprises an attribute feature enhancement network based on a Transform architecture, a text feature enhancement network based on a Bi-LSTM architecture and a cross attention mechanism, and an accident level embedded network, and is used for performing feature enhancement; and the feature fusion module is used for splicing the output of the feature enhancement module to obtain a comprehensive feature vector. Compared with the prior art, the method can further improve the prevention and coping capacity of safety production accidents.

Description

Technical Field

[0001] The present invention belongs to the field of industrial health and safety, and in particular, relates to an industrial safety and health risk prediction method, device and medium based on machine learning. Background Art

[0002] With the acceleration of industrialization and the popularization of modern manufacturing technology, how to effectively prevent and manage industrial accidents has become a focus of attention in the industrial community. At present, most commonly used industrial safety management technologies are based on the "after-the-fact analysis" model, relying on accident reports, empirical analysis and statistical methods. Although these methods have provided basic support for industrial safety management in the past few decades, they also have obvious limitations, especially when facing complex and changing industrial environments, it is difficult to play an effective preventive role. The following are the main problems in the development of existing technologies:

[0003] 1. Insufficient data utilization: In the process of modern industrial production, enterprises have accumulated a large amount of safety-related data, including equipment failure reports, worker health records, accident investigation reports, etc. However, these data are often only used as post-archive materials, and have not been deeply mined and analyzed, resulting in a large amount of valuable information being ignored. Existing analysis methods are mostly based on simple statistics, which cannot fully utilize the potential information in these data and make it difficult to fully control potential risks.

[0004] 2. Insufficient real-time monitoring and response capabilities: In the current industrial safety management system, although some companies have begun to introduce sensor networks and Internet of Things technologies for safety monitoring, the existing system still has obvious deficiencies in real-time performance and response speed. This lagging response mechanism is particularly insufficient in a rapidly changing industrial environment, especially when responding to sudden accidents, there is not enough time to take effective preventive measures.

[0005] 3. Low level of intelligence and poor industry adaptability: Although modern industry has adopted a large number of automated equipment and smart factory technologies, in the field of safety management, the application of intelligence is still at a low level and lacks the ability to learn and improve autonomously. In addition, traditional safety management methods are often general and lack customized solutions for industry characteristics. For example, the risks of high-altitude operations in the construction industry, the risks of hazardous substance leakage in the chemical industry, and the risks of mechanical injuries in the manufacturing industry, the risk factors and accident types of various industries are significantly different, but the existing technology is difficult to adjust according to the specific needs of different industries, resulting in limited effects in practical applications.

[0006] Therefore, it is necessary to design a new type of industrial safety and health risk prediction scheme to further improve the prevention and response capabilities of production safety accidents. Summary of the invention

[0007] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide an industrial safety and health risk prediction method, equipment and medium based on machine learning, make full use of industrial safety-related data, and provide reliable and timely early warning for potential risks.

[0008] The purpose of the present invention can be achieved by the following technical solutions:

[0009] The present invention provides an industrial safety and health risk prediction method based on machine learning, comprising the following steps:

[0010] Obtaining industrial scene descriptions, generating structured data through a large language model, inputting the structured data into a trained industrial safety and health risk prediction model, and outputting predicted potential risk levels, wherein the structured data includes accident descriptions, attribute extraction words, description extraction words, and accident levels;

[0011] Wherein, the industrial safety and health risk prediction model includes a feature extraction module, a feature enhancement module and a feature fusion module;

[0012] The feature extraction module is used to extract text features from the accident description and the description extraction words, and extract attribute features from the attribute extraction words;

[0013] The feature enhancement module includes a parallel attribute feature enhancement network, a text feature enhancement network and an accident level embedding network. The attribute feature enhancement network is built based on the Transformer architecture and is used to enhance the attribute features; the text feature enhancement network is built based on the Bi-LSTM architecture and the cross attention mechanism and is used to perform word-level and sentence-level joint enhancement on the text features; the accident level embedding network is used to generate an accident level embedding vector according to the accident level;

[0014] The feature fusion module is used to splice the outputs of each network in the feature enhancement module to obtain a comprehensive feature vector, and identify the potential risk level based on the comprehensive feature vector.

[0015] Furthermore, the specific process of generating structured data through a large language model is as follows:

[0016] The industrial scene description is embedded in the prompt word template, and the large language model extracts information from the industrial scene description based on the prompt word template;

[0017] The extracted information is converted into structured data in a list structure through the eval function.

[0018] Furthermore, the attribute feature enhancement network includes a splicing layer, a global average pooling layer, a Transformer encoder and an attention layer connected in sequence; the attribute feature includes multiple sub-attribute features, and after the multiple sub-attribute features are spliced ​​through the splicing layer, a compact feature representation is generated through the global average pooling layer and the Transformer encoder; the attention layer performs weighted calculation on the compact feature representation through an attention mechanism to generate enhanced attribute features.

[0019] Furthermore, a mean layer and a random dropout layer are set between the Transformer encoder and the attention layer.

[0020] Furthermore, the text feature enhancement network includes a cross attention layer, and a parallel word-level feature layer and a sentence-level feature layer;

[0021] The word-level feature layer and the sentence-level feature layer are both constructed based on the Bi-LSTM architecture. The word-level feature layer is used to extract word-level attention vectors from the text features, and the sentence-level feature layer is used to extract sentence-level attention vectors from the accident description; the cross-attention layer is used to perform weighted calculation on the word-level attention vectors and the sentence-level attention vectors through a cross-attention mechanism to generate enhanced text features.

[0022] Furthermore, the accident level embedding network includes a ReLU activation function layer and a random inactivation layer connected in sequence.

[0023] Furthermore, the industrial safety and health risk prediction model is trained by back propagation and the model parameters are updated using the Adam optimizer.

[0024] Furthermore, the industrial safety and health risk prediction model uses a cross entropy loss function during the training process.

[0025] The present invention also provides an electronic device, comprising a memory, a processor, and a program stored in the memory, wherein the processor implements the above method when executing the program.

[0026] The present invention also provides a computer-readable storage medium on which a computer program is stored, and the program implements the above method when executed by a processor.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. The present invention proposes an industrial safety and health risk prediction method based on machine learning. First, an industrial scene description is obtained, and structured data is generated through a large language model; then, the structured data is input into a trained industrial safety and health risk prediction model to output a predicted potential risk level; the industrial safety and health risk prediction model includes a feature extraction module, a feature enhancement module and a feature fusion module, wherein the feature extraction module is used to extract text features from accident descriptions and description extraction words, and to extract attribute features from attribute extraction words. The present invention combines multiple data sources such as accident descriptions, attribute extraction words, description extraction words and accident levels to achieve deep fusion of multimodal features, and more comprehensively capture feature information related to accident risks, thereby greatly improving the accuracy of prediction and adaptability to diversified industrial scenarios; the feature enhancement module includes a parallel attribute feature enhancement network, a text feature enhancement network and an accident level embedding network. The attribute feature enhancement network is built based on the Transformer architecture and is used to enhance attribute features. The Transformer architecture has efficient parallel computing capabilities, especially when processing future attributes and long sequence data, it can significantly improve computing speed and efficiency; the text feature enhancement network is built based on the Bi-LSTM architecture and the cross-attention mechanism, which is used to jointly enhance text features at the word level and sentence level. The Bi-LSTM architecture can capture long-range dependencies in the data and help the model better understand the complex temporal information in the accident description. The text features at the word level and sentence level can be aligned through the cross-attention mechanism to capture the correlation between the two and make full use of the input data; the accident level embedding network is used to generate an accident level embedding vector according to the accident level; the feature fusion module is used to splice the outputs of each network in the feature enhancement module to obtain a comprehensive feature vector, and identify the potential risk level based on the comprehensive feature vector; through the above method, we can make full use of industrial safety-related data, provide reliable and timely warnings for potential risks, and further improve the prevention and response capabilities of production safety accidents.

[0029] 2. The present invention embeds the acquired industrial scene description into a prompt word template. The large language model extracts information from the industrial scene description according to the prompt word template, and then converts the extracted information into structured data in a list structure through the eval function. Through the large language model technology, the present invention can make full use of the large amount of security-related raw data accumulated by the enterprise, deeply mine the potential information in these data, and achieve comprehensive control of potential risks.

[0030] 3. In the present invention, the attribute feature enhancement network includes a splicing layer, a global average pooling layer, a Transformer encoder, a mean layer, a random dropout layer and an attention layer connected in sequence. The attribute feature includes multiple sub-attribute features. After the multiple sub-attribute features are spliced ​​through the splicing layer, a compact feature representation is generated through the global average pooling layer and the Transformer encoder; the mean layer and the random dropout layer can prevent the network from overfitting and improve the generalization ability of the model; the attention layer performs weighted calculation on the compact feature representation processed by the mean layer and the random dropout layer through the attention mechanism to generate enhanced attribute features. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the structure of the industrial safety and health risk prediction model;

[0032] Figure 2 Schematic diagram of the structure of the attribute feature enhancement network;

[0033] Figure 3 Schematic diagram of the structure of the text feature enhancement network;

[0034] Figure 4 It is a schematic diagram of the structure of the accident level embedding network;

[0035] Figure 5 Schematic diagram of the training process of the industrial safety and health risk prediction model. DETAILED DESCRIPTION

[0036] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0037] Example:

[0038] This embodiment provides an industrial safety and health risk prediction method based on machine learning. The method runs in the GPU, integrates multimodal features such as accident descriptions and attribute extraction words extracted by a large language model, and uses a variety of deep learning architectures (including Bi-LSTM, Transformer and Attention mechanisms) to predict potential risk levels. It has efficient feature extraction capabilities and powerful classification performance, and can process large-scale industrial data in real time, thereby helping enterprises take effective preventive measures before accidents occur. The method mainly includes the following steps:

[0039] Step 1: Process historical industrial data through a large language model to generate a data set.

[0040] First, obtain historical industrial data in the form of industrial scene descriptions; then, embed the industrial scene descriptions into prompt word templates and send them to the large language model, which can extract information from the industrial scene descriptions based on the prompt word templates; finally, use the eval function to convert the extracted information into structured data in a list structure, which specifically includes accident descriptions (x_description), attribute extraction words (x_f_attribute), description extraction words (x_f_description), accident levels (x_accident_level), and potential risk level labels (y_potential_accident_level). Repeat the above steps to complete the data set construction, and then construct the training data set, validation data set, and test data set in a ratio of 6:2:2.

[0041] Through the large language model technology, we can make full use of the large amount of raw data related to security accumulated by the enterprise, deeply mine the potential information in these data, and achieve comprehensive control of potential risks. The following is an example of a prompt word template:

[0042] You are an assistant responsible for extracting entities from the text information I provide. The scenario is a description of the industrial safety risk field, including: country, local, industry, accident level, type, employee or third party, hazard level, and description. The requirements are as follows:

[0043] 1. I will provide a description, and you need to extract the following aspects:

[0044] a.country represents a word or phrase that can represent a country

[0045] b.local represents words or phrases that can express local

[0046] c.industry_sector represents words or phrases that can represent the industrial field, such as Mining, Metals

[0047] e. Genre represents words or phrases that can indicate the target gender, such as she, he, him, her, it, his, hers

[0048] f.employee_or_third_party represents words or phrases that can represent employee types, such as Employee, Third Party, Third Party (Remote)

[0049] g.critical_risk represents a word or short phrase that can represent any risk, such as Pressurized Systems, Fall prevention (same level), Confined space

[0050] 2. The output should be in json format, with the key being the corresponding noun, such as "country" or "local", and the value being a list containing all the extracted words or phrases

[0051] 3. Output sample:

[0052]

[0053] Step 2: Build an industrial safety and health risk prediction model and use the data set constructed in step 1 for model training.

[0054] like Figure 1 As shown in the figure, the industrial safety and health risk prediction model includes a feature extraction module, a feature enhancement module and a feature fusion module. Among them, the feature extraction module is used to extract text features from the accident description and description extraction words. This process uses sentence embedding technology to convert the text into a high-dimensional embedding vector to retain the semantic and contextual information in the text. The feature extraction module is also used to extract attribute features from attribute extraction words. The attribute features mainly include critical risks (Critical Risk), industry fields (Industry Sector), employee or third party identity (Employee or Third Party), category (Genre) and other information related to accident risks.

[0055] The feature enhancement module (FEM) is used to perform relevance feature extraction (Relevance Feature Extraction) and entity relationship extraction (Entity Relationship Extraction) on text features and attribute features, and output enhanced feature vectors. The feature enhancement module includes a parallel attribute feature enhancement network, a text feature enhancement network, and an accident level embedding network. The attribute feature enhancement network is built based on the Transformer architecture and is used to enhance attribute features. The text feature enhancement network is built based on the Bi-LSTM architecture and the cross-attention mechanism and is used to jointly enhance text features at the word level and sentence level. The accident level embedding network is used to generate an accident level embedding vector based on the accident level. The specific settings of each network in the feature enhancement module (FEM) are as follows:

[0056] like Figure 2As shown in the figure, the attribute feature enhancement network includes a concatenation layer, a global average pooling layer, a Transformer encoder, a mean layer, a random dropout layer, and an attention layer connected in sequence. The attribute feature includes multiple sub-attribute features, which are first concatenated through the concatenation layer and then generated into a compact feature representation through the global average pooling layer and the Transformer encoder. The Transformer structure has strong parallel processing capabilities, can capture complex feature interactions in the data, and efficiently process long sequence data. The compact feature representation then passes through the mean layer and the random dropout layer in sequence, which can prevent the network from overfitting and improve the generalization ability of the model. The attention layer performs weighted calculation on the compact feature representation processed by the mean layer and the random dropout layer through the attention mechanism to generate enhanced attribute features.

[0057] like Figure 3 As shown in the figure, the text feature enhancement network includes a word-level feature layer, a sentence-level feature layer, and a cross-attention layer, where both the word-level feature layer and the sentence-level feature layer are built based on the Bi-LSTM architecture. The word-level feature layer is used to extract word-level attention vectors from text features, the sentence-level feature layer is used to extract sentence-level attention vectors from accident descriptions, and the cross-attention layer is used to perform weighted calculations on word-level attention vectors and sentence-level attention vectors through a cross-attention mechanism to generate enhanced text features. The cross-attention mechanism can capture long-range dependencies in time series data, help the model better understand the complex time series information of accident descriptions, and extract global context information with higher relevance through weighted calculations.

[0058] like Figure 4 As shown in Figure 1, the accident level embedding network includes a ReLU activation function layer and a random dropout layer connected in sequence.

[0059] The output and input feature fusion modules of each network in the feature enhancement module (FEM) are spliced ​​to obtain a comprehensive feature vector. The spliced ​​comprehensive feature vector contains all the features of the input information. The potential risk level is obtained by identifying the comprehensive feature vector.

[0060] The training process of the industrial safety and health risk prediction model is as follows Figure 5 As shown in the figure, first, data is obtained from the training set and used as the input of the forward propagation for linear calculation and activation function processing. After all linear calculations are completed, the loss and back-propagation error are calculated, and the model parameters are updated through the Adam optimizer until the upper limit of the training times is reached.

[0061] In this example, the cross entropy loss function is used:

[0062]

[0063] in, andi are the predicted value and the true value respectively, N represents the number of samples (that is, the number of instances in the data set), and C represents the number of classes (that is, the number of categories in the classification task).

[0064] Step 3: Obtain real-time industrial scene descriptions, generate structured data through a large language model, input the structured data into a trained industrial safety and health risk prediction model, and output the predicted potential risk level. In a preferred embodiment, after the potential risk level is identified, the results can be input into a large language model (LLM) to generate specific prevention recommendations to provide targeted solutions.

[0065] Compared with the prior art, the above method has the following improvements and beneficial effects:

[0066] (1) Traditional industrial safety risk assessment methods often rely on only a single data source (such as historical accident data). The present invention combines multiple data sources such as text descriptions, attribute extraction words, description extraction words, and accident levels to achieve deep fusion of multimodal features, which can more comprehensively capture the characteristic information of accident risks, thereby greatly improving the accuracy of predictions and adaptability to diversified industrial scenarios, and solving the problem of inaccurate predictions caused by traditional single data sources.

[0067] (2) The Transformer network structure used in the model has efficient parallel computing capabilities, which can significantly improve computing speed and efficiency, especially when processing future attributes and long sequence data. Compared with the traditional single-step processing method, the parallel computing capability of Transformer enables the model of the present invention to process large-scale data in a short time and has the ability of real-time prediction, which is especially important for risk assessment scenarios that require rapid response in industrial environments, thereby improving the real-time performance and efficiency of the overall system.

[0068] (3) The present invention uses large language model (LLM) technology to provide enterprises with targeted solutions in real time after completing risk identification. Compared with traditional safety management measures, the real-time solution generation capability of the present invention enables enterprises to take preventive measures in time before accidents occur, effectively reducing the probability of accidents. By generating specific prevention suggestions, the model not only helps enterprises predict risks, but also provides them with actionable solutions, thereby improving the comprehensiveness and timeliness of industrial safety management.

[0069] To verify the effectiveness of the above method, this embodiment selected two highly correlated industrial accident risk data sets as experimental objects and conducted in-depth analysis. The experimental results are shown in Table 1 and Table 2. The empirical distribution method was adopted in the experiment to process the observed data. This method does not rely on any preset theoretical distribution, but gives equal weight to each data point in the sample, intuitively reveals the actual distribution characteristics of the data, and thus provides a unique perspective to observe and analyze the data.

[0070] In order to deeply compare the performance of the present invention with the prior art, this embodiment adopts a variety of common machine learning models and deep learning models, and records the four core indicators of accuracy, precision, recall and F1 score for each model. It can be seen from the results of Table 1 and Table 2 that the performance of the method of the present invention on both data sets is significantly better than that of the prior art, especially in terms of accuracy, precision, recall and F1 score. The performance of the method of the present invention far exceeds other traditional machine learning models and deep learning models, which proves that the present invention is not only advanced in theory, but also has extremely high operability and reliability in practical applications.

[0071] Table 1 Experimental results of the Industrial Safety and Health Analytics Database dataset

[0072]

[0073]

[0074] Table 2 Experimental results of OSHA Accident and Injury Data dataset

[0075]

[0076]

[0077] By comparing with multiple benchmark models, it can be clearly found that the method of the present invention has superior performance in dealing with industrial accident risk assessment problems. This significant advantage further verifies the technological breakthrough brought by the present invention. The present invention realizes data-driven adaptive capabilities through deep learning training of large-scale accident data. The model can continuously absorb new data and make self-improvements and adjustments. This data-driven adaptive capability enables the model to adjust to the dynamic changes in the industrial scene, with greater flexibility and adaptability. When faced with new changes and new risk factors in the industrial environment, the model can adjust its prediction strategy in a timely manner, thereby maintaining efficient risk assessment capabilities.

[0078] If the above method is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0079] The above description of the embodiments is to facilitate the understanding and use of the invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A method for predicting industrial safety and health risks based on machine learning, characterized in that: The following steps are involved: Obtaining industrial scene descriptions, generating structured data through a large language model, inputting the structured data into a trained industrial safety and health risk prediction model, and outputting predicted potential risk levels, wherein the structured data includes accident descriptions, attribute extraction words, description extraction words, and accident levels; Wherein, the industrial safety and health risk prediction model includes a feature extraction module, a feature enhancement module and a feature fusion module; The feature extraction module is used to extract text features from the accident description and the description extraction words, and extract attribute features from the attribute extraction words; The feature enhancement module includes a parallel attribute feature enhancement network, a text feature enhancement network and an accident level embedding network. The attribute feature enhancement network is built based on the Transformer architecture and is used to enhance the attribute features; the text feature enhancement network is built based on the Bi-LSTM architecture and the cross attention mechanism and is used to perform word-level and sentence-level joint enhancement on the text features; the accident level embedding network is used to generate an accident level embedding vector according to the accident level; The feature fusion module is used to splice the outputs of each network in the feature enhancement module to obtain a comprehensive feature vector, and identify the potential risk level based on the comprehensive feature vector.

2. The method for predicting industrial safety and health risks based on machine learning according to claim 1, characterized in that: The specific process of generating structured data through a large language model is as follows: The industrial scene description is embedded in the prompt word template, and the large language model extracts information from the industrial scene description based on the prompt word template; The extracted information is converted into structured data in a list structure through the eval function.

3. The method for predicting industrial safety and health risks based on machine learning according to claim 1, characterized in that: The attribute feature enhancement network includes a splicing layer, a global average pooling layer, a Transformer encoder and an attention layer connected in sequence; the attribute feature includes multiple sub-attribute features, and after the multiple sub-attribute features are spliced ​​through the splicing layer, a compact feature representation is generated through the global average pooling layer and the Transformer encoder; the attention layer performs weighted calculation on the compact feature representation through an attention mechanism to generate enhanced attribute features.

4. The method for predicting industrial safety and health risks based on machine learning according to claim 3 is characterized in that: A mean layer and a random dropout layer are set between the Transformer encoder and the attention layer.

5. The method for predicting industrial safety and health risks based on machine learning according to claim 1, characterized in that: The text feature enhancement network includes a cross attention layer, and a parallel word-level feature layer and a sentence-level feature layer; The word-level feature layer and the sentence-level feature layer are both constructed based on the Bi-LSTM architecture. The word-level feature layer is used to extract word-level attention vectors from the text features, and the sentence-level feature layer is used to extract sentence-level attention vectors from the accident description; the cross-attention layer is used to perform weighted calculation on the word-level attention vectors and the sentence-level attention vectors through a cross-attention mechanism to generate enhanced text features.

6. The method for predicting industrial safety and health risks based on machine learning according to claim 1, characterized in that: The accident level embedding network includes a ReLU activation function layer and a random loss layer connected in sequence.

7. The method for predicting industrial safety and health risks based on machine learning according to claim 1, characterized in that: The industrial safety and health risk prediction model is trained by back propagation and the Adam optimizer is used to update the model parameters.

8. The method for predicting industrial safety and health risks based on machine learning according to claim 1, characterized in that: The industrial safety and health risk prediction model uses a cross entropy loss function during the training process.

9. An electronic device comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images