AI-based mine safety monitoring system
By applying AI technology in mine safety monitoring systems, including deep belief networks, autoencoders and Bayesian networks, the shortcomings of data processing and risk prediction in traditional systems are solved, and more accurate risk identification and timely emergency response are achieved. Responsiveness improves the efficiency and accuracy of mine safety management.
Patent Information
- Application Number
- CN202410194158.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-21
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-02-21
AI Technical Summary
Traditional mine safety monitoring systems rely on simple statistical methods and manual judgments in data processing and risk prediction, resulting in misjudgment and untimely response, lack of efficient automation tools, making it difficult to accurately identify potential risks, affecting the accuracy and response efficiency of security management. .
Adopt AI-based mine safety monitoring system, including geological monitoring module, mine abnormality analysis module, risk trend prediction module, emergency warning system module, safety risk mapping module, emergency management strategy module, operation area safety monitoring module and mine risk management module , using deep belief networks, autoencoders, recurrent neural networks, genetic algorithms, Bayesian networks and variational autoencoders to carry out feature extraction, anomaly detection, risk prediction and emergency response strategy formulation.
It has improved the accuracy of geological stability assessment, timely identification of abnormal operating models, accuracy of risk prediction, response speed of emergency warning and visualization of safety risks, and enhanced the scientific nature of emergency management and the effectiveness of risk control.
Smart Images

Figure CN118037047B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of mine safety monitoring, and in particular to an AI-based mine safety monitoring system. Background Art
[0002] The field of mine safety monitoring technology is a field that focuses on using advanced technologies to ensure the safety of the mine working environment, reduce accidents, and improve the health and safety of miners. This technology field integrates a variety of technologies such as automation, machine vision, data analysis, and machine learning, aiming to achieve real-time monitoring of the mine environment, the construction of early warning systems, early identification of accident risks, and the formulation of preventive measures. Through continuous observation and data collection of the internal environment of the mine, these systems can help mine managers better understand the working environment and formulate more effective safety measures.
[0003] Among them, the AI-based mine safety monitoring system is a monitoring system that integrates artificial intelligence technology and aims to improve the safety and efficiency of mine operations. Its main purpose is to predict and prevent safety accidents through real-time monitoring and data analysis, protect the lives of miners, and reduce economic losses. The system uses the powerful data processing and pattern recognition capabilities of artificial intelligence to conduct in-depth analysis of complex mining environments, thereby achieving early warning of potential dangers and improving the intelligence level of mine safety management.
[0004] Traditional mine safety monitoring systems rely on simple statistical methods and manual judgment in data processing and risk prediction, which not only limits the depth and breadth of data analysis, but also leads to misjudgment due to human factors. In addition, traditional systems lack efficient automated tools for anomaly detection and risk assessment, making it difficult to identify potential risks in a timely and accurate manner, increasing the probability of accidents. In terms of emergency warning, traditional methods often fail to effectively prevent or reduce accidents due to insufficient timely or accurate responses. The mapping and visualization of safety risks are also relatively rough in traditional systems, insufficient to provide clear risk awareness and decision-making support. In addition, the formulation of emergency management strategies often relies on experience and lacks scientific analysis and evaluation, which affects the effectiveness of emergency response. These shortcomings have led to inaccurate risk management, slow response, and unscientific decision-making in actual operations, increasing the risk of safety accidents and limiting the efficiency and safety of mine operations. Summary of the invention
[0005] The purpose of the present invention is to solve the shortcomings of the prior art and propose an AI-based mine safety monitoring system.
[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions: the AI-based mine safety monitoring system includes a geological monitoring module, a mine anomaly analysis module, a risk trend prediction module, an emergency warning system module, a safety risk mapping module, an emergency management strategy module, an operating area safety monitoring module, and a mine risk management and control module;
[0007] The geological monitoring module uses a deep belief network to extract features based on geological data, and uses a cluster analysis method to classify features, identify multiple geological stability levels, and generate geological stability analysis results;
[0008] The mine anomaly analysis module uses an autoencoder to feature encode and reconstruct mine operation data based on the geological stability analysis results, performs trend analysis and anomaly detection on the encoded data through time series analysis, and generates an abnormal operation mode;
[0009] The risk trend prediction module uses a recurrent neural network to perform time series modeling on abnormal data based on abnormal operation modes, estimates future risk development trends through predictive analysis technology, and generates risk development prediction results;
[0010] The emergency warning system module uses a genetic algorithm to optimize the parameters of the warning mechanism based on the risk development prediction results, and uses a rule engine to set warning rules according to the optimized parameters to generate a warning signal system;
[0011] The safety risk mapping module uses a Bayesian network to construct a correlation diagram between risk factors based on the early warning signal system, and uses spatial analysis technology to map the correlation in the geographic information system to generate a safety risk map;
[0012] The emergency management strategy module is based on the safety risk map, uses a causal reasoning model to analyze the causal relationship between risk factors, combines the strategy planning method to formulate an emergency response strategy, and generates an emergency response plan;
[0013] The work area safety monitoring module is based on the emergency response plan, tracks the safety status of the work area in real time through a dynamic graph network, and uses real-time monitoring technology to analyze data and generate real-time safety monitoring data;
[0014] The mine risk management and control module is based on real-time safety monitoring data, uses a variational autoencoder to extract features and detect anomalies in risk data, and uses a risk assessment method to analyze and evaluate the detection results to generate a risk management and control strategy.
[0015] As a further scheme of the present invention, the geological stability analysis results include rock formation stability rating, fault activity risk index, and hydrogeological hazardous area identification results; the abnormal operation mode includes equipment operation abnormality indicators, production process deviation values, and energy consumption abnormal trends; the risk development prediction results include risk level assessment, risk development speed estimation, and potential impact area prediction; the early warning signal system includes risk level classification standards, early warning trigger conditions, and emergency response guidance measures; the safety risk map includes key risk source location, risk area definition, and risk intensity distribution map; the emergency response plan includes emergency evacuation route planning, accident handling flowchart, and key emergency resource allocation; the real-time safety monitoring data includes real-time location distribution of employees, key equipment status monitoring, and environmental safety index; the risk management strategy includes a list of risk reduction measures, implementation of safety regulations, and dynamic update of risk levels.
[0016] As a further solution of the present invention, the geological monitoring module includes a geological data analysis submodule, a geological risk identification submodule, and a geological stability assessment submodule;
[0017] The geological data analysis submodule is based on geological data, adopts a deep belief network, uses TensorFlow and Keras libraries to build a network, initializes the model Sequential(), adds multiple fully connected layers Dense() for feature extraction, sets the number of nodes in each layer, including 100 nodes in the first layer and 50 nodes in the second layer, uses the ReLU activation function, sets the optimizer to Adam, selects mean_squared_error as the loss function, uses the model.compile() method to compile the model, applies the model.fit() method to train the model, sets the training batch batch_size to 32, sets the number of iterations epochs to 10, and generates feature extraction data;
[0018] The geological risk identification submodule extracts data based on features, adopts cluster analysis method, uses Scikit-learn library to execute K-means algorithm, initializes KMeans class, sets the number of clusters n_clusters to 5, selects k-means++ as init parameter, sets the maximum number of iterations max_iter to 300, calls model.fit() method to cluster data, uses labels_ attribute to obtain cluster labels, and generates geological stability grade classification;
[0019] The geological stability assessment submodule performs stability assessment based on geological stability grade classification, uses weighted scoring method, defines weights of differentiated risk factors, including fault activity weight 0.3, soil liquefaction potential weight 0.2, writes scoring function calculate_score() for each assessment indicator, applies weighted formula to calculate total score, grades geological stability according to total score, and generates geological stability analysis results.
[0020] As a further solution of the present invention, the mine abnormality analysis module includes an operation data analysis submodule, an abnormal pattern recognition submodule, and an abnormal index extraction submodule;
[0021] The operation data analysis submodule uses an autoencoder to encode data based on the geological stability analysis results, including using TensorFlow and Keras libraries in a Python environment, initializing a Sequential model, sequentially adding encoding layers Dense, setting the number of nodes to 64, using a ReLU activation function, adding a decoding layer Dense, setting the number of nodes to the number of original data features, restoring data, selecting Adam as the optimizer, setting the loss function to mean_squared_error, training the model through the model.fit() method, setting batch_size to 32, and setting epochs to 50, and generating encoded reconstructed data;
[0022] The abnormal pattern recognition submodule is based on the encoded reconstructed data, uses time series analysis to recognize abnormal patterns, uses Pandas for data preprocessing, creates time series decomposition using the Statsmodels library, tests data stability using the ADF test, and uses the SARIMA model to predict abnormal trends. The model parameters are set to p=2, d=1, and q=2, and the model.fit() method is used for model training. The forecast() method is used to predict future trends and generate trend analysis results.
[0023] The abnormal indicator extraction submodule performs anomaly detection based on the trend analysis results, defines the anomaly detection threshold as 3 times the standard deviation of the data point, uses the Numpy library to calculate the standard deviation of the data point, identifies the data points that exceed the threshold, and generates an abnormal operation mode.
[0024] As a further solution of the present invention, the risk trend prediction module includes a risk model construction submodule, a risk prediction analysis submodule, and a trend judgment submodule;
[0025] The risk model construction submodule is based on the abnormal operation mode and adopts the recurrent neural network algorithm. The model is constructed using the tf.keras.Sequential method in the TensorFlow library. The LSTM layer is added through tf.keras.layers.LSTM, wherein 128 neurons are set in each layer, and the activation function is set to sigmoid in the gate control through the activation=sigmoid parameter. The output layer is added through tf.keras.layers.Dense, and the activation function is set to tanh through the activation=tanh parameter. The model is compiled through the model.compile() method, the optimizer is set to Adam through the optimizer=adam parameter, and the loss function is set to the mean square error through the loss=mean_squared_error parameter. The model is trained through the model.fit() method, and the time step is set to 10 time points of data to generate time series modeling results;
[0026] The risk prediction analysis submodule continues to apply the RNN model for risk prediction based on the time series modeling results. The model input configuration predicts the risk level at a future time point through the model.predict() method, based on the data of the past 10 time points as input. The learning rate is set by the tf.keras.optimizers.Adam(learning_rate=0.001) parameter of the Adam optimizer during model compilation. The batch size and number of iterations are set by the batch_size=32 and epochs=200 parameters of the model.fit() method during model training to generate risk prediction results.
[0027] The trend judgment submodule is based on the risk prediction results, adopts the time series analysis method, uses the numpy and pandas libraries to process data, uses numpy.diff to calculate the changes in risk levels at consecutive time points, uses the pandas.Series.rolling() method to analyze trends, and uses the scipy.stats.linregress() method to perform linear regression analysis. It analyzes the trend of predicted values changing over time and generates risk development prediction results.
[0028] As a further solution of the present invention, the emergency warning system module includes a warning mechanism design submodule, a genetic algorithm optimization submodule, and a warning signal release submodule;
[0029] The early warning mechanism design submodule performs the initial design of the early warning mechanism based on the risk development prediction results. The Python script is used to define the set_warning_levels function, which accepts the risk value as an input parameter. The risk is classified into three levels: low, medium, and high according to the preset thresholds of 0.2, 0.5, and 0.8 through the if-elif-else structure. The level is returned as a string, including if risk<=0.2: return Low. The level is used as the basic input for genetic algorithm optimization to generate the early warning parameter design result.
[0030] The genetic algorithm optimization submodule is based on the warning parameter design result, uses Python's DEAP library to define individuals and fitness types through creator.create, uses base.Toolbox() to create a genetic algorithm toolbox, initializes the population size to 50, encodes individual genes as warning parameter thresholds through tools.initRepeat, registers individual generation functions through toolbox.register() method, sets the crossover probability to 0.7, registers crossover operations through toolbox.register() method, sets the mutation probability to 0.01, registers mutation operations through toolbox.register() method, uses tools.selRoulette() as the selection function for roulette selection, and executes 100 iterations through algorithms.eaSimple() method to generate optimized warning parameters;
[0031] The warning signal issuing submodule builds a rule engine based on the optimized warning parameters using the PyKnow library, defines the rule trigger_warning, and uses the KnowledgeEngine.rule decorator to compare the real-time risk assessment data with the medium risk threshold, including @Rule(AS.fact<<Risk(risk_level=Medium,risk_value=P(lambda x:x> threshold))), when the condition is met, the execution part triggers the corresponding warning signal, and the action defined by def then(self, fact) is used to generate the warning signal system.
[0032] As a further solution of the present invention, the security risk mapping module includes a risk network construction submodule, a risk distribution analysis submodule, and a risk level mapping submodule;
[0033] The risk network construction submodule is based on the early warning signal system, uses the Bayesian network to construct the association graph between risk factors, uses Python's pgmpy library, and defines the Bayesian model structure through the BayesianModel() function, including BayesianModel([(geological stability, meteorological conditions), (meteorological conditions, personnel distribution)]), where each tuple represents a directed edge between nodes, and uses the TabularCPD() function to define the conditional probability distribution for each node, including TabularCPD(variable=geological stability, variable_card=2, values=[[0.1, 0.9]]), to generate a risk association network;
[0034] The risk distribution analysis submodule is based on the risk association network and uses the ArcPy library for spatial analysis. The arcpy.management.AddLayer() function is first used to import the geographic location data of risk factors into GIS, and then the arcpy.sa.HotSpotAnalysis() function is applied to perform hot spot analysis. The neighborhood size of the analysis is defined as 3 map units by setting parameters including neighborhood=NbrCircle(3, "MAP") to generate risk distribution analysis results.
[0035] The risk level mapping submodule uses mapping tools in ArcGIS to visualize the risk level based on the risk distribution analysis results, creates a layer through the arcpy.mapping.Layer() function, and applies the arcpy.mapping.UpdateLayer() function to set the symbolization of differentiated risk levels to generate a safety risk map.
[0036] As a further solution of the present invention, the emergency management strategy module includes an emergency strategy planning submodule, a causal relationship analysis submodule, and a management measure formulation submodule;
[0037] The emergency strategy planning submodule plans emergency strategies based on the security risk map, selects response strategies using a decision tree algorithm, uses the DecisionTreeClassifier class in the scikit-learn library, takes the risk level in the risk map as input features, trains a decision tree model based on historical response measures data corresponding to each risk level, sets the model parameter max_depth=3 to limit the maximum depth of the tree, and uses the model.fit() method to train the model to generate emergency strategy planning results;
[0038] The causal relationship analysis submodule uses a Bayesian network to perform causal relationship analysis based on the emergency strategy planning results, creates a Bayesian model through the BayesianModel class of the pgmpy library, defines nodes and edges through BayesianModel([(weather conditions, geological stability), (personnel distribution, emergency response)]) to represent risk factors and causal relationships, and uses the TabularCPD class to define the conditional probability distribution of each node, including TabularCPD(variable=weather conditions, variable_card=3, values=[[0.3, 0.4, 0.3]]) to set the probability distribution of the meteorological condition node, and generate causal relationship analysis results;
[0039] The management measures formulation submodule formulates emergency management measures based on the causal analysis results and combines strategic planning methods to formulate emergency flow charts for multiple risk situations. At the same time, it uses Visio to draw emergency response flow charts and selects key emergency measures under multiple risk levels, including immediate initiation of emergency evacuation in risk situations, enhanced monitoring in medium risk situations, and generation of emergency response plans.
[0040] As a further solution of the present invention, the work area safety monitoring module includes a real-time monitoring submodule, a dynamic safety analysis submodule, and a safety status update submodule;
[0041] The real-time monitoring submodule is based on the emergency response plan, uses the OpenCV library to create a computer vision algorithm, executes the command cv2.absdiff() to perform differential motion detection on continuous video frames, adjusts the threshold parameter threshold=25 to determine motion sensitivity, uses cv2.findContours() to detect motion areas, and combines the Internet of Things technology to collect sensor data through the command read(), uses numpy.mean() and numpy.std() to calculate the average and standard deviation of the temperature, pressure, and smoke sensor readings to generate real-time monitoring data;
[0042] The dynamic security analysis submodule uses the scikit-learn library to create a support vector machine algorithm based on real-time monitoring data, executes the SVC() function to create a model, sets the kernel function parameter kernel=rbf, the penalty parameter C=1.0, the kernel function coefficient gamma=scale, executes the model.fit() method to train the model, and applies cross_val_score() to perform a five-fold cross validation to generate a security risk assessment;
[0043] The security status update submodule is based on security risk assessment, uses the Pandas library for data analysis, executes pandas.read_csv() to read assessment data, sets security thresholds, filters risk events through DataFrame.query(), adjusts warning levels, uses DataFrame.to_csv() to update emergency measures records, publishes current security information, and generates real-time security monitoring data.
[0044] As a further solution of the present invention, the mine risk management and control module includes a risk assessment submodule, a risk detection and identification submodule, and a management and control strategy implementation submodule;
[0045] The risk assessment submodule is based on real-time security monitoring data and adopts a variational autoencoder, in which the encoder part uses the command encoder.add(Dense(units=256, activation=relu)) to add a dense connection layer, setting the number of units to 256 and the activation function to ReLU. The decoder part uses the same configuration, and the loss function is set to vae_loss=xent_loss+kl_loss, where xent_loss is the reconstruction loss and kl_loss is the KL divergence. Use model.fit() to train the model, extract data features, and generate a feature data set;
[0046] The risk detection and identification submodule applies the local outlier factor algorithm based on the feature data set, executes LocalOutlierFactor(n_neighbors=20, contamination=0.1), sets the neighborhood size to 20, the contamination parameter to 0.1, uses the fit_predict() method to fit the feature data and predict outliers, identifies outlier data points, and generates a potential risk list;
[0047] The control strategy implementation submodule classifies risks and assesses potential impacts based on the potential risk list, applies DataFrame.groupby() to group risk categories, uses DataFrame.apply() to evaluate each group of risks, and formulates risk control measures.
[0048] Compared with the prior art, the advantages and positive effects of the present invention are:
[0049] In the present invention, deep belief networks and cluster analysis are used to effectively extract and classify geological data, thereby improving the accuracy of geological stability assessment. Autoencoders and time series analysis achieve more sophisticated operation data feature encoding and anomaly detection in mine anomaly analysis, and promote the timely identification of abnormal operation modes. In terms of risk trend prediction, recurrent neural networks enhance the ability to predict future risk development trends by modeling the time series of abnormal data. The application of genetic algorithms and rule engines optimizes the emergency warning system and improves the response speed and accuracy of the warning mechanism. Bayesian networks and spatial analysis techniques jointly construct a correlation diagram between risk factors in safety risk mapping, and effectively map them in geographic information systems, making the visualization of safety risks more intuitive and easier to understand. The combination of causal reasoning models and strategic planning methods provides a scientific basis for formulating emergency response strategies in the emergency management strategy module, and enhances the efficiency of responding to emergencies. The use of dynamic graph networks and real-time monitoring technology achieves accurate tracking and analysis of real-time data in safety monitoring of the operating area, and updates safety conditions in a timely manner. In the mine risk management and control module, the variational autoencoder and risk assessment method provide strong technical support for feature extraction and anomaly detection of risk data, ensuring the accuracy of risk assessment and the effectiveness of management and control strategies. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] Figure 1 is a system flow chart of the present invention;
[0051] Figure 2 It is a schematic diagram of the system framework of the present invention;
[0052] Figure 3 It is a flow chart of the geological monitoring module of the present invention;
[0053] Figure 4 This is a flow chart of the mine anomaly analysis module of the present invention;
[0054] Figure 5 This is a flow chart of the risk trend prediction module of the present invention;
[0055] Figure 6 It is a module flow chart of the emergency warning system of the present invention;
[0056] Figure 7 It is a flow chart of the security risk mapping module of the present invention;
[0057] Figure 8 It is a flow chart of the emergency management strategy module of the present invention;
[0058] Fig. 9 It is a flow chart of the working area safety monitoring module of the present invention;
[0059] Fig.10This is a flow chart of the mine risk management and control module of the present invention. DETAILED DESCRIPTION
[0060] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0061] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.
[0062] Embodiment 1
[0063] See also Figure 1 to Figure 2 , the AI-based mine safety monitoring system includes geological monitoring module, mine anomaly analysis module, risk trend prediction module, emergency warning system module, safety risk mapping module, emergency management strategy module, operation area safety monitoring module, and mine risk management and control module;
[0064] The geological monitoring module uses a deep belief network to extract features based on geological data, and uses cluster analysis methods to classify features, identify multiple geological stability levels, and generate geological stability analysis results;
[0065] The mine anomaly analysis module uses an autoencoder to feature encode and reconstruct mine operation data based on the results of geological stability analysis, and performs trend analysis and anomaly detection on the encoded data through time series analysis to generate abnormal operation patterns;
[0066] The risk trend prediction module is based on the abnormal operation mode, applies recurrent neural network to model the time series of abnormal data, estimates the future risk development trend through predictive analysis technology, and generates risk development prediction results;
[0067] The emergency warning system module uses genetic algorithms to optimize the parameters of the warning mechanism based on the risk development prediction results, and uses the rule engine to set warning rules according to the optimized parameters to generate a warning signal system;
[0068] The safety risk mapping module is based on the early warning signal system and uses the Bayesian network to construct a correlation diagram between risk factors. It also uses spatial analysis technology to map the correlation in the geographic information system to generate a safety risk map.
[0069] The emergency management strategy module is based on the safety risk map, uses a causal reasoning model to analyze the causal relationship between risk factors, and combines strategic planning methods to formulate emergency response strategies and generate emergency response plans;
[0070] The work area safety monitoring module is based on the emergency response plan, and tracks the safety status of the work area in real time through a dynamic graph network. It also uses real-time monitoring technology to analyze data and generate real-time safety monitoring data.
[0071] The mine risk management and control module is based on real-time safety monitoring data. It uses variational autoencoders to extract features and detect anomalies in risk data. It also uses risk assessment methods to analyze and evaluate the detection results to generate risk management strategies.
[0072] The results of geological stability analysis include rock stability rating, fault activity risk index, and hydrogeological hazardous area identification results. Abnormal operation modes include abnormal equipment operation indicators, production process deviation values, and abnormal energy consumption trends. Risk development prediction results include risk level assessment, risk development speed estimation, and potential impact area prediction. The early warning signal system includes risk level classification standards, early warning trigger conditions, and emergency response guidance measures. The safety risk map includes the location of key risk sources, the definition of risk areas, and the risk intensity distribution map. The emergency response plan includes emergency evacuation route planning, accident handling flowchart, and key emergency resource allocation. Real-time safety monitoring data includes real-time location distribution of employees, key equipment status monitoring, and environmental safety index. Risk management strategies include a list of risk reduction measures, the implementation of safety regulations, and dynamic updates of risk levels.
[0073] In the geological monitoring module, the feature extraction of geological data is performed through the deep belief network (DBN). The collected data format includes information such as stratigraphic structure, rock type, groundwater level, etc., which are converted into standardized numerical inputs. DBN is established by stacking multiple layers of restricted Boltzmann machines (RBMs), and each layer of RBM is responsible for extracting different levels of data features. During the training process, the RBM is first trained layer by layer without supervision, and then supervised fine-tuning is performed to ensure the accuracy of feature extraction. The extracted features use cluster analysis, such as the K-means algorithm, to classify geological data according to feature similarity, identify different levels of geological stability, and generate detailed geological stability analysis results, providing a basis for subsequent mine anomaly analysis.
[0074] In the mine anomaly analysis module, autoencoders are used to encode and reconstruct the features of mine operation data, including mine equipment operating parameters, environmental monitoring indicators, etc. The autoencoder consists of an input layer, a hidden layer, and an output layer. The hidden layer efficiently encodes the input data, and the output layer reconstructs the original input. In this way, the autoencoder captures the key features in the data and reduces the interference of noise. Time series analysis such as the ARIMA model is applied to the encoded data to analyze the time dependency and trend of the data. Anomaly detection is achieved by identifying mutations in the data sequence or behaviors that do not conform to historical patterns, and generating detailed reports of abnormal operating patterns. These reports help identify potential operational problems or risks.
[0075] In the risk trend prediction module, a recurrent neural network (RNN) is applied to model the time series of abnormal operation patterns. RNN is particularly suitable for processing sequence data because it can consider the temporal dynamic characteristics of data. Through the long short-term memory (LSTM) network, a variant of RNN, the module is able to learn and memorize the long-term dependencies of abnormal data, avoiding the gradient vanishing problem of traditional RNN. Predictive analysis techniques, such as regression analysis, are applied to the output of RNN to estimate the development trend of future risks and generate risk development forecast results. These results are crucial for formulating preventive measures and emergency response strategies.
[0076] In the emergency warning system module, genetic algorithms are used to optimize the parameters of the warning mechanism. Genetic algorithms simulate the natural selection process and find the optimal solution through iteration. In this module, the algorithm evaluates the effect of various warning parameter combinations, such as warning thresholds and response times, and selects the best combination based on the fitness function. The rule engine sets the warning rules based on the optimized parameters. The rule engine analyzes the matching of real-time data with preset rules and generates warning signals. These signal systems effectively guide emergency response and reduce potential risks.
[0077] In the safety risk mapping module, a Bayesian network is used to construct a correlation map between risk factors, and then mapped in a geographic information system (GIS) through spatial analysis technology. The Bayesian network represents the relationship between different risk factors in a probabilistic form, providing a method to quantify and visualize risk associations. Spatial analysis techniques such as buffer analysis and hot spot analysis are implemented in GIS, combining risk correlation maps with actual geographic locations to generate intuitive safety risk maps. These maps make risk management more clear and targeted.
[0078] In the emergency management strategy module, causal reasoning models and strategic planning methods are combined. Causal reasoning models analyze the causal relationships between risk factors identified in the security risk map to reveal potential risk drivers. Strategic planning methods, such as decision tree analysis, develop emergency response strategies based on these causal relationships. This approach comprehensively considers various risk scenarios and response measures to generate a comprehensive and practical emergency response plan.
[0079] In the work area safety monitoring module, the safety status of the work area is tracked in real time through the dynamic graph network. The dynamic graph network can process and analyze network data that changes over time and effectively track the safety dynamics of the work area. Real-time monitoring technologies such as video monitoring and sensor data analysis collect safety-related data of the work area in real time. Through real-time analysis of these data, real-time safety monitoring reports are generated. These reports provide real-time safety status of the work area, which is crucial for rapid response and decision-making.
[0080] In the mine risk management module, a variational autoencoder (VAE) is used to extract features and detect anomalies in risk data. VAE converts input data into potential representations through the encoder, and then reconstructs the data through the decoder. In this process, VAE learns the distribution characteristics of the data and effectively identifies abnormal data. Risk assessment methods such as risk matrix analysis evaluate the anomalies detected by VAE to determine the risk level and impact. In this way, the generated risk management strategy provides in-depth insights and effective management of potential risks in mines.
[0081] See also Figure 3 ,The geological monitoring module includes a geological data analysis submodule, a geological risk identification submodule, and a geological stability assessment submodule;
[0082] The geological data analysis submodule is based on geological data and adopts a deep belief network. The network is built using the TensorFlow and Keras libraries. The model Sequential() is initialized, and multiple fully connected layers Dense() are added for feature extraction. The number of nodes in each layer is set to 100 nodes in the first layer and 50 nodes in the second layer. The ReLU activation function is used, and the optimizer is set to Adam. The loss function is selected as mean_squared_error. The model is compiled using the model.compile() method, and the model is trained using the model.fit() method. The training batch batch_size is set to 32, and the number of iterations epochs is set to 10 to generate feature extraction data.
[0083] The geological risk identification submodule extracts data based on features, adopts cluster analysis method, uses Scikit-learn library to execute K-means algorithm, initializes KMeans class, sets the number of clusters n_clusters to 5, selects k-means++ as init parameter, sets the maximum number of iterations max_iter to 300, calls model.fit() method to cluster data, uses labels_ attribute to obtain cluster labels, and generates geological stability grade classification;
[0084] The geological stability assessment submodule performs stability assessment based on the geological stability grade classification. It uses a weighted scoring method to define the weights of differentiated risk factors, including a fault activity weight of 0.3 and a soil liquefaction potential weight of 0.2. It writes a scoring function calculate_score() for each evaluation indicator, applies a weighted formula to calculate the total score, and classifies the geological stability according to the total score to generate a geological stability analysis result.
[0085] In the geological data analysis submodule, the geological data is feature extracted by using a deep belief network (DBN) combined with TensorFlow and Keras libraries. The geological data includes multi-dimensional information such as stratigraphic structure, rock type, and groundwater level, which are formatted as numerical data input. When using TensorFlow and Keras to build a DBN model, first initialize the Sequential() model, which is a linear stacked model that allows network layers to be added layer by layer. Add multiple fully connected layers Dense() to the model, each of which is responsible for capturing different levels of data features. The first layer is set with 100 nodes as the input layer to receive the raw data, and the second layer is set with 50 nodes to further refine the features. The ReLU activation function is used in these layers to help the model learn nonlinear relationships. Adam is set as the optimizer, which has the ability to adjust the learning rate adaptively, making training more efficient. The loss function selects mean_squared_error to evaluate the error between the model output and the true value. Use the model.compile() method to compile the model and integrate the set layers and configurations. Next, the model.fit() method is used to train the model, and the batch_size is set to 32 and the number of epochs is set to 10 to ensure that the data can be effectively learned. Through this process, the model can extract key features from the original geological data and generate feature extraction data, which provides the basis for subsequent geological risk identification and assessment.
[0086] In the geological risk identification submodule, cluster analysis is performed to identify the geological stability level based on the feature extraction data obtained from the geological data analysis submodule. The K-means algorithm, a widely used clustering method, is implemented using the Scikit-learn library. When initializing the KMeans class, the number of clusters n_clusters is set to 5, which means that the data is divided into five stability levels. The init parameter selects k-means++, which can improve the clustering quality. The maximum number of iterations max_iter is set to 300 to ensure that the algorithm has enough iterations to converge. Then, the model.fit() method is called to cluster the feature extraction data. The algorithm assigns the data points to the nearest cluster by calculating the distance between the data points and the cluster center. After clustering is completed, the labels_ attribute is used to obtain the cluster labels of each data point, which represent different geological stability levels. In this way, the submodule is able to group geological data into different stability levels and generate a classification of geological stability levels, which provides an important basis for geological stability assessment.
[0087] In the geological stability assessment submodule, stability assessment is performed, which is based on the results of the previous geological risk identification submodule. A weighted scoring method is used to assign different weights to different geological risk factors. For example, the weight of fault activity is 0.3, and the weight of soil liquefaction potential is 0.2. These weights reflect the degree of influence of each risk factor on geological stability. A scoring function calculate_score() is written for each evaluation indicator to calculate the score based on the actual data. A weighted formula is used to combine all scores to calculate the total score. Based on the total score, the geological stability is graded, such as assigning areas with high total scores to higher stability levels. This evaluation method can quantify geological stability risks and generate geological stability analysis results. These results are crucial for guiding mining plans and safety monitoring in mines, and provide scientific data support for decision makers.
[0088] Suppose there is a batch of geological data, including rock hardness, soil moisture and other indicators. The simulated value range of hardness is 5-9, and the simulated value range of moisture is 10-40%. In the geological data analysis submodule, these data are processed by the deep belief network to extract features related to geological stability, such as the feature extraction result of a data point is [7.5, 25%]. Then, in the geological risk identification submodule, these feature data are divided into 5 stability levels by the K-means algorithm, and it is assumed that the data point is classified into the third level. Finally, in the geological stability assessment submodule, the geological stability score of the area is calculated by combining factors such as fault activity and soil liquefaction potential, such as a score of 75. These processes together constitute the key part of geological data analysis and evaluation in the mine safety monitoring system, providing in-depth analysis and real-time monitoring capabilities based on data, and providing strong technical support for mine safety management.
[0089] See also Figure 4 ,The mine abnormality analysis module includes the operation data analysis submodule, the abnormal pattern recognition submodule, and the abnormal indicator extraction submodule;
[0090] The operational data analysis submodule uses an autoencoder to encode data based on the geological stability analysis results, including using the TensorFlow and Keras libraries in the Python environment, initializing the Sequential model, sequentially adding the encoding layer Dense, setting the number of nodes to 64, using the ReLU activation function, adding the decoding layer Dense, setting the number of nodes to the number of original data features, restoring the data, selecting Adam as the optimizer, setting the loss function to mean_squared_error, training the model through the model.fit() method, setting batch_size to 32, and epochs to 50, and generating encoded reconstructed data;
[0091] The abnormal pattern recognition submodule reconstructs data based on coding, uses time series analysis for abnormal pattern recognition, uses Pandas for data preprocessing, creates time series decomposition using the Statsmodels library, tests data stability using the ADF test, and uses the SARIMA model to predict abnormal trends. The model parameters are set to p=2, d=1, and q=2, and the model.fit() method is used for model training. The forecast() method is used to predict future trends and generate trend analysis results.
[0092] The abnormal indicator extraction submodule performs anomaly detection based on the trend analysis results, defines the anomaly detection threshold as 3 times the standard deviation of the data point, uses the Numpy library to calculate the standard deviation of the data point, identifies the data points that exceed the threshold, and generates an abnormal operation mode.
[0093] In the operational data analysis submodule, an autoencoder is used to encode data based on the results of geological stability analysis. This process involves the use of TensorFlow and Keras libraries in the Python environment. The results of geological stability analysis include various parameters such as rock hardness and soil moisture, which are presented in numerical format. First, in the Python environment, the Sequential model is initialized using TensorFlow and Keras libraries. The Sequential model is a linear model that can be constructed by sequentially adding network layers. In the model, first add an encoding layer Dense, which is a fully connected layer with the number of nodes set to 64, in order to compress the original data into a lower-dimensional representation. The ReLU activation function is used in this layer, which helps capture nonlinear features and improve the expressiveness of the model. Next, add a decoding layer Dense with the number of nodes set to the same as the number of original data features in order to restore to the original data space. During the training of the autoencoder, the Adam optimizer is selected, and its adaptive learning rate feature helps improve training efficiency and stability. The loss function is set to mean_squared_error, which is used to calculate the error between the reconstructed data and the original data. The model is trained using the fit method, with batch_size set to 32 and epochs set to 50 to ensure that the model can fully learn the data features. Through these steps, the autoencoder can effectively extract the key features of the operational data, reconstruct the data during the decoding process, and generate encoded reconstructed data. These data provide the basis for subsequent abnormal pattern recognition.
[0094] In the abnormal pattern recognition submodule, time series analysis is used to identify abnormal patterns based on the encoded reconstructed data. First, Pandas is used for data preprocessing to convert the encoded reconstructed data into a format suitable for time series analysis. Next, the Statsmodels library is used to create a time series decomposition model, which can decompose the time series into components such as trend, seasonality, and residuals, which helps to better understand the structure of the data. The ADF test is performed on the data to test its stationarity, which is an important step in time series analysis because many time series models require the data to be stationary. According to the test results, the SARIMA model is used to model the data and predict abnormal trends. The SARIMA model is a model widely used in time series forecasting. The model parameters are set as p=2, d=1, and q=2. These parameters control different aspects of the model, such as seasonal differences, autoregressive terms, and moving average terms. The model is trained using the fit method, and future trends are predicted using the forecast method. Through this process, the model is able to identify abnormal patterns in the data and generate trend analysis results. These results are crucial to understanding abnormal situations in the operation process and help to promptly discover and deal with potential operational problems.
[0095] In the abnormal indicator extraction submodule, anomaly detection is performed, which is based on the results of the previous trend analysis. The threshold for anomaly detection is defined as 3 times the standard deviation of the data point, which is a commonly used anomaly detection standard that can effectively identify extreme data points. The standard deviation of the data points is calculated using the Numpy library, which is a commonly used numerical calculation library in Python that can efficiently process large amounts of data. After calculating the standard deviation, data points that exceed the threshold are identified and are considered abnormal. In this way, abnormal operating patterns can be identified, such as equipment failures, operational errors, etc. The generated abnormal operating patterns are very important for operational management. They provide key information that helps to adjust operational strategies in a timely manner and prevent potential risks.
[0096] Assume that there is a batch of operational data, including indicators such as equipment operating time and energy consumption. The simulated value range of operating time is 8-12 hours, and the simulated value range of energy consumption is 1000-1500 kWh. In the operational data analysis submodule, these data are encoded and reconstructed through the autoencoder. For example, the encoding and reconstruction result of a set of data points is [10 hours, 1200 kWh]. In the abnormal pattern recognition submodule, through time series analysis, it is found that the energy consumption on a certain day suddenly increased to 1800 kWh. This anomaly is captured by the SARIMA model and predicted as an abnormal trend. In the abnormal indicator extraction submodule, this data point is identified as an abnormality based on the standard deviation calculation. This series of processes shows how to extract key information from operational data, identify abnormal patterns, and warn of potential operational risks, thereby playing an important role in the mine safety monitoring system.
[0097] See also Figure 5 ,The risk trend prediction module includes a risk model building submodule, a risk prediction and analysis submodule, and a trend judgment submodule;
[0098] The risk model construction submodule is based on the abnormal operation mode and adopts the recurrent neural network algorithm. The model is constructed using the tf.keras.Sequential method in the TensorFlow library. The LSTM layer is added through tf.keras.layers.LSTM, where 128 neurons are set in each layer, and the activation function is set to sigmoid in the gate control through the activation=sigmoid parameter. The output layer is added through tf.keras.layers.Dense, and the activation function is set to tanh through the activation=tanh parameter. The model is compiled through the model.compile() method, the optimizer is set to Adam through the optimizer=adam parameter, and the loss function is set to the mean square error through the loss=mean_squared_error parameter. The model is trained through the model.fit() method, and the time step is set to 10 time points to generate time series modeling results;
[0099] The risk prediction analysis submodule continues to apply the RNN model for risk prediction based on the time series modeling results. The model input configuration predicts the risk level at a future time point through the model.predict() method, based on the data of the past 10 time points as input. The learning rate is set through the tf.keras.optimizers.Adam(learning_rate=0.001) parameter of the Adam optimizer during model compilation. The batch size and number of iterations are set through the batch_size=32 and epochs=200 parameters of the model.fit() method during model training to generate risk prediction results.
[0100] The trend judgment submodule is based on the risk prediction results and adopts the time series analysis method. It processes data through the numpy and pandas libraries, uses numpy.diff to calculate the changes in risk levels at consecutive time points, uses the pandas.Series.rolling() method to analyze trends, and uses the scipy.stats.linregress() method to perform linear regression analysis. It analyzes the trend of predicted values over time and generates risk development prediction results.
[0101] In the risk model construction submodule, based on the abnormal operation mode, the recurrent neural network (RNN) algorithm, especially the long short-term memory network (LSTM), is used to build the time series model. This process is carried out in the TensorFlow library. First, the model is initialized using the tf.keras.Sequential method. Sequential is a linear model that allows different layers to be added in sequence. In the model, several LSTM layers are added through tf.keras.layers.LSTM. Each LSTM layer is set with 128 neurons, which can capture long-term dependencies in time series data. In the LSTM layer, the activation function is set to sigmoid through the activation=sigmoid parameter. This gate control mechanism helps the network remember and forget information. The output layer of the model is added through tf.keras.layers.Dense, and the activation function is set to tanh, which is achieved through the activation=tanh parameter. This setting helps the model output prediction values between -1 and 1, which is suitable for the characteristics of time series data. In the model compilation stage, through the model.compile method, the optimizer is set to Adam, which is called through the optimizer=adam parameter. It is an efficient gradient descent optimization algorithm. The loss function is set to mean squared error (mean_squared_error), which is used to evaluate the gap between the model's predicted value and the actual value. Model training is performed through the model.fit method, setting the time step to 10 time points, which means that the model will use the data of the past 10 time points to predict the value of the next time point. Through these steps, the LSTM model can learn the dynamic characteristics of the time series from abnormal operating patterns and generate time series modeling results, which are crucial for understanding and predicting the development of risks.
[0102] In the risk prediction analysis submodule, the RNN model is continued to be applied for risk prediction. This process is based on the LSTM model built previously. The model input configuration is based on the data of the past 10 time points to predict the risk level at a future time point. When the model is compiled, the learning rate is set by the tf.keras.optimizers.Adam(learning_rate=0.001) parameter of the Adam optimizer. This learning rate parameter has a significant impact on the speed and effect of model training. The batch size and number of iterations are set by the batch_size=32 and epochs=200 parameters of the model.fit method during the model training phase to ensure that the model has enough data and number of iterations to learn the dynamic changes of risks. Through the model.predict method, the model can predict the risk level at a future time point. These prediction results are crucial for identifying and responding to potential safety risks in advance, and help mine managers take preventive measures in a timely manner.
[0103] In the trend judgment submodule, based on the risk prediction results, time series analysis is performed to determine the development trend of the risk. First, the numpy and pandas libraries are used to process the data. The numpy library is used to perform mathematical operations, such as using numpy.diff to calculate the changes in risk levels at consecutive time points. The pandas library is used for data processing and analysis, and the pandas.Series.rolling method is used to analyze the trend of risk levels over time. Linear regression analysis is performed using the scipy.stats.linregress method, which is able to analyze the trend of predicted values over time. Through these methods, the change pattern of risk levels over time can be identified to determine whether the risk is growing, decreasing, or remaining stable. This analysis is crucial to understanding the long-term development trend of risks and helping managers make more informed decisions. The generated risk development forecast results provide valuable insights for mine management and help optimize safety strategies and emergency plans.
[0104] Assume that there is a series of data about mining equipment failures and environmental anomalies. These data are presented in the form of time series, including equipment temperature, vibration intensity, etc. The simulated values of temperature range from 20°C to 50°C, and the simulated values of vibration intensity range from 0.5 to 5.0. In the risk model construction submodule, these data are analyzed by the LSTM model. For example, the model predicts that the equipment temperature will rise to 55°C in the next hour, exceeding the safety threshold. In the risk prediction analysis submodule, the model continues to predict the continuation of this trend in the next few hours, indicating the risk of equipment overheating. Finally, in the trend judgment submodule, time series analysis and linear regression are used to determine that the trend of equipment temperature increase is intensifying. This series of processes shows that the monitoring system can effectively predict and analyze potential risks in mining operations and provide early warnings so that timely measures can be taken to reduce risks.
[0105] See also Figure 6 ,The emergency warning system module includes the warning mechanism design submodule, the genetic algorithm optimization submodule, and the warning signal release submodule;
[0106] The early warning mechanism design submodule performs the initial design of the early warning mechanism based on the risk development prediction results. It uses Python scripts to define the set_warning_levels function, accepts the risk value as an input parameter, and uses the if-elif-else structure to classify the risk into three levels: low, medium, and high according to the preset thresholds of 0.2, 0.5, and 0.8. The level is returned as a string, including if risk<=0.2:return Low. The level is used as the basic input for genetic algorithm optimization to generate the early warning parameter design results.
[0107] The genetic algorithm optimization submodule is based on the warning parameter design results. It uses the Python DEAP library to define individuals and fitness types through creator.create, uses base.Toolbox() to create a genetic algorithm toolbox, initializes the population size to 50, encodes individual genes as warning parameter thresholds through tools.initRepeat, registers the individual generation function through the toolbox.register() method, sets the crossover probability to 0.7, registers the crossover operation through the toolbox.register() method, sets the mutation probability to 0.01, registers the mutation operation through the toolbox.register() method, uses tools.selRoulette() as the selection function for roulette selection, and executes 100 iterations through the algorithms.eaSimple() method to generate optimized warning parameters.
[0108] The warning signal release submodule uses the PyKnow library to build a rule engine based on the optimized warning parameters, defines the rule trigger_warning, and uses the KnowledgeEngine.rule decorator to compare the real-time risk assessment data with the medium risk threshold, including @Rule(AS.fact<<Risk(risk_level=Medium,risk_value=P(lambda x:x> threshold))), when the condition is met, the execution part triggers the corresponding warning signal, and the action defined by def then(self, fact) is used to generate the warning signal system.
[0109] In the early warning mechanism design submodule, Python language is used to implement the early warning level setting based on risk assessment. This process uses conditional judgment statements to divide the risk into three levels: low, medium and high according to the input risk value. The implementation process first defines a function set_warning_levels, which accepts the risk value as input. The risk value is a floating point number that represents the size of the assessed risk. The if-elif-else structure is used inside the function to set three thresholds: 0.2, 0.5, and 0.8. When the input risk value is less than or equal to 0.2, the function returns the string 'Low', indicating a low risk level; when the risk value is greater than 0.2 and less than or equal to 0.5, it returns 'Medium', indicating a medium risk; and when the risk value exceeds 0.8, it returns 'High', indicating a high risk level. This process ensures that the quantitative assessment of risk can be converted into a clear early warning level, providing a basic input for subsequent genetic algorithm optimization. The advantage of this method is its simplicity and intuitiveness, and it can quickly assign appropriate early warning levels according to risk values. In addition, since this design allows flexible adjustment of thresholds, it has high applicability in different application scenarios.
[0110] In the genetic algorithm optimization submodule, Python's DEAP library is used to implement the genetic algorithm for optimizing the warning parameters. The core of this module is to define the individual and fitness types, as well as the various operations of the genetic algorithm. First, the individual type and fitness type are created through the creator.create function in the DEAP library. The individual type contains the warning parameter threshold as its genetic code. Then, a genetic algorithm toolbox is created using base.Toolbox(), and the population size is initialized to 50 individuals through tools.initRepeat. The genetic code of each individual, i.e., the warning parameter threshold, is generated by the individual generation function registered by the toolbox.register method. Then, the crossover probability is set to 0.7 and the mutation probability is set to 0.01, and the crossover and mutation operations are registered by the toolbox.register method respectively. The selection function uses tools.selRoulette to implement the roulette selection mechanism. Finally, 100 iterations are performed through the algorithms.eaSimple method to optimize the warning parameters. The key to this process is to iteratively evolve warning parameters that are more adaptable to the environment by simulating natural selection and genetic mechanisms. The advantage of this approach is that it can find efficient solutions in complex parameter spaces and due to its iterative nature, the solutions can be continuously optimized.
[0111] In the warning signal release submodule, the PyKnow library is used to build a rule engine to process the optimized warning parameters. The core of this module is to define the rule trigger_warning, which is used to trigger the warning signal under certain conditions. The condition part of the rule is to compare the real-time risk assessment data with the medium risk threshold. In the specific implementation, the KnowledgeEngine.rule decorator and the Rule class are used. For example, @Rule(AS.fact<<Risk(risk_level=Medium,risk_value=P(lambdax:x> threshold))) This rule indicates that when there is a fact with a medium risk level and a risk value greater than the preset threshold, the rule is triggered. After the rule is triggered, the corresponding warning signal is issued through the action defined in defthen(self, fact). The key point of this process is to be able to dynamically issue warning signals based on real-time risk assessment data, ensuring the timeliness and accuracy of the warning system. The advantage of using a rule engine is that it can handle complex logical relationships and is easy to adjust and maintain rules.
[0112] In an AI-based mine safety monitoring system, these three submodules are used to process and respond to safety risks in mines. For example, real-time monitoring data shows that the gas concentration in a specific area is 0.18%, while the set low-risk threshold is 0.2%. In the early warning mechanism design submodule, the input gas concentration data is evaluated as low risk. Subsequently, in the genetic algorithm optimization submodule, through iterative optimization, the most suitable early warning parameter settings for the current mine conditions are found. Finally, in the early warning signal release submodule, when the real-time data shows that the gas concentration exceeds the optimized medium-risk threshold, the rule engine will trigger the corresponding early warning signal and promptly notify the management personnel to take measures, thereby effectively improving the safety level of the mine.
[0113] See also Figure 7 ,The security risk mapping module includes a risk network construction submodule, a risk ,distribution analysis submodule, and a risk level mapping submodule;
[0114] The risk network construction submodule is based on the early warning signal system and uses the Bayesian network to construct the association diagram between risk factors. The Python pgmpy library is used to define the Bayesian model structure through the BayesianModel() function, including BayesianModel([(geological stability, meteorological conditions), (meteorological conditions, personnel distribution)]), where each tuple represents a directed edge between nodes. At the same time, the TabularCPD() function is used to define the conditional probability distribution for each node, including TabularCPD(variable=geological stability, variable_card=2, values=[[0.1, 0.9]]), to generate a risk association network;
[0115] The risk distribution analysis submodule uses the ArcPy library for spatial analysis based on the risk association network. First, the geographic location data of risk factors are imported into GIS using the arcpy.management.AddLayer() function. Then, the arcpy.sa.HotSpotAnalysis() function is used for hot spot analysis. The neighborhood size of the analysis is defined as 3 map units by setting parameters including neighborhood=NbrCircle(3, "MAP") to generate the risk distribution analysis results.
[0116] Based on the results of risk distribution analysis, the risk level mapping submodule uses mapping tools in ArcGIS to visualize the risk level, creates a layer through the arcpy.mapping.Layer() function, and applies the arcpy.mapping.UpdateLayer() function to set the symbolization of differentiated risk levels to generate a safety risk map.
[0117] In the risk network construction submodule, the association graph between risk factors is constructed through Bayesian network. This process uses Python's pgmpy library, and Bayesian network is used to express the conditional dependency between risk factors. First, the BayesianModel() function is used to define the model structure, such as BayesianModel([('geological stability', 'meteorological conditions'), ('meteorological conditions', 'personnel distribution')]), where each tuple represents a directed edge, indicating the direct impact of one factor on another factor. Then, the TabularCPD() function is used to define the conditional probability distribution for each node. For example, TabularCPD(variable='geological stability', variable_card=2, values=[[0.1, 0.9]]) defines the probability distribution of the "geological stability" node, where variable_card=2 indicates that this node has two states (stable and unstable), and values=[[0.1, 0.9]] indicates the probability of each state of the node given the parent node state. In this way, the Bayesian network can capture the relationships and interactions between risk factors, providing a quantitative framework for risk assessment and decision-making. The constructed Bayesian network can represent complex causal relationships and provide a deep understanding of the interactions between risk factors through probabilistic reasoning.
[0118] In the risk distribution analysis submodule, spatial analysis was performed using the ArcPy library to reveal the distribution characteristics of risk factors in geographic space. First, the data containing the geographical locations of risk factors were imported into the Geographic Information System (GIS) using the arcpy.management.AddLayer() function. These data are identified by geographic coordinates or specific map locations. Then, the arcpy.sa.HotSpotAnalysis() function was applied to perform hot spot analysis, which defined the neighborhood size of the analysis as 3 map units by setting the parameter neighborhood=NbrCircle(3, "MAP") at each location. This process uses statistical methods to analyze the density of risk factors at each location and identify hot spots, that is, those areas where the risk level is significantly higher than the average level. This analysis helps to identify areas where risks are concentrated and provides a basis for risk prevention and resource allocation. The generated risk distribution analysis results can be presented in the form of graphics or charts, making the geographical distribution of risks clear at a glance.
[0119] In the risk level mapping submodule, the risk levels are visualized using the mapping tools in ArcGIS. This process first creates a map layer using the arcpy.mapping.Layer() function, and then sets the symbolic representation of different risk levels using the arcpy.mapping.UpdateLayer() function. The purpose of this is to visually display the security risk levels of different regions on the map. Different colors, icons, or patterns are used to represent different risk levels, such as red for high-risk areas and green for low-risk areas. In this way, the risk map not only provides a visual representation of the risk distribution, but also helps decision makers and relevant personnel quickly identify and respond to high-risk areas. The generated security risk map can serve as an important part of the early warning system to assist real-time monitoring and emergency response.
[0120] In an AI-based mine safety monitoring system, these three submodules are integrated and applied. Assume that there is geological stability data showing that the stability index of a part of the area is 0.7 (representing relative instability), meteorological condition data shows that the wind speed is 15 kilometers per hour, and personnel distribution data shows that there are 20 workers in the area. In the risk network construction submodule, the constructed Bayesian network can reveal the probabilistic relationship between these factors, such as high wind speed increases the danger of geologically unstable areas. In the risk distribution analysis submodule, spatial analysis identifies areas with higher risks, such as areas with unstable geology and dense personnel. Finally, in the risk level mapping submodule, this information is converted into a risk map, in which high-risk areas are marked in red, intuitively indicating areas that need special attention. This comprehensive analysis and visualization method provides scientific decision-making support for mine safety management and enhances the awareness and response capabilities of potential risks.
[0121] See also Figure 8 ,The emergency management strategy module includes the emergency strategy planning submodule, the causal relationship analysis submodule, and the management measures formulation submodule;
[0122] The emergency strategy planning submodule plans emergency strategies based on the security risk map, uses the decision tree algorithm to select response strategies, uses the DecisionTreeClassifier class in the scikit-learn library, takes the risk level in the risk map as the input feature, and trains the decision tree model based on the historical response measures data corresponding to each risk level. The model parameter max_depth=3 is set to limit the maximum depth of the tree, and the model.fit() method is used to train the model to generate the emergency strategy planning results;
[0123] The causal relationship analysis submodule uses the Bayesian network to perform causal relationship analysis based on the emergency strategy planning results. The Bayesian model is created through the BayesianModel class of the pgmpy library. The nodes and edges are defined through BayesianModel([(meteorological conditions, geological stability), (personnel distribution, emergency response)]) to represent risk factors and causal relationships. The conditional probability distribution of each node is defined using the TabularCPD class, including TabularCPD(variable=meteorological conditions, variable_card=3, values=[[0.3, 0.4, 0.3]]) to set the probability distribution of the meteorological condition node and generate the causal relationship analysis results.
[0124] The management measures formulation submodule formulates emergency management measures based on the results of causal analysis and combines strategic planning methods to develop emergency flow charts for multiple risk situations. At the same time, it uses Visio to draw emergency response flow charts and selects key emergency measures under multiple risk levels, including immediate initiation of emergency evacuation in risk situations, enhanced monitoring in medium-risk situations, and generation of emergency response plans.
[0125] In the emergency strategy planning submodule, the decision tree algorithm is used to select and plan the response strategy. The DecisionTreeClassifier class in the scikit-learn library is used to train the decision tree model based on the risk level in the security risk map as the input feature. In this process, historical response data is first collected, including the emergency measures taken for each risk level. These data are input into the decision tree model as a training set. When using DecisionTreeClassifier, the model parameter max_depth=3 is set. This parameter limits the maximum depth of the tree to prevent the model from overfitting and maintain the interpretability of the results. Then, the decision tree model is trained using historical data through the fit method. In this process, the model learns the response measures under different risk levels and forms a set of decision rules. The generated emergency strategy planning results are a series of emergency response measures based on risk levels, which can provide guidance for real-time emergency response. The advantage of this method is that it can provide a clear decision path and corresponding strategy selection, making emergency response more systematic and efficient.
[0126] In the causal relationship analysis submodule, causal relationship analysis was performed through Bayesian network. A Bayesian model was created using the BayesianModel class of the pgmpy library, which defines the causal relationship between risk factors and emergency response. For example, the model structure is defined by BayesianModel([('weather conditions', 'geological stability'), ('personnel distribution', 'emergency response')]), which shows how meteorological conditions affect geological stability and how personnel distribution affects emergency response decisions. The conditional probability distribution of each node is defined by the TabularCPD class, for example, TabularCPD(variable='weather conditions', variable_card=3, values=[[0.3, 0.4, 0.3]]) sets the probability distribution of the meteorological condition node, where variable_card=3 indicates that meteorological conditions have three states. Such analysis can reveal the mutual influence and dependency between different factors, and provide a scientific basis for the formulation of risk management and emergency strategies. The advantage of Bayesian network is that it can quantify uncertainty and conditional dependence, providing an intuitive and effective analysis tool for causal relationships in complex decision-making environments.
[0127] In the management measures formulation submodule, emergency management measures were formulated by combining the strategic planning method and the results of causal analysis. By using Visio to draw an emergency response flowchart, an emergency process for multi-risk situations was formulated. This process includes the selection of key emergency measures under different risk levels, such as immediate initiation of emergency evacuation in high-risk situations and enhanced monitoring in medium-risk situations. The advantage of this approach is that it provides a clear visual representation that helps managers quickly understand and implement emergency plans. The generated emergency response plan not only guides the implementation of specific emergency measures, but also enhances the ability to respond to emergencies and improves the overall safety management efficiency.
[0128] In the AI-based mine safety monitoring system, the emergency strategy planning submodule uses historical data, which contains the risk levels and corresponding emergency measures of multiple earthquake events. For example, the data shows that in the case of minor earthquakes (low risk level), the measures taken in history are to strengthen monitoring; in the case of moderate earthquakes (medium risk level), the measures taken are local evacuation. The decision tree model trained with these data provides clear response strategies for earthquake events of different levels. In the causal relationship analysis submodule, the impact of factors such as meteorological conditions, geological stability and personnel distribution on emergency response is analyzed. For example, the analysis shows that in strong wind weather, geological instability increases, and more stringent safety measures need to be taken. Finally, in the management measures formulation submodule, these analysis results are combined to formulate a complete set of emergency response processes, including specific emergency measures under different meteorological conditions and geological stability conditions, such as immediate evacuation in the case of strong winds and geological instability. Such a system provides a comprehensive safety management solution to ensure the safe operation of mines under various risk conditions.
[0129] See also Fig. 9 ,The operation area safety monitoring module includes a real-time monitoring submodule, a dynamic safety analysis submodule, and a safety status update submodule;
[0130] The real-time monitoring submodule is based on the emergency response plan and uses the OpenCV library to create a computer vision algorithm. It executes the command cv2.absdiff() to detect motion by differential detection of continuous video frames, adjusts the threshold parameter threshold=25 to determine motion sensitivity, and uses cv2.findContours() to detect motion areas. At the same time, it combines the Internet of Things technology to collect sensor data through the command read(), and uses numpy.mean() and numpy.std() to calculate the average and standard deviation of the temperature, pressure, and smoke sensor readings to generate real-time monitoring data.
[0131] The dynamic security analysis submodule uses the scikit-learn library to create a support vector machine algorithm based on real-time monitoring data, executes the SVC() function to create a model, sets the kernel function parameter kernel=rbf, the penalty parameter C=1.0, and the kernel function coefficient gamma=scale, executes the model.fit() method to train the model, and applies cross_val_score() to perform a five-fold cross validation to generate a security risk assessment.
[0132] The security status update submodule is based on security risk assessment, uses the Pandas library for data analysis, executes pandas.read_csv() to read assessment data, sets security thresholds, filters risk events through DataFrame.query(), adjusts warning levels, uses DataFrame.to_csv() to update emergency measures records, publishes current security information, and generates real-time security monitoring data.
[0133] In the real-time monitoring submodule, computer vision algorithms are implemented through the OpenCV library to monitor and analyze video frames. First, cv2.absdiff() is used to differentiate consecutive video frames to detect motion. This method identifies moving objects by comparing the pixel differences between two adjacent frames. Areas with differences greater than a set threshold (for example, threshold=25) are considered to be moving areas. This threshold determines the sensitivity of motion detection. A lower threshold makes the system more sensitive to small movements. Next, cv2.findContours() is used to detect the contours of these moving areas and identify the position and shape of moving objects. At the same time, combined with the Internet of Things technology, environmental data such as temperature, pressure, and smoke are collected through sensors. For these sensor data, the mean() and std() functions of numpy are used to calculate the mean and standard deviation to evaluate the stability and change trend of the environmental state. The advantage of this method is that it can monitor and analyze environmental changes in real time, detect abnormal situations in a timely manner, and provide important data support for emergency response.
[0134] In the dynamic security analysis submodule, the scikit-learn library is used to create a support vector machine (SVM) algorithm for security risk assessment. First, an SVM model is created using the SVC() function, and the kernel function parameter is set to kernel=rbf. This radial basis function (RBF) kernel is an effective method for processing nonlinear data. The penalty parameter C=1.0 is used to control the model's tolerance for misclassification, and gamma=scale automatically adjusts the coefficient of the kernel function. Then, the model.fit() method is used to train the model. The training data comes from the data collected by the real-time monitoring submodule, including video monitoring analysis results and sensor readings. Next, cross_val_score() is applied to perform a five-fold cross validation to evaluate the performance and accuracy of the model. The advantage of this method is that it can effectively handle high-dimensional data and ensure the generalization ability of the model through cross-validation, thereby providing accurate security risk assessment.
[0135] In the security status update submodule, the Pandas library is used for data analysis and information update. First, execute pandas.read_csv() to read the data of the previous security risk assessment. Then, according to the set security threshold, use DataFrame.query() to filter out high-risk events. Adjust the warning level according to these risk events to respond to the current security situation. Finally, use DataFrame.to_csv() to update the record of emergency measures and publish the current security information. This process can ensure the real-time update of security information and the timely adjustment of emergency measures, improving the response speed and efficiency of security management.
[0136] In an AI-based mine safety monitoring system, the real-time monitoring submodule uses OpenCV to analyze the video captured by the surveillance camera. If abnormal movement is detected in a specific area (such as equipment movement or unauthorized entry), the system will immediately identify and mark it. At the same time, sensor data shows that the temperature in the area suddenly rises to 50 degrees Celsius and the smoke sensor readings are abnormal. In the dynamic safety analysis submodule, the SVM model analyzes this data, identifies high-risk conditions, and ensures the accuracy of the analysis through five-fold cross-validation. Then, in the safety status update submodule, these analysis results are recorded and updated to the safety information system, emergency warnings are issued, and emergency measures are adjusted, such as starting the automatic fire extinguishing system and evacuation instructions. Such implementation ensures that the mine safety monitoring system can respond to potential safety threats in real time, take timely countermeasures, and protect the safety of workers and the integrity of equipment.
[0137] See also Fig.10 ,The mine risk management and control module includes the risk assessment submodule, the risk detection and identification submodule, and the management ,strategy implementation submodule;
[0138] The risk assessment submodule is based on real-time security monitoring data and uses a variational autoencoder. The encoder part uses the command encoder.add(Dense(units=256, activation=relu)) to add a dense connection layer, setting the number of units to 256 and the activation function to ReLU. The decoder part uses the same configuration and the loss function is set to vae_loss=xent_loss+kl_loss, where xent_loss is the reconstruction loss and kl_loss is the KL divergence. Use model.fit() to train the model, extract data features, and generate a feature data set.
[0139] The risk detection and identification submodule applies the local outlier factor algorithm based on the feature data set, executes LocalOutlierFactor(n_neighbors=20, contamination=0.1), sets the neighborhood size to 20, the contamination parameter to 0.1, uses the fit_predict() method to fit the feature data and predict outliers, identifies outlier data points, and generates a potential risk list;
[0140] The control strategy implementation submodule classifies risks and assesses potential impacts based on the potential risk list. It uses DataFrame.groupby() to group risk categories, uses DataFrame.apply() to evaluate each group of risks, and formulates risk control measures.
[0141] In the risk assessment submodule, a variational autoencoder (VAE) is used to extract features from real-time security monitoring data. A variational autoencoder is a generative model that can learn the latent feature representation of data. In the implementation of this submodule, a neural network is first constructed as an encoder. The network uses encoder.add(Dense(units=256, activation=relu)) to add a densely connected layer with 256 units, and the activation function is ReLU. The purpose is to map the input data (such as the multi-dimensional features of the monitoring data) to a latent representation space. Next, the decoder part is constructed, which has a similar configuration to the encoder and is used to remap the data in the latent space back to the original data space. During the model training process, the loss function is set to vae_loss=xent_loss+kl_loss, where xent_loss is the reconstruction loss (measures the difference between the reconstructed data and the original data), and kl_loss is the Kullback-Leibler divergence (measures the difference between the latent representation and the prior distribution). The model is trained through the model.fit() method to extract the latent features of the data and generate a feature dataset. The key to this process is to be able to extract useful feature representations from complex monitoring data, providing a basis for subsequent risk detection.
[0142] In the risk detection and identification submodule, the Local Outlier Factor (LOF) algorithm is applied to identify abnormal data points. The Local Outlier Factor algorithm is an unsupervised anomaly detection method that identifies outliers by calculating the local density deviation between a data point and its neighbors. In the implementation of this submodule, the neighborhood size is set to 20 and the contamination parameter is set to 0.1 using LocalOutlierFactor(n_neighbors=20, contamination=0.1), which means that 10% of the data points are expected to be outliers. Then, the feature data extracted from the VAE is fitted and outliers are predicted by the fit_predict() method, thereby identifying potential risk points and generating a potential risk list. The advantage of this method is that it can effectively identify data points that behave abnormally in the feature space, i.e., risk events, thus providing key information for risk management.
[0143] In the control strategy implementation submodule, risk levels and potential impact assessments were performed based on the potential risk list. In this process, Pandas' data processing function was used to group risk categories using DataFrame.groupby(), and then each group of risks was evaluated using the DataFrame.apply() method to develop corresponding risk control measures. This approach makes it possible to take targeted management measures for different types of risks, improving the effectiveness and pertinence of risk response. For example, emergency response measures can be developed for events identified as high-risk, while more conventional monitoring and prevention measures can be taken for low-risk events. This refined risk management strategy ensures the effective allocation of resources and minimizes the impact of potential risks.
[0144] In an AI-based mine safety monitoring system, the risk assessment submodule processes multiple sensor data including video surveillance, temperature, pressure, etc. through a variational autoencoder. For example, the feature data extracted from the monitoring video includes the object's moving speed, shape change, etc., and the temperature and pressure sensor data include the time series of real-time readings. After training, the VAE model is able to capture the key features of these data. In the risk detection and identification submodule, the LOF algorithm identifies several abnormal data points, such as a sudden increase in temperature in a certain area, or abnormal movement in the monitoring video. In the control strategy implementation submodule, these abnormal events are classified according to their potential danger level, and corresponding emergency response measures are formulated, such as increasing the inspection frequency of the area or starting the automatic fire extinguishing system. Such implementation ensures that the mine safety monitoring system can not only monitor the mine status in real time, but also detect and respond to potential risk events in a timely manner, effectively ensuring the safe operation of the mine.
[0145] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.
Claims
1. AI-based mine safety monitoring system, characterized by: The system includes a geological monitoring module, a mine anomaly analysis module, a risk trend prediction module, an emergency warning system module, a safety risk mapping module, an emergency management strategy module, an operating area safety monitoring module, and a mine risk management and control module; The geological monitoring module uses a deep belief network to extract features based on geological data, and uses a cluster analysis method to classify features, identify multiple geological stability levels, and generate geological stability analysis results; The mine anomaly analysis module uses an autoencoder to feature encode and reconstruct mine operation data based on the geological stability analysis results, performs trend analysis and anomaly detection on the encoded data through time series analysis, and generates an abnormal operation mode; The risk trend prediction module uses a recurrent neural network to perform time series modeling on abnormal data based on abnormal operation modes, estimates future risk development trends through predictive analysis technology, and generates risk development prediction results; The emergency warning system module uses a genetic algorithm to optimize the parameters of the warning mechanism based on the risk development prediction results, and uses a rule engine to set warning rules according to the optimized parameters to generate a warning signal system; The safety risk mapping module uses a Bayesian network to construct a correlation diagram between risk factors based on the early warning signal system, and uses spatial analysis technology to map the correlation in the geographic information system to generate a safety risk map; The emergency management strategy module is based on the safety risk map, uses a causal reasoning model to analyze the causal relationship between risk factors, combines the strategy planning method to formulate an emergency response strategy, and generates an emergency response plan; The work area safety monitoring module is based on the emergency response plan, tracks the safety status of the work area in real time through a dynamic graph network, and uses real-time monitoring technology to analyze data and generate real-time safety monitoring data; The mine risk management module uses a variational autoencoder to extract features and detect anomalies from risk data based on real-time safety monitoring data, and uses a risk assessment method to analyze and evaluate the detection results to generate a risk management strategy; The geological stability analysis results include rock formation stability rating, fault activity risk index, and hydrogeological hazardous area identification results; the abnormal operation mode includes equipment operation abnormality indicators, production process deviation values, and energy consumption abnormal trends; the risk development prediction results include risk level assessment, risk development speed estimation, and potential impact area prediction; the early warning signal system includes risk level classification standards, early warning trigger conditions, and emergency response guidance measures; the safety risk map includes key risk source location, risk area definition, and risk intensity distribution map; the emergency response plan includes emergency evacuation route planning, accident handling flowchart, and key emergency resource allocation; the real-time safety monitoring data includes real-time employee location distribution, key equipment status monitoring, and environmental safety index; the risk management strategy includes a risk reduction measures list, safety specification implementation, and risk level dynamic update; The security risk mapping module includes a risk network construction submodule, a risk distribution analysis submodule, and a risk level mapping submodule; The risk network construction submodule is based on the early warning signal system, uses the Bayesian network to construct the association graph between risk factors, uses Python's pgmpy library, and defines the Bayesian model structure through the BayesianModel() function, including BayesianModel([(geological stability, meteorological conditions), (meteorological conditions, personnel distribution)]), where each tuple represents a directed edge between nodes, and uses the TabularCPD() function to define the conditional probability distribution for each node, including TabularCPD(variable=geological stability, variable_card=2, values=[[0.1, 0.9]]), to generate a risk association network; The risk distribution analysis submodule is based on the risk association network and uses the ArcPy library for spatial analysis. The arcpy.management.AddLayer() function is first used to import the geographic location data of risk factors into GIS, and then the arcpy.sa.HotSpotAnalysis() function is applied to perform hot spot analysis. The neighborhood size of the analysis is defined as 3 map units by setting parameters including neighborhood=NbrCircle(3, "MAP") to generate risk distribution analysis results. The risk level mapping submodule uses mapping tools in ArcGIS to visualize the risk level based on the risk distribution analysis results, creates a layer through the arcpy.mapping.Layer() function, and uses the arcpy.mapping.UpdateLayer() function to set the symbolization of the differentiated risk level to generate a safety risk map; The emergency management strategy module includes an emergency strategy planning submodule, a causal relationship analysis submodule, and a management measures formulation submodule; The emergency strategy planning submodule plans emergency strategies based on the security risk map, selects response strategies using a decision tree algorithm, uses the DecisionTreeClassifier class in the scikit-learn library, takes the risk level in the risk map as input features, trains a decision tree model based on historical response measures data corresponding to each risk level, sets the model parameter max_depth=3 to limit the maximum depth of the tree, and uses the model.fit() method to train the model to generate emergency strategy planning results; The causal relationship analysis submodule uses a Bayesian network to perform causal relationship analysis based on the emergency strategy planning results, creates a Bayesian model through the BayesianModel class of the pgmpy library, defines nodes and edges through BayesianModel([(weather conditions, geological stability), (personnel distribution, emergency response)]) to represent risk factors and causal relationships, and uses the TabularCPD class to define the conditional probability distribution of each node, including TabularCPD(variable=weather conditions, variable_card=3, values=[[0.3, 0.4, 0.3]]) to set the probability distribution of the meteorological condition node, and generate causal relationship analysis results; The management measures formulation submodule formulates emergency management measures based on the causal relationship analysis results and combines strategic planning methods to formulate emergency flow charts for multiple risk situations, and uses Visio to draw emergency response flow charts, selects key emergency measures under multiple risk levels, including immediate initiation of emergency evacuation under risk situations, strengthening monitoring under medium risk situations, and generating emergency response plans; The work area safety monitoring module includes a real-time monitoring submodule, a dynamic safety analysis submodule, and a safety status update submodule; The real-time monitoring submodule is based on the emergency response plan, uses the OpenCV library to create a computer vision algorithm, executes the command cv2.absdiff() to perform differential motion detection on continuous video frames, adjusts the threshold parameter threshold=25 to determine motion sensitivity, uses cv2.findContours() to detect motion areas, and combines the Internet of Things technology to collect sensor data through the command read(), uses numpy.mean() and numpy.std() to calculate the average and standard deviation of the temperature, pressure, and smoke sensor readings to generate real-time monitoring data; The dynamic security analysis submodule uses the scikit-learn library to create a support vector machine algorithm based on real-time monitoring data, executes the SVC() function to create a model, sets the kernel function parameter kernel=rbf, the penalty parameter C=1.0, the kernel function coefficient gamma=scale, executes the model.fit() method to train the model, and applies cross_val_score() to perform a five-fold cross validation to generate a security risk assessment; The security status update submodule is based on security risk assessment, uses the Pandas library for data analysis, executes pandas.read_csv() to read assessment data, sets security thresholds, filters risk events through DataFrame.query(), adjusts warning levels, uses DataFrame.to_csv() to update emergency measures records, publishes current security information, and generates real-time security monitoring data.
2. The AI-based mine safety monitoring system according to claim 1, characterized in that: The geological monitoring module includes a geological data analysis submodule, a geological risk identification submodule, and a geological stability assessment submodule; The geological data analysis submodule is based on geological data, adopts a deep belief network, uses TensorFlow and Keras libraries to build a network, initializes the model Sequential(), adds multiple fully connected layers Dense() for feature extraction, sets the number of nodes in each layer, including 100 nodes in the first layer and 50 nodes in the second layer, uses the ReLU activation function, sets the optimizer to Adam, selects mean_squared_error as the loss function, uses the model.compile() method to compile the model, applies the model.fit() method to train the model, sets the training batch batch_size to 32, sets the number of iterations epochs to 10, and generates feature extraction data; The geological risk identification submodule extracts data based on features, adopts cluster analysis method, uses Scikit-learn library to execute K-means algorithm, initializes KMeans class, sets the number of clusters n_clusters to 5, selects k-means++ as init parameter, sets the maximum number of iterations max_iter to 300, calls model.fit() method to cluster data, uses labels_ attribute to obtain cluster labels, and generates geological stability grade classification; The geological stability assessment submodule performs stability assessment based on geological stability grade classification, uses weighted scoring method, defines weights of differentiated risk factors, including fault activity weight 0.3, soil liquefaction potential weight 0.2, writes scoring function calculate_score() for each assessment indicator, applies weighted formula to calculate total score, grades geological stability according to total score, and generates geological stability analysis results.
3. The AI-based mine safety monitoring system according to claim 1, characterized in that: The mine abnormality analysis module includes an operation data analysis submodule, an abnormal pattern recognition submodule, and an abnormal index extraction submodule; The operation data analysis submodule uses an autoencoder to encode data based on the geological stability analysis results, including using TensorFlow and Keras libraries in a Python environment, initializing a Sequential model, sequentially adding encoding layers Dense, setting the number of nodes to 64, using a ReLU activation function, adding a decoding layer Dense, setting the number of nodes to the number of original data features, restoring data, selecting Adam as the optimizer, setting the loss function to mean_squared_error, training the model through the model.fit() method, setting batch_size to 32, and setting epochs to 50, and generating encoded reconstructed data; The abnormal pattern recognition submodule is based on the encoded reconstructed data, uses time series analysis to recognize abnormal patterns, uses Pandas for data preprocessing, creates time series decomposition using the Statsmodels library, tests data stability using the ADF test, and uses the SARIMA model to predict abnormal trends. The model parameters are set to p=2, d=1, and q=2, and the model.fit() method is used for model training. The forecast() method is used to predict future trends and generate trend analysis results. The abnormal indicator extraction submodule performs anomaly detection based on the trend analysis results, defines the anomaly detection threshold as 3 times the standard deviation of the data point, uses the Numpy library to calculate the standard deviation of the data point, identifies the data points that exceed the threshold, and generates an abnormal operation mode.
4. The AI-based mine safety monitoring system according to claim 1, characterized in that: The risk trend prediction module includes a risk model building submodule, a risk prediction analysis submodule, and a trend judgment submodule; The risk model construction submodule is based on the abnormal operation mode and adopts the recurrent neural network algorithm. The model is constructed using the tf.keras.Sequential method in the TensorFlow library. The LSTM layer is added through tf.keras.layers.LSTM, wherein 128 neurons are set in each layer, and the activation function is set to sigmoid in the gate control through the activation=sigmoid parameter. The output layer is added through tf.keras.layers.Dense, and the activation function is set to tanh through the activation=tanh parameter. The model is compiled through the model.compile() method, the optimizer is set to Adam through the optimizer=adam parameter, and the loss function is set to the mean square error through the loss=mean_squared_error parameter. The model is trained through the model.fit() method, and the time step is set to 10 time points of data to generate time series modeling results; The risk prediction analysis submodule continues to apply the RNN model for risk prediction based on the time series modeling results. The model input configuration predicts the risk level at a future time point through the model.predict() method, based on the data of the past 10 time points as input. The learning rate is set by the tf.keras.optimizers.Adam(learning_rate=0.001) parameter of the Adam optimizer during model compilation. The batch size and number of iterations are set by the batch_size=32 and epochs=200 parameters of the model.fit() method during model training to generate risk prediction results. The trend judgment submodule is based on the risk prediction results, adopts the time series analysis method, performs data processing through the numpy and pandas libraries, uses numpy.diff to calculate the changes in risk levels at consecutive time points, uses the pandas.Series.rolling() method to analyze trends, and performs linear regression analysis through the scipy.stats.linregress() method to analyze the trend of predicted values changing over time and generate risk development prediction results.
5. The AI-based mine safety monitoring system according to claim 1, characterized in that: The emergency warning system module includes a warning mechanism design submodule, a genetic algorithm optimization submodule, and a warning signal release submodule; The early warning mechanism design submodule performs the initial design of the early warning mechanism based on the risk development prediction results. The Python script is used to define the set_warning_levels function, which accepts the risk value as an input parameter. The risk is classified into three levels: low, medium, and high according to the preset thresholds of 0.2, 0.5, and 0.8 through the if-elif-else structure. The level is returned as a string, including if risk<=0.2: return Low. The level is used as the basic input for genetic algorithm optimization to generate the early warning parameter design result. The genetic algorithm optimization submodule is based on the warning parameter design result, uses Python's DEAP library to define individuals and fitness types through creator.create, uses base.Toolbox() to create a genetic algorithm toolbox, initializes the population size to 50, encodes individual genes as warning parameter thresholds through tools.initRepeat, registers individual generation functions through toolbox.register() method, sets the crossover probability to 0.7, registers crossover operations through toolbox.register() method, sets the mutation probability to 0.01, registers mutation operations through toolbox.register() method, uses tools.selRoulette() as the selection function for roulette selection, and executes 100 iterations through algorithms.eaSimple() method to generate optimized warning parameters; The warning signal issuing submodule builds a rule engine based on the optimized warning parameters using the PyKnow library, defines the rule trigger_warning, and uses the KnowledgeEngine.rule decorator to compare the real-time risk assessment data with the medium risk threshold, including @Rule(AS.fact<<Risk(risk_level=Medium,risk_value=P(lambda x:x> threshold))), when the condition is met, the execution part triggers the corresponding warning signal, and the action defined by def then(self, fact) is used to generate the warning signal system.
6. The AI-based mine safety monitoring system according to claim 1, characterized in that: The mine risk management and control module includes a risk assessment submodule, a risk detection and identification submodule, and a management and control strategy implementation submodule; The risk assessment submodule is based on real-time security monitoring data and adopts a variational autoencoder, in which the encoder part uses the command encoder.add(Dense(units=256, activation=relu)) to add a dense connection layer, setting the number of units to 256 and the activation function to ReLU. The decoder part uses the same configuration, and the loss function is set to vae_loss=xent_loss+kl_loss, where xent_loss is the reconstruction loss and kl_loss is the KL divergence. Use model.fit() to train the model, extract data features, and generate a feature data set; The risk detection and identification submodule applies the local outlier factor algorithm based on the feature data set, executes LocalOutlierFactor(n_neighbors=20, contamination=0.1), sets the neighborhood size to 20, the contamination parameter to 0.1, uses the fit_predict() method to fit the feature data and predict outliers, identifies outlier data points, and generates a potential risk list; The control strategy implementation submodule classifies risks and assesses potential impacts based on the potential risk list, applies DataFrame.groupby() to group risk categories, uses DataFrame.apply() to evaluate each group of risks, and formulates risk control measures.
