AI-based mine risk hidden danger intelligent management and control system

Abstract

The application discloses a kind of AI-based mine risk hidden danger intelligent management and control system, comprising: problem entry module, for obtaining user problem text;Multi-intention recognition module, for user problem text as input, through the intention recognition model constructed based on BERT model and LSTM model carries out intention recognition, outputs corresponding multiple intention categories and the description target of each intention category;Question and answer processing module, for according to the description target of each intention category is retrieved, obtains the search information corresponding to each intention category, and then user problem text and the search information of all intention categories are input into the question and answer model based on multilayer Transform, output corresponding problem answer text;Hidden danger analysis module, for realizing risk assessment and hidden danger investigation according to problem answer text.The application accurately identifies multiple intention types in user problem through the combination model of BERT and LSTM, and simultaneously realizes risk assessment and hidden danger investigation through the question and answer results generated by question and answer model.

Description

Technical Field

[0001] This invention relates to the field of Internet big data and risk management, specifically to an AI-based intelligent management and control system for mine risks and hazards. Background Technology

[0002] With the development of the times, energy resources have become increasingly important in social development, and mines are the main source of mineral resources in my country. The core content of safety management in coal mining enterprises mainly includes risk classification and control, and hidden danger investigation and management.

[0003] Currently, the risk assessment phase primarily employs a combination of qualitative and quantitative methods, comprehensively considering factors such as the potential accident types, impact levels, and likelihood of occurrence of risk sources to rate the risks. This process requires identifying potential hazards based on relevant laws and regulations, safety production management requirements, equipment operating status, and work environment data. These hazards are then analyzed qualitatively and quantitatively to assess their probability of occurrence, scope of impact, and degree, thereby determining the risk level, formulating relevant handling and control measures, and continuously improving them. In the hazard identification phase, the main goal is to prevent accidents by discovering and eliminating hidden dangers. Currently, hazard identification primarily involves relevant personnel conducting inspections and assessments based on applicable laws, regulations, and technical standards, using various methods and means to identify and rectify existing hazards, and then inspecting and accepting the findings.

[0004] However, the applicant found that existing methods of risk assessment and hazard identification conducted manually have the following problems: 1) Due to the limitations of human knowledge, comprehensive, scientific, and practical assessments, identifications, and handling are impossible. Furthermore, the dissemination and sharing of human experience is very difficult, resulting in poor accuracy in current risk assessments and hazard identification. 2) Risk assessment and hazard identification involve a large amount of repetitive work, relying heavily on the experience of the personnel involved. This means that human interference and subjectivity significantly affect risk assessment and hazard identification, leading to poor objectivity. Therefore, improving the accuracy and objectivity of risk assessment and hazard identification is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide an AI-based intelligent management and control system for mine risks and hazards, which accurately identifies multiple intent types in user questions through a combination of BERT and LSTM models, and achieves risk assessment and hazard investigation through question-and-answer results generated by a question-and-answer model, thereby improving the accuracy and objectivity of risk assessment and hazard investigation.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] The AI-based intelligent management and control system for mine risks and hazards includes:

[0008] The issue entry module is used to obtain user issue texts that contain risk assessment and hazard investigation intentions;

[0009] The multi-intent recognition module takes user question text as input, performs intent recognition through an intent recognition model built based on BERT and LSTM models, and outputs the corresponding multiple intent categories and the description target of each intent category.

[0010] The question-answering module is used to retrieve information based on the description target of each intent category, obtain the retrieval information corresponding to each intent category, and then input the user's question text and the retrieval information of all intent categories into the question-answering model to output the corresponding question answer text.

[0011] The hazard analysis module is used to perform risk assessment and hazard identification based on the text of the answer to the question.

[0012] Preferably, the processing steps of the intent recognition model are as follows:

[0013] S101: Input the user's question text into the encoder of the BERT model to generate the corresponding hidden vector;

[0014] The formula is expressed as:

[0015] c i =BERT-Encoder(ω 0:n ,i);

[0016] In the formula: c i Represents the hidden vector; ω 0:n This represents the user's question text, where 1 ≤ i ≤ n;

[0017] S102: Input the hidden vector into the feedforward neural network and obtain the intention category probability distribution through the Sigmoid function;

[0018] The formula is expressed as:

[0019] p = Sigmoid(MLP(c i ));

[0020] In the formula: p represents the probability distribution of intent categories, which is a 1*n dimensional vector, where each dimension represents the probability of identifying an intent category; MLP is a fully connected neural network;

[0021] S103: Connect the intent category probability distribution with the hidden vector to obtain the fused probability distribution;

[0022] The formula is expressed as:

[0023]

[0024] Where: h i Represents the fusion probability distribution;

[0025] S104: Input the fusion probability distribution into the LSTM model to generate the corresponding slot filling probability distribution, and output the intent category of each word in the user question text, i.e. the description target, through the Sigmoid function;

[0026] The formula is expressed as:

[0027] r i =Sigmoid(LSTM(h) i ));

[0028] In the formula: r i This represents the intent category of the i-th word in the user's question text, i.e., the i-th descriptive target.

[0029] Preferably, the loss function for training the intent recognition model is as follows:

[0030]

[0031] in

[0032] In the formula: p i r i This represents the probability distribution of intent categories predicted by the intent recognition model and describes the target. This represents the probability distribution of intent categories and the true labels describing the target; y i For r i or p i ; for or

[0033] Preferably, the processing steps of the question-and-answer module are as follows:

[0034] S201: Take multiple intent categories and the description target of each intent category as input;

[0035] S202: Determine the information source for the retrieved information based on the intent category;

[0036] S203: Based on the description target of each intent category, search the corresponding information sources to obtain the search information corresponding to each intent category;

[0037] S204: Input the user's question text and the retrieved information into the question-answering model, and output the corresponding question answer text.

[0038] Preferably, in step S204, the processing steps of the question-answering model are as follows:

[0039] S2041: Encode the user question text and the retrieved information through the embedding layer to obtain the embedded representation of the user question text and the retrieved information, so as to convert the user question text and the retrieved information into a form that the question answering model can understand;

[0040] S2042: The embedded representation of the answer is output by stacking the user question text and retrieval information into multiple layers of Transformers.

[0041] S2043: Decode the embedded representation of the answer to generate the question answer text.

[0042] Preferably, in step S202, the intent categories include safety knowledge base, risk level classification, risk measure formulation, accident hazard investigation and hazard disposal measures;

[0043] The information sources for retrieval include six types of information sources corresponding to safety knowledge base, risk level classification, risk measure formulation, accident hazard investigation and hazard disposal measures.

[0044] Preferably, in step S202, the information source is constructed through the following steps:

[0045] S2021: Obtain public and private documents related to coal mine production business scenarios;

[0046] S2022: Preprocess public and private documents to obtain preprocessed text; the preprocessing includes noise reduction, cleaning, format unification and text segmentation;

[0047] S2023: Segment the preprocessed text to obtain text fragments;

[0048] S2024: Generate text content based on text fragments using a trained large-scale mine safety model; the text content includes text content structure and text content attributes;

[0049] S2025: Extract keywords and determine text type based on text content structure and attributes, and finally classify text fragments into safety knowledge base, risk level classification, risk measure formulation, accident hazard investigation and hazard disposal measures.

[0050] Preferably, in step S202, during retrieval, the cosine similarity between each vector in the description target of the intent category and each vector in the information source is calculated, and the k vectors with the highest cosine similarity in the information source are selected as the retrieval information for the intent category.

[0051] Preferably, in step S204, the question answer text output by the question-answering model includes any one or more of the following: a risk identification list to be confirmed, a risk control list to be confirmed, a hazard investigation standard library, risk and hazard handling measures, automatic risk and hazard classification, and knowledge base question results.

[0052] Preferably, the hazard analysis module can perform one or more of the following functions based on the text of the question answer: risk identification, risk control, hazard investigation, risk and hazard handling, risk and hazard classification, and safety question response.

[0053] Compared with existing technologies, the AI-based intelligent management and control system for mine risks and hazards in this invention has the following advantages:

[0054] The intelligent risk assessment and hazard identification system of this invention predicts user questions' intentions using an intent recognition model built on BERT and LSTM models, resulting in multiple intent categories and descriptive targets for each category. First, BERT captures contextual information from the text, while LSTM excels at handling long-term dependencies in sequential data. Therefore, the combined BERT and LSTM model can more accurately identify various types of intents in user question text, thereby improving the accuracy of risk assessment and hazard identification. Second, by identifying multiple intent categories, the system can handle a wider range of more complex user questions, enhancing its versatility and practicality. Furthermore, each intent category is accompanied by a descriptive target, enabling the system to provide more specific and easily understood feedback, enhancing its interpretability and user-friendliness. Finally, automatic intent identification reduces the need for manual intervention, improving the automation level of hazard identification and risk management.

[0055] This invention, based on the identification of multiple intent categories and the descriptive targets of each intent category, retrieves corresponding information for each intent and inputs the user's question text and all retrieved information into a question-answering model built on a multi-layer Transformer architecture to obtain the question answer text. First, by retrieving information related to the intent category, the question-answering model can generate answers based on broader and more comprehensive information, improving the accuracy and richness of the answers, thereby further enhancing the accuracy of risk assessment and hazard identification. Second, the multi-layer Transformer structure allows the question-answering model to better capture the contextual information of the text, generating more context-appropriate answers, which is beneficial for improving the accuracy of risk assessment and hazard identification. Finally, based on automatic intent recognition, it further achieves automated retrieval and question answering. On the one hand, it can quickly respond to user questions, improving the system's response speed and efficiency; on the other hand, the entire process does not rely on human work experience and is free from subjective human interference, thereby improving the objectivity of risk assessment and hazard identification.

[0056] This invention generates corresponding hazard assessment results and hazard mitigation measures based on the text of the answer to a question. Firstly, the risk and hazard assessment results enable the timely identification of potential risks, providing early warning information for safe production in coal mines. Secondly, the risk and hazard mitigation measures provide a scientific basis for coal mine decision-makers, guiding the formulation of effective risk prevention and control strategies. Finally, timely investigation and handling of hazards can reduce or prevent accidents, ensuring the production safety of coal mines and the personal safety of employees. Attached Figure Description

[0057] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0058] Figure 1 This is a network structure diagram of the multi-intent recognition module in the system;

[0059] Figure 2 This is a network structure diagram of the question-and-answer processing module in the system.

[0060] Figure 3 This is a diagram illustrating the source of the information. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0062] It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the figures, or the orientation or positional relationship commonly used when the product is in use. They are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance. In addition, the terms "horizontal," "vertical," etc., do not mean that the component is required to be absolutely horizontal or suspended, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0063] The following detailed explanation illustrates the specific implementation methods:

[0064] Example:

[0065] This embodiment discloses an AI-based intelligent management and control system for mine risks and hazards.

[0066] The AI-based intelligent management and control system for mine risks and hazards includes:

[0067] The issue entry module is used to obtain user issue texts that contain risk assessment and hazard investigation intentions;

[0068] In this embodiment, the user's question text is similar to "If the methane sensor is suspended vertically, is it in compliance with regulations if it is more than 300mm from the top plate? If not, please provide a rectification plan."

[0069] Combination Figure 1As shown, the multi-intent recognition module takes the user's question text as input, performs intent recognition through an intent recognition model built based on BERT and LSTM models, and outputs the corresponding multiple intent categories and the description target of each intent category.

[0070] In this embodiment, taking the user question text "If the methane sensor is suspended vertically, is it compliant with regulations if it is more than 300mm from the ceiling? If not, please provide a rectification plan." as an example, the intent categories output by the multi-intent recognition module include safety knowledge base query and hazard handling measures. The description target of the safety knowledge base query is: 1) Query target: methane sensor, 2) Status: vertically suspended, 3) Status data: more than 300mm from the ceiling. The description target of the hazard handling measures is: 1) Rectification target description: if the methane sensor is suspended vertically, 2) more than 300mm from the ceiling.

[0071] Combination Figure 2 As shown, the question-answering processing module is used to retrieve information based on the description target of each intent category, obtain the retrieval information corresponding to each intent category, and then input the user's question text and the retrieval information of all intent categories into the question-answering model built on a multi-layer Transformer, and output the corresponding question answer text.

[0072] In this embodiment, the user question text "If the methane sensor is suspended vertically, is it compliant with regulations if it is more than 300mm from the ceiling? If not, please provide a rectification plan." is used as an example. The intent categories output by the multi-intent recognition module include safety knowledge base query and hazard handling measures. The description target of the safety knowledge base query is: 1) Query target: methane sensor, 2) Status: vertically suspended, 3) Status data: more than 300mm from the ceiling. The description target of the hazard handling measures is: 1) Rectification target description: if the methane sensor is suspended vertically, 2) more than 300mm from the ceiling.

[0073] The question and answer processing module outputs the following text: The vertical suspension of the methane sensor complies with regulations, but the distance between the methane sensor and the top plate is more than 300mm, which does not comply with regulations and poses a slight risk and hazard. The distance between the methane sensor and the top plate needs to be rectified to within 250mm.

[0074] The hazard analysis module is used to perform risk assessment and hazard identification based on the text of the answer to the question.

[0075] In this embodiment, the user question text is used as an example: "If the methane sensor is suspended vertically, is it compliant with regulations if it is more than 300mm from the top plate? If not, please provide a rectification plan." The question and answer processing module outputs the following answer text: The vertical suspension of the methane sensor is compliant with regulations, but the distance of the methane sensor from the top plate being more than 300mm is not compliant, as there is a slight wind direction and potential hazard. The distance of the methane sensor from the top plate needs to be rectified to within 250mm.

[0076] The hazard analysis module assessed that the methane sensor was vertically suspended, and the distance from the ceiling to the roof was more than 300mm, which did not meet the regulations and posed a slight risk and hazard. The distance between the methane sensor and the ceiling should be reduced to within 250mm.

[0077] The intelligent risk assessment and hazard identification system of this invention predicts user questions' intentions using an intent recognition model built on BERT and LSTM models, resulting in multiple intent categories and descriptive targets for each category. First, BERT captures contextual information from the text, while LSTM excels at handling long-term dependencies in sequential data. Therefore, the combined BERT and LSTM model can more accurately identify various types of intents in user question text, thereby improving the accuracy of risk assessment and hazard identification. Second, by identifying multiple intent categories, the system can handle a wider range of more complex user questions, enhancing its versatility and practicality. Furthermore, each intent category is accompanied by a descriptive target, enabling the system to provide more specific and easily understood feedback, enhancing its interpretability and user-friendliness. Finally, automatic intent identification reduces the need for manual intervention, improving the automation level of hazard identification and risk management.

[0078] This invention, based on the identification of multiple intent categories and the descriptive targets of each intent category, retrieves corresponding information for each intent and inputs the user's question text and all retrieved information into a question-answering model built on a multi-layer Transformer architecture to obtain the question answer text. First, by retrieving information related to the intent category, the question-answering model can generate answers based on broader and more comprehensive information, improving the accuracy and richness of the answers, thereby further enhancing the accuracy of risk assessment and hazard identification. Second, the multi-layer Transformer structure allows the question-answering model to better capture the contextual information of the text, generating more context-appropriate answers, which is beneficial for improving the accuracy of risk assessment and hazard identification. Finally, based on automatic intent recognition, it further achieves automated retrieval and question answering. On the one hand, it can quickly respond to user questions, improving the system's response speed and efficiency; on the other hand, the entire process does not rely on human work experience and is free from subjective human interference, thereby improving the objectivity of risk assessment and hazard identification.

[0079] This invention generates corresponding hazard assessment results and hazard mitigation measures based on the text of the answer to a question. Firstly, the risk and hazard assessment results enable the timely identification of potential risks, providing early warning information for safe production in coal mines. Secondly, the risk and hazard mitigation measures provide a scientific basis for coal mine decision-makers, guiding the formulation of effective risk prevention and control strategies. Finally, timely investigation and handling of hazards can reduce or prevent accidents, ensuring the production safety of coal mines and the personal safety of employees.

[0080] In practice, the processing steps of the intent recognition model are as follows:

[0081] S101: Input the user's question text into the encoder of the BERT model to generate the corresponding hidden vector;

[0082] The formula is expressed as:

[0083] c i =BERT-Encoder(ω 0:n ,i);

[0084] In the formula: c i Represents the hidden vector; ω 0:n This represents user input (which can be either user-generated questions or commands), where 1 ≤ i ≤ n;

[0085] In this embodiment, the BERT (Bidirectional Encoder Representations from Transformers) model is a pre-trained language representation model based on the Transformer architecture, proposed by Google AI in 2018. The core feature of BERT is that it uses a bidirectional Transformer encoder to learn deep semantic representations of text data.

[0086] The BERT model includes:

[0087] Input layer: The input user question text is converted into a format that the BERT model can process, including word embeddings and positional encoding.

[0088] Transformer Encoder: The BERT model uses a multi-layer Transformer encoder to process the input data. Each Transformer encoder contains a self-attention mechanism and a feed-forward neural network.

[0089] Self-attention mechanism: This is the core of Transformer, allowing the model to take into account all other words in the sentence when processing each word, thereby capturing contextual information.

[0090] Feedforward Neural Network: After each self-attention layer, BERT uses a feedforward neural network to further process the representation at each location.

[0091] Output layer: After pre-training, the BERT model can be used for various downstream tasks, such as text classification, question answering, and named entity recognition. Typically, one or more task-specific layers are added on top of the BERT model for fine-tuning.

[0092] S102: Input the hidden vector into the feed-forward neural network and obtain the intention category probability distribution through the Sigmoid function;

[0093] The formula is expressed as:

[0094] p = Sigmoid(MLP(c i ));

[0095] In the formula: p represents the probability distribution of intent categories, which is a 1*n dimensional vector, where each dimension represents the probability of identifying an intent category; MLP is a fully connected neural network;

[0096] S103: Connect the intent category probability distribution with the hidden vector to obtain the fused probability distribution;

[0097] The formula is expressed as:

[0098]

[0099] Where: h i Represents the fusion probability distribution;

[0100] S104: Input the fusion probability distribution into the LSTM model to generate the corresponding slot filling probability distribution, and output the intent category (one-hot encoding of each token) of each word in the user question text through the Sigmoid function, i.e. the description target;

[0101] The formula is expressed as:

[0102] r i =Sigmoid(LSTM(h) i ));

[0103] In the formula: r i This represents the intent category of the i-th word in the user's question text, i.e., the i-th descriptive target.

[0104] In this embodiment, the LSTM (Long Short-Term Memory) model is a special recurrent neural network (RNN) architecture specifically designed to address the vanishing and exploding gradient problems encountered by traditional RNNs when processing long sequences of data. LSTM maintains long-term dependencies by introducing a series of structures called "gates," enabling the network to effectively learn and remember information in long sequences.

[0105] LSTM models include:

[0106] Cell states are the core of LSTM, running throughout the entire sequence and carrying long-term information. A cell state can be viewed as a "memory" unit, capable of storing or forgetting information.

[0107] The Forget Gate determines which information should be forgotten or retained from the cell state. It outputs a value between 0 and 1 via a sigmoid function, where 0 represents complete forgetting and 1 represents complete retention.

[0108] The input gate determines which new information will be stored in the cell state. It consists of two parts: a sigmoid layer that determines which values ​​will be updated, and a tanh layer that creates a new vector of candidate values ​​that will be added to the state.

[0109] The output gate determines the final output based on the updated cell state. It uses the sigmoid function to determine which parts of the cell state will be output, then processes the cell state through the tanh function (resulting in a value between -1 and 1), and multiplies it by the output of the sigmoid gate to determine the final output.

[0110] This invention employs an intent recognition model based on BERT and LSTM models for intent prediction. Firstly, BERT captures contextual information from the text, while LSTM excels at handling long-term dependencies in sequential data. Therefore, the combined BERT and LSTM model can more accurately identify various types of intent in user question texts, thereby improving the accuracy of subsequent risk prediction. Secondly, by recognizing multiple intent categories, the system can handle a wider range of more complex user questions, enhancing its versatility and practicality. Furthermore, each intent category is accompanied by a descriptive target, enabling the system to provide more specific and easily understood feedback, thus enhancing its interpretability and user-friendliness. Finally, automatic intent recognition reduces the need for manual intervention, improving the automation level of hazard identification and risk management.

[0111] In the specific implementation process, the loss function used when training the intent recognition model is as follows:

[0112]

[0113] in

[0114] In the formula: p i r i This represents the probability distribution of intent categories predicted by the intent recognition model and describes the target. The probability distribution of the intent category and the true labels describing the target (i.e., slot filling); y i For r i or p i ; for or

[0115] In this embodiment, the various intents input by the user are represented as a 1*N dimensional vector. For example, a 1*5 vector can represent whether the user wants to use one or more functions among the following: security knowledge base, risk level classification, risk measure formulation, accident hazard investigation, and hazard handling measures. For example, [1,0,0,0] means the user needs to use the security knowledge base function. Slot filling refers to predicting the probability that each word in the user's input statement represents a certain intent. It can be considered as a probability distribution for intent recognition, but this probability distribution describes the intent category of a certain word under a certain intent.

[0116] In the specific implementation process, the processing steps of the question-and-answer module are as follows:

[0117] S201: Take multiple intent categories and the description target of each intent category as input;

[0118] S202: Determine the information source for the retrieved information based on the intent category;

[0119] S203: Based on the description target of each intent category, search in the corresponding information source to obtain the search information corresponding to each intent category; during the search, calculate the cosine similarity between each vector in the description target of the intent category and each vector in the information source, and select the k vectors with the highest cosine similarity in the information source as the search information for the intent category.

[0120] The formula for calculating cosine similarity is as follows:

[0121]

[0122] In the formula: A and B are two different vectors; ||A|| and ||B|| represent the magnitudes of vector A and vector B, respectively.

[0123] This invention utilizes cosine similarity for information retrieval. Cosine similarity is an efficient similarity calculation method, especially when processing high-dimensional or text data. It assesses the similarity between two vectors by calculating the cosine of the angle between them, without considering the absolute length of the vectors, making the calculation process faster. Secondly, cosine similarity can be applied to various types of information retrieval scenarios, including text retrieval and image retrieval, where it effectively measures the similarity between queries and documents or images. Finally, the formula for calculating cosine similarity is simple and straightforward, easy to understand and implement. It only requires calculating the sum of the products of corresponding elements of the two vectors and then dividing by the product of the magnitudes of the two vectors, enabling the rapid development and deployment of information retrieval systems based on cosine similarity.

[0124] S204: Input the user's question text and the retrieved information into the question-answering model built on a multi-layer Transformer, and output the corresponding question answer text.

[0125] In this embodiment, the question answer text output by the question-answering model includes any one or more of the following: a risk identification list to be confirmed, a risk control list to be confirmed, a hazard investigation standard library, risk and hazard handling measures, automatic risk and hazard classification, and knowledge base question results.

[0126] The hazard analysis module can perform one or more of the following functions based on the text of the answer to the question: risk identification, risk control, hazard investigation, risk and hazard handling, risk and hazard classification, and safety question response.

[0127] In specific implementation, the intent categories include safety knowledge base, risk level classification, risk measure formulation, accident hazard investigation, and hazard handling measures. Combined with... Figure 3 As shown, the information sources for retrieval include six types of information sources corresponding to safety knowledge base, risk level classification, risk measure formulation, accident hazard investigation and hazard disposal measures.

[0128] In this embodiment, if a web search engine is used for retrieval, the search engine's API is first invoked, and the first X (adjustable) links are selected. The text of these links is then segmented into N (unknown) text blocks and vectorized. Finally, k text blocks are selected as the retrieval information. Specifically, a cosine similarity algorithm can be used to select the k most relevant text blocks as the retrieval information.

[0129] Specifically, information sources are constructed through the following steps:

[0130] S2021: Obtain public and private documents related to coal mine production business scenarios;

[0131] In this embodiment, documents are categorized into public and private documents based on their availability. Public documents primarily include mandatory and recommended safety-related standards from national, industry, ministerial, and safety supervision bureau levels. Private documents mainly include coal mining enterprises' own safety management systems, records, and experiences. These documents are used for learning and enhancing the foundation of the large-scale mine safety model.

[0132] This system supports various document types (images, PDFs, Excels, Words) as input. For images, OCR technology is first used to convert the images into text. Then, word segmentation and text cleaning tools are used to perform basic processing on the text to generate standardized text. The system can also segment the text to prevent information loss or confusion and improve reasoning efficiency.

[0133] S2022: Preprocess public and private documents to obtain preprocessed text; the preprocessing includes noise reduction, cleaning, format unification and text segmentation;

[0134] In this embodiment, irrelevant information in the document, such as advertisements, watermarks, headers and footers, is removed; the document is converted to a uniform format, such as PDF to text or XML, to facilitate subsequent processing; and the text is segmented to facilitate subsequent keyword extraction and text classification.

[0135] S2023: Segment the preprocessed text to obtain text fragments (words or vocabulary units);

[0136] S2024: Text content is generated from text fragments using a pre-trained mine safety model. This text content includes text structure and text attributes. The text structure is primarily differentiated across four dimensions: personnel, equipment or devices, environment, and management. It describes the location, equipment, personnel, and behaviors involved, along with their consequences (good or bad), and the presence of danger. Text attributes mainly include attributes related to hazard identification and assessment, risk level evaluation, hazard mitigation measures, hazard system planning, hazard identification methods, hazard identification for each production subsystem, safety training, worker operating procedures, and safety standardization.

[0137] In this embodiment, the large-scale mine safety model can be any existing mature model. The large-scale mine safety model can be constructed and trained using the following methods:

[0138] 1) Data Collection and Preprocessing: Collect data related to mine safety, including public documents, private documents, mine structure, personnel, equipment status, production conditions, environmental parameters, etc. Clean and preprocess the data to ensure its accuracy and completeness. This may include removing duplicate data, handling missing values, and outliers. If necessary, perform feature engineering on the data to extract meaningful features to better describe the mine's safety status.

[0139] 2) Establish Model Structure: Based on the problem requirements and data characteristics, select an appropriate model structure. A large-scale mine safety model can be a deep learning model, such as a neural network or recurrent neural network. First, determine the model's input and output. The input is preprocessed mine safety-related data, and the output is the predicted or classified results of the mine safety status. Then, design the model's hierarchical structure, including the number of hidden layers and the number of neurons in each layer. The selection of these parameters will affect the model's performance and generalization ability.

[0140] 3) Model Training and Optimization: The preprocessed data is divided into training, validation, and test sets. The training set is used to train the model, the validation set is used to tune model parameters and hyperparameters, and the test set is used to evaluate model performance. First, a suitable loss function and optimizer are selected. The loss function measures the difference between the model's predictions and the actual results, and the optimizer tunes the model parameters to minimize the loss function. Then, the model is trained using the training set, updating the model parameters through backpropagation. During training, the validation set can be used for model validation and early stopping to prevent overfitting. Next, model parameters and hyperparameters, such as learning rate, batch size, and regularization terms, are adjusted based on the performance of the validation set. The selection of these parameters will affect the training speed and performance of the model. Finally, the model's performance is evaluated using the test set (evaluation metrics include accuracy, recall, and F1 score), and the model is further adjusted and optimized based on the evaluation results.

[0141] S2025: Based on the text content structure and attributes, extract keywords and determine the text type. Finally, categorize the text fragments into the safety knowledge base, risk level classification, risk measure formulation, accident hazard investigation, and hazard handling measures. This requires vectorization of the text fragments. It also involves designing and constructing the corresponding data structure for storing and managing data for the safety knowledge base, risk level classification, risk measure formulation, accident hazard investigation, and hazard handling measures.

[0142] In this embodiment, keyword extraction refers to using natural language processing (NLP) technology to extract keywords related to safe production from documents; keyword classification refers to matching the extracted keywords with predefined categories (such as safety knowledge base, risk level classification, etc.) to preliminarily determine the category to which the document belongs.

[0143] In practice, the question-answering model is built on the RAG (Retrieval-Augmented Generation) architecture, which combines two key NLP (Natural Language Processing) tasks: retrieval and generation. RAG aims to improve model performance by combining relevant information retrieved from large text datasets with a generation module to produce output text, particularly in tasks involving domain-specific knowledge or requiring external information.

[0144] In the RAG architecture, the question-answering model typically includes the following main steps:

[0145] Indexing: Documents are split into smaller parts (often called "chunks"), these chunks are encoded into vectors, and then stored in a vector database. This step allows models to quickly retrieve document portions relevant to a query.

[0146] Retrieval: Based on the user's query, the model uses some similarity metric (such as cosine similarity) to retrieve the top k most relevant chunks. These chunks contain the information needed to answer the query.

[0147] Generation: The original query and the retrieved chunks are fed into a large language model (LLM) to generate the final answer. This step leverages the text generation capabilities of the LLM to produce a natural and fluent response based on the retrieved relevant information.

[0148] The RAG architecture offers numerous advantages in question-answering systems. First, by combining retrieval and generation tasks, RAG can better utilize information from external knowledge bases and corpora to answer complex questions. Second, the RAG architecture boasts high flexibility and scalability, adapting to different domains and question types. Finally, because RAG incorporates cutting-edge NLP technologies such as Transformer and BERT, it can generate more accurate and natural answers.

[0149] Specifically, the processing steps of the question-answering model in this invention are as follows:

[0150] S2041: The user question text and the retrieved information are encoded through the embedding layer to obtain the embedded representation of the user question text and the retrieved information, so as to convert the user question text and the retrieved information into a form that the question answering model can understand;

[0151] In this embodiment, the question-answering model can also perform additional processing on the retrieved information, such as enhancing the retrieved information through an attention mechanism to highlight the part of the retrieved information that is most relevant to the user's question.

[0152] S2042: The embedded representations of the user's question text and the retrieved information are input into a multi-layered Transformer, and the output is the embedded representation of the answer. The Transformer model uses a self-attention mechanism to capture the dependencies between different positions in the user's question text and the retrieved information, achieving global context modeling. This means the model can consider all information in both the question and the retrieved information simultaneously to more accurately understand the intent and context of the question. The generated answer's embedded representation can be directly generated text (in generative question answering systems) or fragments selected from the retrieved information (in extractive question answering systems). For generative question answering systems, the model outputs a probability distribution representing the likelihood of each word or character being part of the answer, and then selects the sequence with the highest probability as the answer. For extractive question answering systems, the model evaluates the relevance of each fragment in the retrieved information to the question and selects the most relevant fragment as the answer.

[0153] In this embodiment, a self-attention mechanism is used in each layer of the Transformer to process the input. This allows the model to consider the relationship between each position in the input sequence and all other positions, and update the representation of each position based on these relationships. This helps the model capture key information in the question and retrieved information, as well as the relationships between them. In addition to the self-attention mechanism, each layer of the Transformer also includes a feedforward neural network for further processing of the input data. To ensure that the model can be trained deeper without encountering gradient vanishing or exploding problems, the Transformer uses residual connections and layer normalization techniques. By stacking multiple layers of Transformers, this invention allows the model to learn more complex feature representations, improving the accuracy of question answers.

[0154] S2043: Decode the embedded representation of the answer to generate the question answer text.

[0155] In this embodiment, when generating answer text, the question-answering model may employ strategies to ensure that the generated text is relevant to the question, fluent, and meaningful. For example, algorithms such as greedy search and beam search can be used to generate candidate answers, and scoring functions (such as language model scores, BLEU scores, etc.) can be used to evaluate the quality of these candidate answers, selecting the best answer as the output.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. An AI-based intelligent management and control system for mine risks and hazards, characterized in that: include: The issue entry module is used to obtain user issue texts that contain risk assessment and hazard investigation intentions; The multi-intent recognition module takes user question text as input, performs intent recognition through an intent recognition model built based on BERT and LSTM models, and outputs the corresponding multiple intent categories and the description target of each intent category. The processing steps of the intent recognition model are as follows: S101: Input the user's question text into the encoder of the BERT model to generate the corresponding hidden vector; The formula is expressed as: ； In the formula: Represents the hidden vector; This represents the user's question text. ; S102: Input the hidden vector into the feedforward neural network and obtain the intention category probability distribution through the Sigmoid function; The formula is expressed as: ； In the formula: This represents the probability distribution of intent categories, which is a... A dimensional vector, where each dimension represents the probability of recognizing an intent category; It is a fully connected neural network; S103: Connect the intent category probability distribution with the hidden vector to obtain the fused probability distribution; The formula is expressed as: ； In the formula: Represents the fusion probability distribution; S104: Input the fusion probability distribution into the LSTM model to generate the corresponding slot filling probability distribution, and output the intent category of each word in the user question text, i.e. the description target, through the Sigmoid function; The formula is expressed as: ； In the formula: This indicates the first part of the user's question text. The intent category of the word, i.e., the first word One description target; The question-answering module is used to retrieve information based on the description target of each intent category, obtain the retrieval information corresponding to each intent category, and then input the user's question text and the retrieval information of all intent categories into the question-answering model to output the corresponding question answer text. The hazard analysis module is used to perform risk assessment and hazard identification based on the text of the answer to the question.

2. The AI-based intelligent management and control system for mine risks and hazards as described in claim 1, characterized in that: The loss function for training the intent recognition model is as follows: ； in ; In the formula: , This represents the probability distribution of intent categories predicted by the intent recognition model and describes the target. , Represents the probability distribution of intent categories and the true labels describing the target; for or ; for or .

3. The AI-based intelligent management and control system for mine risks and hazards as described in claim 1, characterized in that: The processing steps of the question-and-answer module are as follows: S201: Take multiple intent categories and the description goal of each intent category as input; S202: Determine the information source for the retrieved information based on the intent category; S203: Based on the description target of each intent category, search the corresponding information sources to obtain the search information corresponding to each intent category; S204: Input the user's question text and the retrieved information into the question-answering model, and output the corresponding question answer text.

4. The AI-based intelligent management and control system for mine risks and hazards as described in claim 3, characterized in that: In step S204, the processing steps of the question-answering model are as follows: S2041: Encode the user question text and the retrieved information through the embedding layer to obtain the embedded representation of the user question text and the retrieved information, so as to convert the user question text and the retrieved information into a form that the question answering model can understand; S2042: The embedded representation of the answer is output by stacking the user question text and retrieval information into multiple layers of Transformers. S2043: Decode the embedded representation of the answer to generate the question answer text.

5. The AI-based intelligent management and control system for mine risks and hazards as described in claim 3, characterized in that: In step S202, the intent categories include safety knowledge base, risk level classification, risk measure formulation, accident hazard investigation and hazard disposal measures; The information sources for retrieval include six types of information sources corresponding to safety knowledge base, risk level classification, risk measure formulation, accident hazard investigation and hazard disposal measures.

6. The AI-based intelligent management and control system for mine risks and hazards as described in claim 4, characterized in that: In step S202, the information source is constructed through the following steps: S2021: Obtain public and private documents related to coal mine production business scenarios; S2022: Preprocess public and private documents to obtain preprocessed text; Preprocessing includes noise reduction, cleaning, format unification, and text segmentation. S2023: Segment the preprocessed text to obtain text fragments; S2024: Generate text content based on text fragments using a trained large-scale mine safety model; the text content includes text content structure and text content attributes; S2025: Extract keywords and determine text type based on text content structure and attributes, and finally classify text fragments into safety knowledge base, risk level classification, risk measure formulation, accident hazard investigation and hazard disposal measures.

7. The AI-based intelligent management and control system for mine risks and hazards as described in claim 3, characterized in that: In step S202, during retrieval, the cosine similarity between each vector in the description target of this intent category and each vector in the information source is calculated, and the vector with the highest cosine similarity in the information source is selected. Each vector serves as retrieval information for that intent category.

8. The AI-based intelligent management and control system for mine risks and hazards as described in claim 1, characterized in that: In step S204, the question answer text output by the question-answering model includes any one or more of the following: a list of risks to be identified, a list of risks to be controlled, a library of hazard investigation standards, risk and hazard handling measures, automatic risk and hazard classification, and knowledge base question results.

9. The AI-based intelligent management and control system for mine risks and hazards as described in claim 8, characterized in that: The hazard analysis module can perform one or more of the following functions based on the text of the answer to the question: risk identification, risk control, hazard investigation, risk and hazard handling, risk and hazard classification, and safety question response.

Images