A knowledge question-answering method for coal mine production management based on natural language processing

Abstract

The present invention discloses a coal mine production management knowledge question and answer method based on natural language processing, comprising the following steps: S1, collecting environmental data and user behavior data, and constructing a data set; S2, processing the data set in real time; S3, using a density-based spatial clustering algorithm to construct a user portrait; S4, constructing a knowledge graph, using a graph convolutional network and an improved Transformer to construct a cognitive model, and combining logistic regression and XGBoost algorithms for reasoning; S5, constructing an environmental simulation model through historical data analysis; S6, generating personalized natural language answers based on user portraits, knowledge graphs, and environmental state predictions; S7, collecting user feedback data, and using an adaptive learning algorithm to adjust model parameters and optimization strategies; S8, using a relational database to store plan templates, and continuously optimizing plan templates through historical data analysis. The present invention uses natural language processing and deep learning to realize intelligent question and answer for coal mine production management, and has the advantages of high efficiency and strong personalization.

Description

Technical Field

[0001] The present invention relates to the technical fields of natural language processing and deep learning, and particularly relates to a method for answering questions about coal mine production management based on natural language processing. Background Art

[0002] Coal mine production management is a complex and highly dynamic process, involving a large amount of data collection, analysis, and decision-making. Traditional coal mine production management systems mainly rely on manual monitoring and empirical judgment, and have problems such as low efficiency, slow response speed, and insufficient decision-making accuracy. These defects are particularly obvious in the face of emergencies and complex environmental changes, and may lead to serious production safety accidents and economic losses.

[0003] With the development of information technology, more and more coal mine production management systems have begun to introduce automated monitoring and data analysis technologies. For example, environmental data (such as temperature, humidity, equipment status, gas concentration, etc.) at the coal mine production site is monitored in real time through sensor technology, and user behavior data is collected through log analysis technology. Although these technologies have improved the efficiency and accuracy of data collection to a certain extent, there are still significant deficiencies in data processing, analysis, and decision support.

[0004] Existing coal mine production management systems often lack the ability to effectively process and analyze massive data. Especially in scenarios with high real-time requirements, traditional centralized data processing modes have problems such as high latency and low processing efficiency. The introduction of edge computing provides a solution to this problem. By deploying computing nodes at the coal mine site, data processing and analysis can be performed at the data source, significantly improving the response speed and processing efficiency of the system. However, there is still much room for improvement in the complexity and intelligence level of data processing algorithms in existing edge computing technologies.

[0005] In terms of user behavior analysis, traditional analysis methods mainly rely on simple statistical analysis and rule matching, and cannot accurately capture the complexity and diversity of user behavior. In recent years, density-based spatial clustering algorithms (such as DBSCAN) and association rule analysis technologies have gradually been applied to user behavior analysis, and can more accurately describe user behavior patterns and preferences. However, these methods still face challenges in processing high-dimensional and large-scale data, and need to be combined with more advanced data processing algorithms and computing technologies.

[0006] As an emerging knowledge management and representation technology, the knowledge graph can represent complex knowledge in a structured and systematic manner and support efficient knowledge retrieval and reasoning. Through natural language processing and information extraction technologies, knowledge can be automatically extracted from a large number of documents and added to the knowledge graph. However, existing knowledge graph construction and reasoning methods still have deficiencies in dealing with dynamic and uncertain knowledge, and more advanced deep learning models need to be combined to improve the accuracy and efficiency of knowledge reasoning.

[0007] In terms of environmental state prediction, traditional time series analysis methods perform poorly in dealing with complex non-linear and multi-modal data. The combination of mixture density networks and long short-term memory networks provides a more powerful tool for time series prediction, capable of capturing long-term and short-term dependencies and multi-modal characteristics in the data. However, these methods still face problems such as high model complexity and large computational resource requirements in practical applications and need further optimization and improvement.

[0008] Personalized answer generation is one of the important applications of natural language processing technology in coal mine production management. Existing methods are mainly based on simple template matching and rule generation and cannot meet the diverse and personalized needs of users.

[0009] Therefore, how to provide a knowledge Q&A method for coal mine production management based on natural language processing is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0010] An object of the present invention is to propose a knowledge Q&A method for coal mine production management based on natural language processing. The present invention makes full use of natural language processing, edge computing, and deep learning to construct an intelligent knowledge Q&A system for coal mine production management. By collecting and analyzing environmental data and user behavior data in real time, personalized natural language answers are generated, and the emergency plan is dynamically optimized. The system has the advantages of efficient processing, personalized service, and high security, effectively improving the intelligent level of coal mine production management, significantly enhancing the response speed and decision-making accuracy of the system, and providing more efficient and secure support for coal mine production.

[0011] A knowledge Q&A method for coal mine production management based on natural language processing according to an embodiment of the present invention includes the following steps:

[0012] S1. Use sensor technology to collect environmental data at the coal mine production site, and use log analysis technology to collect and analyze user behavior data to construct a data set;

[0013] S2. Deploy edge computing nodes at the coal mine site, perform edge analysis using a data processing algorithm that combines variational autoencoders and generative adversarial networks, and perform real-time processing on the data set;

[0014] S3. Use the density-based spatial clustering algorithm to construct a user profile, analyze user behaviors and preferences through association rules, and combine the environmental data at the production site with the user profile;

[0015] S4. Construct a knowledge graph, perform knowledge extraction through natural language processing and information extraction technologies, use graph convolutional networks and improved Transformers to construct a cognitive model, and perform reasoning by combining logistic regression and XGBoost algorithms;

[0016] S5. Construct an environmental simulation model through historical data analysis, use the combination of edge computing and cloud computing for real-time data fusion, and use a time series prediction algorithm that combines a mixture density network and a long short-term memory network for environmental state prediction;

[0017] S6. Generate personalized natural language answers based on the user profile, knowledge graph, and environmental state prediction;

[0018] S7. Collect user feedback data, analyze it through data analysis tools, and use an adaptive learning algorithm to adjust model parameters and optimization strategies;

[0019] S8. Use a relational database to store the emergency plan template, use a rule engine combined with the knowledge graph to generate a rule-based emergency plan, automatically generate an emergency plan in combination with the environmental simulation and prediction results, and continuously optimize the emergency plan template through historical data analysis.

[0020] Further, the specific steps of S2 include:

[0021] S21. Use edge computing nodes for preliminary data cleaning and preprocessing, including noise filtering, missing value filling, and data standardization processing, to form a preprocessed dataset X;

[0022] S22. Input the preprocessed dataset X into the variational autoencoder model, and use the variational autoencoder to extract features and reduce the dimension of the data to generate a latent variable Z:

[0023] Z = μ + σ ⊙ ∈;

[0024] where μ represents the mean, σ represents the variance, ∈ ~ N(0, I), ⊙ represents element-wise multiplication, and I represents the identity matrix;

[0025] S23. Input the latent variable Z into the generative adversarial network, and the generative adversarial network includes a generator G and a discriminator D;

[0026] The generator G generates a new data sample G(Z) according to the latent variable Z:

[0027] G(Z) = tanh(W g ·Z + b g);

[0028] Among them, W g represents the weight of the generator, and b g represents the bias of the generator;

[0029] The discriminator D compares the generated data sample G(Z) with the real data set X, and judges the authenticity of the sample through the discriminator D:

[0030] D(G(Z)) = f(W d ·G(Z) + b d );

[0031] Among them, W d represents the weight of the discriminator, and b d represents the bias of the discriminator, and f represents the activation function;

[0032] S24. Jointly train the variational autoencoder and the generative adversarial network through the backpropagation algorithm, optimize the parameters of the generator and the discriminator, so that the generated data sample gradually approaches the real data distribution. The specific optimization objective function is:

[0033]

[0034] Among them, p data (x) represents the probability density function of the real data distribution, x represents the data point sampled from the real data distribution, and p z (z) represents the probability distribution of the latent variable in the latent space, and z represents the latent variable sampled from the latent space;

[0035] S25. Process the real-time collected data through the trained variational autoencoder and generative adversarial network to generate high-quality processed data

[0036] Furthermore, the S25 specifically includes:

[0037]

[0038] Among them, α represents the weight of the attention mechanism, and Attention(D, G(Z)) represents the attention-weighted fusion of the data sample G(Z) and the real data.

[0039] Furthermore, the S3 specifically includes:

[0040] S31. Use the density-based spatial clustering algorithm to perform clustering analysis on the data generated in step S25 ;

[0041] Set two parameters: the radius parameter ∈ and the minimum sample parameter MinPts;

[0042] For each data point p in the dataset i , calculate the number N(p i ) of its neighboring data points within a radius ∈:

[0043]

[0044] where dist(p i , p j ) represents the distance between data point p i and data point p j ;

[0045] Classify the data points into core points, boundary points, and noise points. Among them, core points satisfy that the number N(p i ) of neighboring data points is greater than or equal to MinPts, boundary points satisfy that the number of neighboring data points is less than MinPts but within the neighborhood range of core points, and noise points do not satisfy the conditions of both core points and boundary points;

[0046] Cluster the core points and the points within their neighborhoods, and group the connected core points and the points directly density-reachable from them into the same class to form clusters of user behavior patterns;

[0047] S32. Perform association rule analysis on the clustering results;

[0048] Calculate frequent item sets, and generate frequent item set F through the Apriori algorithm or the FP-Growth algorithm. The frequent item set F satisfies the support threshold Support(A):

[0049]

[0050] where Count(A) represents the number of transactions containing item set A, and N represents the total number of transactions;

[0051] Generate association rules. The form of the association rule is A → B, where A and B represent elements in the frequent item set and satisfy the confidence threshold Confidence(A → B):

[0052]

[0053] Evaluate the effectiveness of the association rules, and measure the relevance of the rules using lift(A → B):

[0054]

[0055] S33. Perform time series decomposition on the environmental data, decomposing it into a trend component, a seasonal component, and a residual component, expressed as:

[0056] y t = T t+S t +R t ;

[0057] where y t represents time series data, T t represents the trend component, S t represents the seasonal component, R t represents the residual component;

[0058] S34. Smooth the trend component using the moving average method to identify the long-term change trend. The moving average formula for T t is:

[0059]

[0060] where n represents the size of the moving window;

[0061] S35. Identify the periodic changes in the seasonal component using the fast Fourier transform. The Fourier transform formula is:

[0062]

[0063] where X k represents the frequency domain signal, x t represents the time domain signal, and N represents the signal length;

[0064] S36. Combine the time series features of the user behavior pattern and environmental data to generate a user profile, which includes the user's operation habits, behavior preferences, and response information to different environmental conditions.

[0065] Furthermore, the specific steps of S4 include:

[0066] S41. Process the coal mine production management documents through natural language processing technology, including word segmentation, part-of-speech tagging, and named entity recognition. The word segmentation and part-of-speech tagging use the SpaCy tool, and the named entity recognition adopts the BERT model;

[0067] S42. Extract knowledge from the processed documents, including entities, attributes, and relationships. Adopt information extraction technology and add the extracted knowledge to the knowledge graph to form a complete knowledge graph;

[0068] S43. Use the graph convolutional network to perform embedding representation learning on the knowledge graph to generate a low-dimensional vector representation for each node. Let the graph be G, the node be v, and the neighbor nodes be N(v). Then the representation h v of node v is calculated as:

[0069]

[0070] where σ represents the activation function, W(l) denotes the weight matrix of the l-th layer, and aggregate denotes the aggregation function. denotes the representation of node v at the l-th layer. denotes the representation of node u at the l-th layer. denotes the representation of node v at the l+1-th layer.

[0071] S44. Input the embedding representation result of the graph convolutional network into the improved Transformer model for further processing. The improved Transformer model includes an adaptive multi-head attention mechanism and dynamic positional encoding to generate a more contextually understanding knowledge representation. Among them, the adaptive multi-head attention score α ij The calculation formula is:

[0072]

[0073] where a represents the adaptive attention weight vector, || represents the vector concatenation operation, W represents the weight matrix, and h i denotes the feature representation of node i, and h j denotes the feature representation of node j, and PE(i,j) represents the dynamic positional encoding of nodes i and j:

[0074]

[0075] where i and j represent the indices of the nodes, d represents the position dimension, and D represents the total embedding dimension.

[0076] The improved Transformer model plays the following roles in the coal mine production management knowledge Q&A system:

[0077] Enhance context understanding ability: Through the adaptive multi-head attention mechanism, the improved Transformer can more effectively capture the relationships between different parts of the input data. Especially for complex knowledge graphs and long-distance dependencies, the improved Transformer can better understand and process context information.

[0078] Improve the flexibility of positional encoding: The dynamic positional encoding allows the model to adaptively adjust the positional encoding according to the characteristics of the input data, thereby improving the adaptability of the model in different application scenarios. Compared with traditional fixed positional encoding, the dynamic positional encoding can better handle the relationships and structural information between different nodes.

[0079] Optimize the effect of the attention mechanism: By improving the score calculation formula in the attention mechanism and introducing dynamic positional encoding, the improved Transformer can more accurately calculate the attention weights between nodes, thereby improving the model's ability to identify key nodes and important information.

[0080] Improving the generalization ability of the model: The improved Transformer can better handle different types of input data by introducing adaptive multi-head attention and dynamic positional encoding, enhancing the generalization ability of the model and making it perform more excellently in dealing with diverse problems in coal mine production management.

[0081] The role of the improved Transformer in question answering generation: When generating personalized natural language answers, the improved Transformer can utilize its enhanced context understanding ability and positional encoding mechanism to generate more accurate and context-compliant answers, improving the intelligence level and user satisfaction of the question answering system.

[0082] S45. Classify the nodes in the knowledge graph through the logistic regression algorithm and predict the categories of unknown nodes. The logistic regression prediction formula is:

[0083]

[0084] where x represents the input feature, w represents the weight, b represents the bias, and P(y = 1|x) represents the probability that the node belongs to class 1;

[0085] S46. Use the XGBoost algorithm to perform regression analysis on the nodes and relationships in the knowledge graph. XGBoost makes predictions by weighted averaging the results of multiple decision trees:

[0086]

[0087] where K represents the number of trees, represents the predicted value of sample i, and f k represents the prediction result of the kth tree;

[0088] S47. Combine the results of the logistic regression and XGBoost algorithms to generate the final inference result, improve the inference ability of the knowledge graph, and provide an accurate and comprehensive knowledge basis for the coal mine production management knowledge question answering system.

[0089] Furthermore, the specific steps of S5 are as follows:

[0090] S51. Construct a mixture density network;

[0091] S52. Construct a long short-term memory network. The network structure includes an input layer, an LSTM layer, and an output layer. The input layer receives the parameters output by the MDN. The LSTM layer is used to capture the long-term and short-term dependencies in the time series, and the output layer generates the prediction result;

[0092] S53. Combine the mixture density network and the long short-term memory network to perform environmental state prediction. The prediction formula is:

[0093]

[0094] Among them, represents the predicted value at time point t, y t represents the parameter input generated by the mixture density network, h t-1 represents the hidden state at the previous time point, c t-1 represents the cell state at the previous time point;

[0095] S54. Store the prediction result in the cloud computing platform and provide it to the knowledge Q&A system through the interface for query and use.

[0096] Furthermore, the specific steps of S51 include:

[0097] An input layer, a hidden layer, and an output layer. The input layer receives the fused comprehensive data set. The hidden layer is implemented by a multi-layer perceptron, and the output layer generates the parameters of the mixture density model:

[0098]

[0099] Among them, y t represents the predicted value at time point t, π j represents the weight of the j-th Gaussian distribution, N(μ j ,σ j ) represents the probability density function of the Gaussian distribution, μ j represents the mean of the j-th Gaussian distribution, σ j represents the standard value of the j-th Gaussian distribution.

[0100] Furthermore, the specific steps of S52 include:

[0101] i t =σ(W i ·[h t-1 ,x t +b i );

[0102] f t =σ(W f ·[h t-1 ,x t +b f );

[0103] o t =σ(W o ·[h t-1 ,x t +b o );

[0104] g t =tanh(W g· [h t-1 , x t + b g );

[0105] c t = f t ⊙ c t-1 + i t ⊙ g t ;

[0106] h t = o t ⊙ tanh(c t );

[0107] Wherein, i t represents the activation vector of the input gate, σ represents the activation function, W i , W f , W o and W g represent the weight matrices input to each gate and the candidate memory state, b i , b f , b o and b g represent the bias terms input to each gate and the candidate memory state, h t-1 represents the hidden state at the previous time point, x t represents the input data at the current time point, f t represents the activation vector of the forget gate, o t represents the activation vector of the output gate, g t represents the activation vector of the candidate memory state, c t represents the cell state at the current time point, c t-1 represents the cell state at the previous time point, ⊙ represents the element-wise multiplication operation, h t represents the hidden state at the current time point.

[0108] Furthermore, the S7 specifically includes:

[0109] S71. In the adaptive learning module, construct the reward function R;

[0110] S72. Based on the optimization result of the reward function R, use the adaptive learning algorithm to update the model parameters of the system, so that the system can gradually improve its response ability to user requirements;

[0111] S73. Regularly apply the updated model parameters to the system operation, and monitor and collect the new round of user feedback data to form a feedback loop to ensure that the system can continuously optimize and improve its performance;

[0112] S74. Save the adjusted model parameters and optimization strategies to the parameter storage module to ensure that the system can quickly recover to the optimal state during system updates and maintenance.

[0113] Furthermore, the S71 specifically includes:

[0114]

[0115] Among them, S represents the user satisfaction score, T represents the system response time, F represents the number of new questions raised by users, U represents the number of opinions and suggestions provided by users, and α, β, γ, and δ represent weight coefficients.

[0116] The beneficial effects of the present invention are:

[0117] Through the method for answering questions about coal mine production management based on natural language processing of the present invention, we have achieved significant beneficial effects in multiple aspects. First, by using sensor technology and log analysis technology to collect and analyze the environmental data and user behavior data at the coal mine production site, a data set is constructed to ensure the comprehensiveness and accuracy of the data. Combining edge computing technology, real-time analysis and processing are carried out through the data processing algorithms of variational autoencoders and generative adversarial networks, significantly improving the efficiency and real-time performance of data processing and reducing the system response time.

[0118] In terms of user portrait and behavior analysis, by using the density-based spatial clustering algorithm and association rule analysis technology, it is possible to more accurately capture user behavior patterns and preferences, combine environmental data with the user portrait, and provide more personalized and relevant services. By constructing a knowledge graph and combining natural language processing and information extraction technology, a systematic and structured representation of knowledge is achieved. The combination of graph convolutional networks and improved Transformer models enhances the accuracy and efficiency of knowledge reasoning. Combining logical regression and XGBoost algorithms for reasoning further improves the intelligent level of the system.

[0119] In terms of environmental state prediction, by constructing an environmental simulation model through historical data analysis and using a time series prediction algorithm that combines a mixture density network and a long short-term memory network, it is possible to more accurately predict complex environmental state changes, providing forward-looking information support for the system. Combining user portraits, knowledge graphs, and environmental state predictions to generate personalized natural language answers makes the answer generation more intelligent and flexible, capable of meeting the diverse needs of users.

[0120] By collecting user feedback data and analyzing and optimizing it using adaptive learning algorithms, the system can continuously adjust and improve based on user feedback, forming a feedback loop, which enhances the system's self-optimization ability and user satisfaction. In terms of the generation and optimization of emergency plans, combined with the knowledge graph and environmental simulation results, rules-based emergency plans are automatically generated to ensure the timeliness and effectiveness of emergency responses. By continuously optimizing the plan templates through historical data analysis, the applicability and accuracy of the emergency plans are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0121] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0122] Figure 1 is a flowchart of a method for knowledge answering in coal mine production management based on natural language processing proposed by the present invention;

[0123] Figure 2 is a flowchart for constructing a user profile of a method for knowledge answering in coal mine production management based on natural language processing proposed by the present invention;

[0124] Figure 3 is a flowchart for predicting the environmental state of a method for knowledge answering in coal mine production management based on natural language processing proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0125] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0126] Refer to Figures 1-3 , a method for knowledge answering in coal mine production management based on natural language processing, includes the following steps:

[0127] S1. Use sensor technology to collect environmental data at the coal mine production site, use log analysis technology to collect and analyze user behavior data, and construct a data set;

[0128] S2. Deploy edge computing nodes at the coal mine site, perform edge analysis using a data processing algorithm that combines variational autoencoders and generative adversarial networks, and perform real-time processing on the data set;

[0129] S3. Use the density-based spatial clustering algorithm to construct a user profile, analyze user behavior and preferences through association rules, and combine the environmental data at the production site with the user profile;

[0130] S4. Construct a knowledge graph, extract knowledge through natural language processing and information extraction technologies, build a cognitive model using graph convolutional networks and an improved Transformer, and perform reasoning in combination with logistic regression and XGBoost algorithms;

[0131] S5. Build an environmental simulation model through historical data analysis, use the combination of edge computing and cloud computing for real-time data fusion, and use a time series prediction algorithm that combines a mixture density network and a long short-term memory network for environmental state prediction;

[0132] S6. Generate personalized natural language answers based on the user profile, knowledge graph, and environmental state prediction;

[0133] S7. Collect user feedback data, analyze it through data analysis tools, and use an adaptive learning algorithm to adjust model parameters and optimization strategies;

[0134] S8. Store the emergency plan template in a relational database, use a rule engine combined with the knowledge graph to generate a rule-based emergency plan, automatically generate an emergency plan in combination with the environmental simulation and prediction results, and continuously optimize the emergency plan template through historical data analysis.

[0135] In this embodiment, the specific steps of S2 include:

[0136] S21. Use edge computing nodes for preliminary data cleaning and preprocessing, including noise filtering, missing value filling, and data standardization processing, to form a preprocessed dataset X;

[0137] S22. Input the preprocessed dataset X into a variational autoencoder model, and use the variational autoencoder to extract features and reduce the dimensionality of the data to generate a latent variable Z:

[0138] Z = μ + σ ⊙ ∈;

[0139] where μ represents the mean, σ represents the variance, ∈ ~ N(0, I), ⊙ represents element-wise multiplication, and I represents the identity matrix;

[0140] S23. Input the latent variable Z into a generative adversarial network, and the generative adversarial network includes a generator G and a discriminator D;

[0141] The generator G generates a new data sample G(Z) based on the latent variable Z:

[0142] G(Z) = tanh(W g ·Z + b g );

[0143] where W g represents the weight of the generator, and b g represents the bias of the generator;

[0144] The discriminator D compares the generated data samples G(Z) with the real data set X, and judges the authenticity of the samples through the discriminator D:

[0145] D(G(Z)) = f(W d ·G(Z) + b d );

[0146] where W d represents the weights of the discriminator, b d represents the bias of the discriminator, and f represents the activation function;

[0147] S24. Jointly train the variational autoencoder and the generative adversarial network through the backpropagation algorithm, optimize the parameters of the generator and the discriminator, so that the generated data samples gradually approximate the real data distribution. The specific optimization objective function is:

[0148]

[0149] where p data (x) represents the probability density function of the real data distribution, x represents the data points sampled from the real data distribution, p z (z) represents the probability distribution of the latent variables in the latent space, and z represents the latent variables sampled from the latent space;

[0150] S25. Process the real-time collected data through the trained variational autoencoder and generative adversarial network to generate high-quality processed data

[0151] In this embodiment, the S25 specifically includes:

[0152]

[0153] where α represents the weight of the attention mechanism, and Attention(D, G(Z)) represents the attention-weighted fusion of the data sample G(Z) and the real data.

[0154] In this embodiment, the S3 specifically includes:

[0155] S31. Use the density-based spatial clustering algorithm to perform clustering analysis on the data generated in step S25 ;

[0156] Set two parameters: the radius parameter ∈ and the minimum sample parameter MinPts;

[0157] For each data point p in the data set i , calculate the number of neighboring data points N(p i ) within the radius ∈:

[0158]

[0159] Among them, dist(p i , p j ) represents the distance between data point p i and data point p j ;

[0160] The data points are divided into core points, boundary points, and noise points. Among them, core points satisfy that the number of neighborhood data points N(p i ) is greater than or equal to MinPts. Boundary points satisfy that the number of neighborhood data points is less than MinPts but within the neighborhood range of core points. Noise points do not satisfy the conditions of both core points and boundary points;

[0161] Cluster the core points and the points within their neighborhoods, and group the connected core points and the points directly density-reachable from them into the same class to form clusters of user behavior patterns;

[0162] S32. Perform association rule analysis on the clustering results;

[0163] Calculate frequent item sets, and generate frequent item set F through the Apriori algorithm or the FP-Growth algorithm. Frequent item set F satisfies the support threshold Support(A):

[0164]

[0165] Among them, Count(A) represents the number of transactions containing item set A, and N represents the total number of transactions;

[0166] Generate association rules. The form of the association rule is A → B, where A and B represent elements in the frequent item set and satisfy the confidence threshold Confidence(A → B):

[0167]

[0168] Evaluate the effectiveness of the association rules, and use lift(A → B) to measure the correlation of the rules:

[0169]

[0170] S33. Decompose the environmental data into a time series, decomposed into a trend component, a seasonal component, and a residual component, expressed as:

[0171] y t = T t + S t + R t ;

[0172] Among them, y t represents the time series data, Tt represents the trend component, S t represents the seasonal component, R t represents the residual component;

[0173] S34. Smooth the trend component using the moving average method to identify the long-term change trend. The moving average formula T t is:

[0174]

[0175] where n represents the size of the moving window;

[0176] S35. Identify the periodic changes in the seasonal component using the fast Fourier transform. The Fourier transform formula is:

[0177]

[0178] where X k represents the frequency-domain signal, x t represents the time-domain signal, and N represents the signal length;

[0179] S36. Combine the time series features of the user behavior pattern and environmental data to generate a user profile, which includes the user's operation habits, behavior preferences, and response information to different environmental conditions.

[0180] In this embodiment, the S4 specifically includes:

[0181] S41. Process the coal mine production management documents through natural language processing technology, including word segmentation, part-of-speech tagging, and named entity recognition. The word segmentation and part-of-speech tagging use the SpaCy tool, and the named entity recognition adopts the BERT model;

[0182] S42. Extract knowledge from the processed documents, including entities, attributes, and relationships. Adopt information extraction technology and add the extracted knowledge to the knowledge graph to form a complete knowledge graph;

[0183] S43. Use the graph convolutional network to perform embedded representation learning on the knowledge graph to generate a low-dimensional vector representation for each node. Let the graph be G, the node be v, and the neighbor nodes be N(v). Then the representation h v of the node v has the following calculation formula:

[0184]

[0185] where σ represents the activation function, W (l) represents the weight matrix of the l-th layer, aggregate represents the aggregation function, represents the representation of the node v at the l-th layer, Denote the representation of node u at layer l, Denote the representation of node v at layer l+1;

[0186] S44. Input the embedding representation results of the graph convolutional network into the improved Transformer model for further processing. The improved Transformer model includes an adaptive multi-head attention mechanism and dynamic positional encoding to generate a more contextually understood knowledge representation. Among them, the adaptive multi-head attention score α ij The calculation formula is:

[0187]

[0188] where a represents the adaptive attention weight vector, || represents the vector concatenation operation, W represents the weight matrix, h i Denote the feature representation of node i, h j Denote the feature representation of node j, and PE(i,j) represents the dynamic positional encoding between node i and node j:

[0189]

[0190] where i and j represent the indices of the nodes, d represents the position dimension, and D represents the total embedding dimension;

[0191] The improved Transformer model plays the following roles in the coal mine production management knowledge Q&A system:

[0192] Enhance context understanding ability: Through the adaptive multi-head attention mechanism, the improved Transformer can more effectively capture the relationships between different parts of the input data. Especially for complex knowledge graphs and long-range dependencies, the improved Transformer can better understand and process context information.

[0193] Improve the flexibility of positional encoding: The dynamic positional encoding allows the model to adaptively adjust the positional encoding according to the characteristics of the input data, thereby improving the adaptability of the model in different application scenarios. Compared with traditional fixed positional encoding, the dynamic positional encoding can better handle the relationships and structural information between different nodes.

[0194] Optimize the effect of the attention mechanism: By improving the score calculation formula in the attention mechanism and introducing dynamic positional encoding, the improved Transformer can more accurately calculate the attention weights between each node, thereby improving the model's ability to identify key nodes and important information.

[0195] Enhancing the generalization ability of the model: The improved Transformer can better handle different types of input data by introducing adaptive multi-head attention and dynamic positional encoding, enhancing the model's generalization ability and making it perform more excellently when dealing with diverse problems in coal mine production management.

[0196] The role of the improved Transformer in question answering generation: When generating personalized natural language answers, the improved Transformer can utilize its enhanced context understanding ability and positional encoding mechanism to generate more accurate and context-compliant answers, improving the intelligence level and user satisfaction of the question answering system.

[0197] S45. Classify the nodes in the knowledge graph through the logistic regression algorithm and predict the categories of unknown nodes. The logistic regression prediction formula is:

[0198]

[0199] where x represents the input feature, w represents the weight, b represents the bias, and P(y = 1|x) represents the probability that the node belongs to class 1;

[0200] S46. Use the XGBoost algorithm to perform regression analysis on the nodes and relationships in the knowledge graph. XGBoost makes predictions by weighted averaging the results of multiple decision trees:

[0201]

[0202] where K represents the number of trees, represents the predicted value of sample i, and f k represents the prediction result of the kth tree;

[0203] S47. Combine the results of the logistic regression and XGBoost algorithms to generate the final inference result, improve the inference ability of the knowledge graph, and provide an accurate and comprehensive knowledge basis for the coal mine production management knowledge question answering system.

[0204] In this embodiment, the specific content of S5 includes:

[0205] S51. Construct a mixture density network;

[0206] S52. Construct a long short-term memory network. The network structure includes an input layer, an LSTM layer, and an output layer. The input layer receives the parameters output by the MDN, the LSTM layer is used to capture the long-term and short-term dependencies in the time series, and the output layer generates the prediction result;

[0207] S53. Combine the mixture density network and the long short-term memory network to perform environmental state prediction. The prediction formula is:

[0208]

[0209] Among them, represents the predicted value at time point t, y t represents the parameter input generated by the mixture density network, h t-1 represents the hidden state at the previous time point, c t-1 represents the cell state at the previous time point;

[0210] S54. Store the prediction result in the cloud computing platform and provide it to the knowledge Q&A system for query and use through an interface.

[0211] In this embodiment, S51 specifically includes:

[0212] An input layer, a hidden layer, and an output layer. The input layer receives the fused comprehensive data set. The hidden layer is implemented by a multi-layer perceptron, and the output layer generates the parameters of the mixture density model:

[0213]

[0214] Among them, y t represents the predicted value at time point t, π j represents the weight of the j-th Gaussian distribution, N(μ j , σ j ) represents the probability density function of the Gaussian distribution, μ j represents the mean of the j-th Gaussian distribution, σ j represents the standard value of the j-th Gaussian distribution.

[0215] In this embodiment, S52 specifically includes:

[0216] i t =σ(W i ·[h t-1 , x t +b i );

[0217] f t =σ(W f ·[h t-1 , x t +b f );

[0218] o t =σ(W o ·[h t-1 , x t +b o );

[0219] g t =tanh(W g ·[h t-1 , xt +b g );

[0220] c t =f t ⊙c t-1 +i t ⊙g t ;

[0221] h t =o t ⊙tanh(c t );

[0222] wherein, i t represents the activation vector of the input gate, σ represents the activation function, W i , W f , W o and W g represent the weight matrices input to each gate and the candidate memory state, b i , b f , b o and b g represent the bias terms input to each gate and the candidate memory state, h t-1 represents the hidden state at the previous time point, x t represents the input data at the current time point, f t represents the activation vector of the forget gate, o t represents the activation vector of the output gate, g t represents the activation vector of the candidate memory state, c t represents the cell state at the current time point, c t-1 represents the cell state at the previous time point, ⊙ represents the element-wise multiplication operation, h t represents the hidden state at the current time point.

[0223] In this embodiment, the S7 specifically includes:

[0224] S71. In the adaptive learning module, construct the reward function R;

[0225] S72. Based on the optimization result of the reward function R, use the adaptive learning algorithm to update the model parameters of the system, so that the system can gradually improve its response ability to user requirements;

[0226] S73. Regularly apply the updated model parameters to the system operation, and monitor and collect the new round of user feedback data to form a feedback loop to ensure that the system can continuously optimize and improve its performance;

[0227] S74. Save the adjusted model parameters and optimization strategies to the parameter storage module to ensure that the system can quickly recover to the best state during system update and maintenance.

[0228] In this implementation mode, the S71 specifically includes:

[0229]

[0230] Where S represents the user satisfaction score, T represents the system response time, F represents the number of new questions raised by users, U represents the number of opinions and suggestions provided by users, and α, β, γ and δ represent weight coefficients.

[0231] Example 1:

[0232] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a large coal mine production base. The base has a vast mining area and a complex underground environment. The mining area faces challenges of high temperature, high humidity and changes in toxic gas concentrations all year round. The traditional coal mine production management method mainly relies on manual monitoring and empirical judgment. This method has problems such as slow response speed and insufficient decision-making accuracy when facing emergencies and complex environmental changes, which greatly affects production efficiency and safety.

[0233] In order to improve the level of intelligence in coal mine production management, we introduced a coal mine production management knowledge question-and-answer system based on natural language processing in the mine. The system collects environmental data of the mine in real time through sensor technology, including temperature, humidity, equipment status, and gas concentration. At the same time, the system uses log analysis technology to collect and analyze the behavior data of miners and build a complete data set. These data are transmitted to the edge computing nodes deployed in the mine, and are processed and analyzed in real time by variational autoencoders and generative adversarial networks (GANs).

[0234] Through a density-based spatial clustering algorithm, the system builds detailed user portraits and analyzes the behavior patterns and preferences of miners. Combining these user behavior data with mining environment data, the system can better understand the miners' operating habits and the impact of environmental conditions on their behavior. Then, the system extracts knowledge from a large number of coal mine production management documents through natural language processing and information extraction technology to build a knowledge graph. Graph convolutional networks (GCNs) and improved Transformer models are used to reason about the knowledge graph to provide accurate answers and decision support.

[0235] Through historical data analysis, the system builds an environmental simulation model and uses a time series prediction algorithm that combines a mixed density network (MDN) and a long short-term memory network (LSTM) to make real-time predictions of the environmental status of the mining area. For example, the system can predict changes in temperature and humidity in the mining area within the next 24 hours, as well as possible fluctuations in toxic gas concentrations. Based on these prediction results, the system generates personalized natural language answers to help managers make decisions in advance.

[0236] In addition, the system continuously adjusts and optimizes the model parameters by collecting feedback data from miners. For example, when miners are not satisfied with the answers provided by the system, the system records this feedback and adjusts the model through an adaptive learning algorithm. The optimization strategy of the system is saved through the parameter storage module to ensure that the system remains efficient and intelligent during continuous improvement.

[0237] Table 1 Data related to the coal mine production management knowledge Q&A system

[0238]

[0239]

[0240] First of all, in terms of response speed, after the system was launched, the response time of the mining area management personnel decreased from an average of 15 minutes to 5 minutes. This significant improvement is attributed to the application of edge computing and advanced data processing algorithms, which greatly improve the efficiency of data processing and decision support.

[0241] In terms of the accuracy of environmental prediction, the prediction accuracy of the system for the environmental status of the mining area reached 92%. For example, the system successfully predicted a sudden increase in the concentration of toxic gases and issued a warning 4 hours in advance, avoiding a potential safety accident. Such a high accuracy provides strong support for the safety management of the mining area.

[0242] In terms of user satisfaction, according to the feedback data from miners, the satisfaction score for the answers and suggestions provided by the system has increased from 70 points at the initial stage of launch to 85 points currently. Through continuous adjustment and optimization of the model by the adaptive learning algorithm, the user experience has been significantly improved.

[0243] In terms of the generation time of the emergency plan, the system generated an emergency plan in only 10 seconds during a sudden mine fire, guiding the miners to evacuate quickly. Post-event statistics showed that there were no casualties in the fire accident and the economic loss was reduced by 30%. This shows that the system performs excellently in terms of emergency response speed and effectiveness.

[0244] In terms of production efficiency, after the system was launched, the overall production efficiency of the mining area increased by 15%. Through real-time monitoring of the equipment status and predictive maintenance, the equipment failure rate decreased by 20%. This preventive maintenance measure effectively reduces the equipment downtime and improves the production efficiency.

[0245] In terms of safety, after the system was launched, the incidence of coal mine safety accidents decreased by 40%. This achievement is mainly due to the efficient environmental monitoring of the system and the timely response of the emergency plan, providing a solid guarantee for the safety management of the mining area.

[0246] In summary, the coal mine production management knowledge Q&A system based on natural language processing has achieved remarkable results in improving response speed, enhancing the accuracy of environmental prediction, increasing user satisfaction, optimizing the effectiveness of emergency plans, improving production efficiency, and enhancing safety. These data fully demonstrate the practical value and innovation of the present invention, providing more efficient and safe support for coal mine production management.

[0247] As described above, only the preferred specific embodiments of the present invention are given, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A coal mine production management knowledge question-answering method based on natural language processing, characterized in that: The steps include: S1. Use sensor technology to collect environmental data from coal mine production sites, use log analysis technology to collect and analyze user behavior data, and build a data set; S2. Deploy edge computing nodes on-site at coal mines, use a data processing algorithm combining variational autoencoders and generative adversarial networks to perform edge analysis, and process data sets in real time; S3. Use density-based spatial clustering algorithms to build user profiles, analyze user behaviors and preferences through association rules, and combine environmental data from production sites with user profiles; S4. Build a knowledge graph, extract knowledge through natural language processing and information extraction technology, use graph convolutional networks and improved Transformer to build a cognitive model, and combine logistic regression and XGBoost algorithms for reasoning; S5. Build an environmental simulation model through historical data analysis, use edge computing and cloud computing to combine real-time data fusion, and use a time series prediction algorithm that combines a hybrid density network with a long short-term memory network to predict environmental status; S6, Generate personalized natural language answers based on user portraits, knowledge graphs, and environmental state predictions; S7. Collect user feedback data, analyze it through data analysis tools, and use adaptive learning algorithms to adjust model parameters and optimization strategies; S8. Use relational databases to store emergency plan templates, apply rule engines combined with knowledge graphs to generate rule-based emergency plans, automatically generate emergency plans based on environmental simulation and prediction results, and continuously optimize emergency plan templates through historical data analysis.

2. According to a coal mine production management knowledge question-answering method based on natural language processing according to claim 1, it is characterized in that: The S2 specifically includes: S21. Use edge computing nodes to perform preliminary data cleaning and preprocessing, including noise filtering, missing value filling, and data standardization, to form a preprocessed data set X; S22. Input the preprocessed data set X into the variational autoencoder model, extract features and reduce the dimension of the data through the variational autoencoder to generate latent variables Z: Z=μ+σ⊙∈; Among them, μ represents the mean, σ represents the variance, ∈~N(0,I), ⊙ represents element-by-element multiplication, and I represents the identity matrix; S23, inputting the hidden variable Z into a generative adversarial network, where the generative adversarial network includes a generator G and a discriminator D; The generator G generates new data samples G(Z) based on the hidden variable Z: G(Z)=tanh(W g ·Z+b g ); Among them, W g represents the weight of the generator, b g represents the bias of the generator; The discriminator D compares the generated data sample G(Z) with the real data set X, and uses the discriminator D to determine the authenticity of the sample: D(G(Z))=f(W d ·G(Z)+b d ); Among them, W d represents the weight of the discriminator, b d represents the bias of the discriminator, and f represents the activation function; S24. The variational autoencoder and the generative adversarial network are jointly trained through the back propagation algorithm to optimize the parameters of the generator and the discriminator. The specific optimization objective function is: Among them, p data (x) represents the probability density function of the real data distribution, x represents the data point sampled from the real data distribution, and p z (z) represents the probability distribution of latent variables in the latent space, and z represents the latent variables sampled from the latent space; S25. Process the real-time collected data through the trained variational autoencoder and generative adversarial network to generate high-quality processed data 3. The coal mine production management knowledge question and answer method based on natural language processing according to claim 2 is characterized in that: The S25 specifically includes: Among them, α represents the weight of the attention mechanism, and Attention(D,G(Z)) represents the weighted fusion of attention to the data sample G(Z) and the real data.

4. The coal mine production management knowledge question and answer method based on natural language processing according to claim 1 is characterized in that: The S3 specifically includes: S31, using density-based spatial clustering algorithm to generate the Cluster analysis of the data; Set two parameters: radius parameter ∈ and minimum sample parameter MinPts; For each data point p in the dataset i , calculate the number of neighborhood data points N(p i ): Among them, dist(p i ,p j ) represents the data point p i and data point p j The distance between The data points are divided into core points, boundary points and noise points, where the core points meet the number of neighborhood data points N(p i ) is greater than or equal to MinPts, the boundary point satisfies the condition that the number of neighboring data points is less than MinPts but is within the neighborhood of the core point, and the noise point does not meet the conditions of both the core point and the boundary point; Cluster the core points and the points in their neighborhood, classify the connected core points and the points directly reachable by density into the same category, and form a user behavior pattern cluster; S32, performing association rule analysis on the clustering results; Calculate frequent item sets and generate frequent item sets F through Apriori algorithm or FP-Growth algorithm. Frequent item sets F meet the support threshold Support(A): Among them, Count(A) represents the number of transactions containing item set A, and N represents the total number of transactions; Generate association rules in the form of A→B, where A and B represent elements in the frequent item set and satisfy the confidence threshold Confidence(A→B): To evaluate the effectiveness of association rules, lift (A→B) is used to measure the relevance of the rules: S33. Decompose the environmental data into time series into trend component, seasonal component and residual component, expressed as: y t =T t +S t +R t ; Among them, y t represents time series data, T t represents the trend component, S t represents the seasonal component, R t represents the residual component; S34, use the moving average method to smooth the trend components and identify the long-term trend. The moving average formula is T t for: Where n represents the size of the moving window; S35. Use fast Fourier transform to identify periodic changes in seasonal components. The Fourier transform formula is: Among them, X k represents the frequency domain signal, x t represents the time domain signal, and N represents the signal length; S36. Combine the user behavior pattern and the time series characteristics of the environmental data to generate a user profile, which includes the user's operating habits, behavioral preferences, and reaction information to different environmental conditions.

5. The coal mine production management knowledge question and answer method based on natural language processing according to claim 1 is characterized in that: The S4 specifically includes: S41. Processing coal mine production management documents by natural language processing technology, including word segmentation, part-of-speech tagging, and named entity recognition. The word segmentation and part-of-speech tagging use the SpaCy tool, and the named entity recognition uses the BERT model; S42. Extract knowledge from the processed documents, including entities, attributes and relationships, and use information extraction technology to add the extracted knowledge to the knowledge graph to form a complete knowledge graph; S43. Use graph convolutional networks to learn embedded representations of knowledge graphs and generate low-dimensional vector representations of each node. Let the graph be G, the node be v, and the neighbor node be N(v). Then the representation of node v is h v The calculation formula is: Among them, σ represents the activation function, W (l) represents the weight matrix of the lth layer, aggregate represents the aggregation function, represents the representation of node v at layer l, represents the representation of node u at layer l, represents the representation of node v at layer l+1; S44. Input the embedding representation result of the graph convolutional network into the improved Transformer model for further processing. The improved Transformer model includes an adaptive multi-head attention mechanism and dynamic position encoding. The adaptive multi-head attention score α ij The calculation formula is: Where a represents the adaptive attention weight vector, || represents the vector connection operation, W represents the weight matrix, and h i represents the feature representation of node i, h j represents the feature representation of node j, and PE(i,j) represents the dynamic position encoding of nodes i and j: Among them, i and j represent the index of the node, d represents the position dimension, and D represents the total embedding dimension; S45. Classify the nodes in the knowledge graph through the logistic regression algorithm and predict the category of unknown nodes. The logistic regression prediction formula is: Where x represents the input feature, w represents the weight, b represents the bias, and P(y=1|x) represents the probability that the node belongs to category 1; S46. Use the XGBoost algorithm to perform regression analysis on the nodes and relationships in the knowledge graph. XGBoost makes predictions by weighted averaging the results of multiple decision trees: Where K represents the number of trees, represents the predicted value of sample i, f k Represents the prediction result of the kth tree; S47. Combine the results of the logistic regression and XGBoost algorithms to generate the final inference result.

6. The coal mine production management knowledge question and answer method based on natural language processing according to claim 1 is characterized in that: The S5 specifically includes: S51, build a mixed density network; S52. Build a long short-term memory network. The network structure includes an input layer, an LSTM layer, and an output layer. The input layer receives the parameters output by the MDN. The LSTM layer is used to capture the long-term and short-term dependencies in the time series. The output layer generates prediction results. S53. Combine the mixed density network with the long short-term memory network to predict the environmental state. The prediction formula is: in, represents the predicted value at time point t, y t represents the parameter input generated by the mixed density network, h t-1 represents the hidden state at the previous time point, c t-1 indicates the cell state at the previous time point; S54. The prediction results are stored in the cloud computing platform and provided to the knowledge question answering system through an interface for query and use.

7. The coal mine production management knowledge question and answer method based on natural language processing according to claim 6 is characterized in that: The S51 specifically includes: Input layer, hidden layer and output layer. The input layer receives the fused comprehensive data set, the hidden layer is implemented by a multi-layer perceptron, and the output layer generates the parameters of the mixed density model: Among them, y t represents the predicted value at time point t, π j represents the weight of the j-th Gaussian distribution, N(μ j ,σ j ) represents the probability density function of Gaussian distribution, μ j represents the mean of the jth Gaussian distribution, σ j represents the standard value of the j-th Gaussian distribution.

8. The method for coal mine production management knowledge question and answer based on natural language processing according to claim 6 is characterized in that: The S52 specifically includes: i t =σ(W i ·[h t-1 ,x t ]+b i ); f t =σ(W f ·[h t-1 ,x t ]+b f ); the t =σ(W o ·[h t-1 ,x t ]+b o ); g t =tanh(W g ·[h t-1 ,x t ]+b g ); c t =f t ⊙c t-1 +i t ⊙g t ; h t =o t ⊙tanh(c t ); Among them, i t represents the activation vector of the input gate, σ represents the activation function, W i , W f , W o and W g represents the weight matrix input to each gate and candidate memory state, b i 、b f 、b o and b g represents the bias term input to each gate and candidate memory state, h t-1 represents the hidden state at the previous time point, x t represents the input data at the current time point, f t represents the activation vector of the forget gate, o t represents the activation vector of the output gate, g t represents the activation vector of the candidate memory state, c t represents the cell state at the current time point, c t-1 represents the cell state at the previous time point, ⊙ represents the element multiplication operation, h t Indicates the hidden state at the current time point.

9. The coal mine production management knowledge question and answer method based on natural language processing according to claim 1 is characterized in that: The S7 specifically includes: S71. In the adaptive learning module, construct a reward function R; S72, based on the optimization result of the reward function R, using an adaptive learning algorithm to update the model parameters of the system; S73. Regularly apply the updated model parameters to the system operation, and monitor and collect a new round of user feedback data to form a feedback closed loop; S74. Save the adjusted model parameters and optimization strategies into a parameter storage module.

10. The coal mine production management knowledge question and answer method based on natural language processing according to claim 9, characterized in that: The S71 specifically includes: Among them, S represents the user satisfaction score, T represents the system response time, F represents the number of new questions raised by users, U represents the number of opinions and suggestions provided by users, and α, β, γ, and δ represent weight coefficients.

Images