A method for establishing a coal mine digital twin model based on Bayesian network algorithm
The coal mine digital twin model using a Bayesian network algorithm and adversarial framework addresses the limitations of existing systems by predicting mine disasters, enhancing safety through real-time data analysis and proactive measures.
Patent Information
- Application Number
- CN202410352746.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-26
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-03-26
AI Technical Summary
Existing coal mine safety management systems, such as those using digital twin technology and traditional numerical simulation methods, fail to provide detailed predictions of disaster types and have limited effectiveness in enhancing safety.
A method for establishing a coal mine digital twin model using a Bayesian network algorithm, incorporating real-time sensor data, drone imagery, and machine learning to analyze and predict the probability of mine disasters, and embedding disaster factor analysis and Bayesian models in an adversarial framework to optimize the digital twin model.
Enables advanced prediction of mine disasters, allowing for proactive countermeasures, improves model robustness and generalization, and enhances the safety factor of coal mines by providing real-time situational awareness.
Smart Images

Figure CN118196662B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of machine learning, and in particular to a method for establishing a coal mine digital twin model based on a Bayesian network algorithm. Background Art
[0002] Coal mine safety management plays a very important role in coal production and is a key factor in determining the development of mining areas. Although the current situation of coal mine safety management is improving, coal mine safety is still the top priority of mining activities. Therefore, with the development of science and technology, how to apply modern technology or methods to the safety management of coal mines, reduce the occurrence of safety accidents in mining areas, and achieve the goal of the safety management level of mining areas is a difficult problem that needs to be solved urgently.
[0003] Digital twin technology is an emerging digital technology that uses data such as physical models, sensor updates, and operation history to integrate multi-disciplinary, multi-physical, multi-scale, and multi-probability simulation processes to complete mapping in virtual space, thereby reflecting the entire life cycle of the corresponding physical equipment. Digital twin model technology is mainly used in product design, production, prediction, and health management. By establishing a virtual digital twin model, it can simulate and predict various situations and behaviors in the real world, thereby optimizing product design, improving production efficiency, and reducing maintenance costs.
[0004] A Bayesian network is a graphical network based on probabilistic reasoning. Using probabilistic reasoning, a Bayesian network can effectively handle uncertainty, model and reason about uncertainty and unknown factors, and help understand the inherent uncertainty of data. The learning mechanism of a Bayesian network is efficient and flexible. When conditions change, it can discover potentially useful patterns or relationships without modifying the model. In general, Bayesian networks have advantages in handling uncertainty, performing probabilistic reasoning, flexible learning, prediction, and avoiding data overfitting, making them widely used in various fields.
[0005] The invention patent with Chinese application number 202210386640.5 discloses a coal mine rock burst disaster monitoring and early warning system based on digital twins, which consists of: stress sensors, local data storage center, computer processing center, cloud information storage center, PC user information input client, mobile device client, alarm system, and alarm device. The monitoring objects are the stress conditions of the coal and rock mass in the current coal mining area, the impact hazard of the working face, and the support stress conditions. This prior art uses digital twin technology, combined with traditional numerical simulation methods, to monitor data at designated locations for early warning. However, this solution cannot make detailed predictions of disaster types in advance, and its degree of increasing coal mine safety is limited. Summary of the invention
[0006] In view of this, the present invention provides a method for establishing a digital twin model of a coal mine based on a Bayesian network algorithm, installing various sensors in the coal mine, obtaining real-time sensor data, and regularly taking orthophotos by drones to compare the deformation of the mine. Through real-time sensor data and regular drone images, using technologies such as three-dimensional modeling and machine learning, a digital twin of the physical mirror image of the mine is established to display the mine situation in real time. By incorporating the Bayesian network algorithm, the input data can be analyzed and predicted to obtain the probability of occurrence of mine disasters. For different disasters, countermeasures or rectification measures can be taken in advance to effectively improve the safety factor of coal mines.
[0007] The technical purpose of the present invention is achieved in this way:
[0008] The present invention provides a method for establishing a coal mine digital twin model based on a Bayesian network algorithm, comprising the following steps:
[0009] S1: Determine the target mine, conduct real-time element monitoring of the target mine, and obtain real-time data;
[0010] S2: preprocessing the real-time data to obtain first input data, and using the first input data to build an initial digital twin model;
[0011] S3: Determine the disaster type and corresponding disaster factors of the target mine, calculate the forward normalized index according to the disaster type and disaster factors, and then construct and train a disaster factor analysis model, and use the output of the disaster factor analysis model to obtain the topological relationship of the disaster factors;
[0012] S4: construct a Bayesian model, use the topological relationship of the disaster-causing factors and the first input data as the training parameters of the Bayesian model, use superposition weight analysis to iteratively train the Bayesian model, and optimize the Bayesian model;
[0013] S5: Use the adversarial framework to embed the disaster factor analysis model and the Bayesian model into the initial digital twin model in series to obtain the final digital twin model.
[0014] Based on the above solution, preferably, step S1 includes:
[0015] S11: installing multiple types of sensors in the target mine, and using the sensors to monitor environmental parameters in the target mine in real time;
[0016] S12: Determine the composition, geological structure and physical and chemical properties of the target mine as geological parameters through field sampling and testing;
[0017] S13: Using geographic information system technology to obtain topographic data of the target mine;
[0018] S14: Regularly use drones to perform orthophotography of target mines to obtain drone image data;
[0019] S15: Acquire satellite remote sensing data of the target mine;
[0020] S16: Use environmental parameters, geological parameters, topographic data, drone image data and satellite remote sensing data as real-time data of the target mine.
[0021] Based on the above solution, preferably, step S2 includes:
[0022] S21: preprocessing the real-time data to obtain preprocessed real-time data, where the preprocessing includes denoising, cleaning and standardization;
[0023] S22: performing feature extraction and feature selection on the preprocessed real-time data to obtain feature data, and enhancing the feature data using a convolutional neural network to obtain first input data;
[0024] S23: Obtain demand analysis and geometric properties of the target mine, establish a three-dimensional model of the target mine according to the demand analysis and the geometric properties, input the first input data into the three-dimensional model for verification and optimization, and obtain an initial digital twin model.
[0025] Based on the above solution, preferably, step S3 includes:
[0026] S31: Determine the disaster type of the target mine, determine the type of disaster factor according to the disaster type, match the first input data with the type of disaster factor, use the successfully matched first input data as disaster factor data, and classify the disaster factor data according to the type of disaster factor to which it belongs to, so as to form category data;
[0027] S32: Calculate the positive normalization index of the category data using a multi-criteria decision analysis method to obtain a calculated value of the category data;
[0028] S33: selecting a regression analysis model, taking the calculated value as a characteristic value of the regression analysis model, and using the first input data to train and verify the regression analysis model to obtain a disaster factor analysis model;
[0029] S34: During the training process, the correlation between the disaster-causing factors and the disasters, that is, the topological relationship of the disaster-causing factors, is obtained according to the output of the disaster-causing factor analysis model.
[0030] Based on the above solution, preferably, step S32 includes:
[0031] S321: extracting features from the category data, including numerical features, category features, and text features, to obtain data features;
[0032] S322: Organize the data features into feature channels, where the feature channels represent feature dimensions in the category data, and each feature channel represents a data feature;
[0033] S324: Perform forward normalization index calculation based on the feature channel on each category data to obtain the maximum eigenvalue in each category data, and use the maximum eigenvalue as the calculation value of the corresponding category data.
[0034] Based on the above solution, preferably, step S324 includes:
[0035] Step 1: Select a set of category data and the extracted data features and feature channels;
[0036] Step 2: For each feature channel, calculate the mean and standard deviation of the feature channel in the small batch samples, where the small batch samples are a batch of category data randomly selected from the group of category data;
[0037] Step 3: Normalize the data features of each sample:
[0038]
[0039] Among them, x' is the normalized data feature, x is the original data feature, μ is the mean, and σ is the standard deviation;
[0040] Step 4: Select a normalized data feature;
[0041] Step 5: Based on the normalized data features, learn a scaling parameter γ and an offset parameter β, and perform the following transformation:
[0042] y=γ×x′+β
[0043] Among them, y is the eigenvalue of the output;
[0044] Step 6, repeating steps 4 and 5, transforming all normalized data features of the group of category data, obtaining all eigenvalues of the group of category data, sorting all eigenvalues from large to small, and selecting the first eigenvalue as the maximum eigenvalue of the group of category data;
[0045] Step 7: Repeat steps 1 to 6 to obtain the maximum eigenvalue of all category data, and use the maximum eigenvalue as the calculated value of the corresponding category data.
[0046] Based on the above solution, preferably, step S4 includes:
[0047] S41: taking the disaster type as the evaluation object, obtaining the topological relationship of the disaster-causing factors corresponding to the disaster type, inputting the topological relationship of the disaster-causing factors and the first input data into the Bayesian network, and iteratively training the Bayesian network. During the iteration, the Bayesian network is used to perform probability evaluation on the evaluation object until the objective function converges, thereby obtaining an optimized Bayesian network;
[0048] S42: According to the topological relationship of the disaster-causing factors in the optimized Bayesian network, feedback is provided to adjust the disaster-causing factor analysis model.
[0049] On the basis of the above scheme, preferably, in step S41, when iteratively training the Bayesian network, the probability score of each evaluation object is calculated using the superposition weight analysis method, and the probability score is used as the occurrence probability of the corresponding evaluation object. The steps of the superposition weight analysis method are as follows:
[0050] Determine the evaluation criteria and evaluation indicators of the evaluation object. The evaluation indicators include qualitative indicators and quantitative indicators. After converting the qualitative indicators into quantitative data, update the evaluation indicators.
[0051] Determine the weight of the evaluation index based on expert opinions, where the weight represents the importance of the evaluation index to the disaster assessment;
[0052] The evaluation indicators are processed by using the standardization method to make the dimensions of different evaluation indicators consistent and obtain the standardized value of each evaluation indicator;
[0053] Multiply the standardized value of each evaluation indicator by the corresponding weight to calculate the weighted score of each evaluation indicator;
[0054] The weighted scores of all evaluation indicators are superimposed to obtain the probability score of the evaluation object.
[0055] Based on the above solution, preferably, the calculation formula of the weighted score is:
[0056] W i =w i ×r′ i
[0057] Where W i is the weighted score of evaluation index i, w i is the weight of evaluation index i, r' i is the standardized value of evaluation index i;
[0058] The superposition formula is:
[0059]
[0060] Where S is the probability score and n is the number of evaluation indicators.
[0061] Based on the above solution, preferably, step S5 includes:
[0062] The disaster factor analysis model and the Bayesian model are embedded in an adversarial framework, and adversarial training is performed on the disaster factor analysis model and the Bayesian model using the training method of the generative adversarial network to optimize the output data of the disaster factor analysis model and the Bayesian model.
[0063] The output data of the optimized disaster factor analysis model and Bayesian model are used as part of the input data of the initial digital twin model. The disaster factor analysis model and the Bayesian network model are respectively embedded in the initial digital twin model in series to obtain the final digital twin model.
[0064] The method of the present invention has the following beneficial effects compared with the prior art:
[0065] (1) The present invention installs various sensors in coal mines to obtain real-time sensor data, and uses drones to regularly take orthophotos to compare the deformation of the mine. Through real-time sensor data and regular drone images, three-dimensional modeling and machine learning technologies are used to establish a digital twin of the mine's physical mirror image to display the mine situation in real time. By incorporating the Bayesian network algorithm, the input data can be analyzed and predicted to obtain the probability of occurrence of mine disasters. For different disasters, countermeasures or rectification measures can be taken in advance to effectively improve the safety factor of coal mines;
[0066] (2) The present invention can reduce the variation range of input data through forward normalization index calculation, making it easier for the model to learn effective feature representations and improving the convergence speed and stability of the model;
[0067] (3) The present invention adopts a method of embedding the disaster factor analysis model and the Bayesian network model into the digital twin model, using serial embedding and adversarial embedding. The output of the disaster factor analysis model and the Bayesian network model is used as part of the input data of the digital twin model. This operation can be regarded as a serial embedding method. However, the output data is not directly input into the digital twin model. The two models of the output data need to be embedded into a general adversarial framework. The idea of using a generative adversarial network can be regarded as adversarial embedding of the model. After being processed by the adversarial framework, the output data of the two models can be further optimized. After being input into the digital twin model, the robustness and generalization ability of the model can be improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0069] Figure 1 is a flow chart of a method according to an embodiment of the present invention;
[0070] Figure 2 This is a technical implementation diagram of an embodiment of the present invention. DETAILED DESCRIPTION
[0071] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0072] like Figure 1 As shown, the present invention provides a method for establishing a coal mine digital twin model based on a Bayesian network algorithm, comprising the following steps:
[0073] S1: Determine the target mine, conduct real-time element monitoring of the target mine, and obtain real-time data;
[0074] S2: preprocessing the real-time data to obtain first input data, and using the first input data to build an initial digital twin model;
[0075] S3: Determine the disaster type and corresponding disaster factors of the target mine, calculate the forward normalized index according to the disaster type and disaster factors, and then construct and train a disaster factor analysis model, and use the output of the disaster factor analysis model to obtain the topological relationship of the disaster factors;
[0076] S4: construct a Bayesian model, use the topological relationship of the disaster-causing factors and the first input data as the training parameters of the Bayesian model, use superposition weight analysis to iteratively train the Bayesian model, and optimize the Bayesian model;
[0077] S5: Use the adversarial framework to embed the disaster factor analysis model and the Bayesian model into the initial digital twin model in series to obtain the final digital twin model.
[0078] See also Figure 2The server operating system of the present invention is Windows 10, the programming language used is Python 3.8, and the programming environment is PyCharm Community; in terms of hardware, the computer processor is Intel (R) Core (TM) i9-10900K CPU @ 3.70GHz, and the running memory is 64GB. The present invention needs to use a large amount of sensor data as input data, including gas, dust, hydrology, temperature, pressure, coal seam GIS and other data in the mine, and construct a digital model of the mine by real-time sensor data to reflect the mine situation in real time.
[0079] The design idea of the present invention is to use the sensor data of the mine as the basic data, use three-dimensional modeling and other technologies, use real-time data as input data, and construct a digital twin model, which is used to synchronize mine information in real time, that is, as a visual management interface for mine safety managers. In the visual digital twin model, according to the existing technology, combined with the actual situation of the mine, the types of disasters that may occur in the mine are enumerated, and then the forward normalized index is calculated according to the topological relationship of the disaster-causing factors, so as to construct a disaster-causing factor analysis model, so that the digital twin can analyze the real-time data of the mine, so as to obtain the safety factor of the mine. Finally, the Bayesian network algorithm is embedded. The eigenvalues in the disaster-causing factor analysis model can only be used for real-time analysis. After inputting the Bayesian network algorithm, the corresponding eigenvalues and real-time data can be predicted, and the type and location of disasters that will occur in the future can be calculated through the network model, so that the management personnel can deal with them in a targeted manner, eliminate the existence of disaster-causing factors or evacuate workers in the disaster location in advance, greatly improving the safety factor of the mine.
[0080] Specifically, in one embodiment of the present invention, step S1 includes:
[0081] S11: Install multiple types of sensors in the target mine, and use the sensors to monitor environmental parameters in the target mine in real time.
[0082] Multiple types of sensors are installed in the target mine, which may include gas sensors, pressure sensors, temperature sensors, etc. These sensors are used to monitor various environmental parameters in the target mine in real time, such as temperature, pressure, gas concentration, dust content, etc. Through these sensors, changes in the internal environment of the mine can be obtained in a timely manner.
[0083] S12: Determine the composition, geological structure and physical and chemical properties of the target mine as geological parameters through field sampling and testing.
[0084] These geological parameters are crucial to understanding the geological characteristics of a mine, and include information such as ore composition, ore layer structure, and geological structure.
[0085] S13: Use geographic information system technology to obtain topographic data of the target mine.
[0086] These data can include information such as terrain elevation, landform features, and surface cover, providing an important geographic information foundation for the subsequent establishment of digital twin models.
[0087] S14: Regularly use drones to perform orthophoto imaging of target mines to obtain drone image data.
[0088] Regularly use drones to perform orthophoto imaging of target mines and obtain high-resolution drone image data. These image data can provide detailed mine surface information and help understand the actual situation and changes of the mine.
[0089] S15: Acquire satellite remote sensing data of the target mine.
[0090] Obtain satellite remote sensing data of the target mine, which can provide mine information in a wider area, including the surrounding environment of the mine, vegetation coverage, etc. Satellite remote sensing data can provide more comprehensive spatial information for the digital twin model.
[0091] S16: Use environmental parameters, geological parameters, topographic data, drone image data and satellite remote sensing data as real-time data of the target mine.
[0092] The integration of these data provides the digital twin model with rich multi-source data, which helps to more accurately simulate and predict the status and changes of the mine.
[0093] Specifically, in one embodiment of the present invention, step S2 involves the process of preprocessing, feature extraction and selection of real-time data, and establishing an initial digital twin model. The steps of constructing the digital twin are: first clarify the mining production process, including the production steps and production equipment, the influence relationship of each link, understand the guiding significance of the digital twin to each process and stage, and clarify the demand analysis; then obtain the geometric properties such as the geometric structure, spatial movement, and geometric association of each object, and establish a 3D model. Combined with the spatial movement law of the physical object, the operation effect is rendered and optimized, and then matched and connected to achieve digital and accurate reproduction of the geometric properties.
[0094] In this embodiment, the specific process of step S2 includes:
[0095] S21: Preprocess the real-time data to obtain preprocessed real-time data, where the preprocessing includes denoising, cleaning and standardization.
[0096] The collected sensor data and image data are preprocessed, including denoising, cleaning and standardization, to improve data quality and accuracy. The specific operations are: denoising the input data to eliminate interference signals; cleaning the data to remove outliers; standardizing the data to keep the data within a uniform scale for subsequent processing.
[0097] S22: Perform feature extraction and feature selection on the preprocessed real-time data to obtain feature data, and enhance the feature data using a convolutional neural network to obtain first input data.
[0098] Extract features from preprocessed data and perform feature selection to reduce data dimensions and extract key information. The specific operations are: use sensor network technology, 3D modeling technology and physical simulation technology to transform multi-source data into a usable 3D data model; extract important features and reduce data volume.
[0099] Use convolutional neural networks to process feature data, improve model accuracy, and make the digital twin model simulate the physical mine conditions more accurately. According to the needs of the actual model building process, appropriate source data and feature data can be selected; use convolutional neural networks to process feature data and enhance data feature information to facilitate subsequent processing and model building. For example, in remote sensing image data, feature data is slope, elevation and other information. Among the large amount of data collected, the data type required to build a mine digital twin is identified as feature data.
[0100] S23: Obtain demand analysis and geometric properties of the target mine, establish a three-dimensional model of the target mine according to the demand analysis and the geometric properties, input the first input data into the three-dimensional model for verification and optimization, and obtain an initial digital twin model.
[0101] Select appropriate modeling methods and tools, such as BIM and CAD, to build a three-dimensional data model of the mine. Clarify the influence relationship between the mine production process and each link, and understand the guiding significance of digital twins to each process and stage; obtain geometric properties and build a 3D model; optimize the rendering of the model to achieve digital and accurate reproduction of geometric properties.
[0102] Train the established model with a large amount of data to improve the efficiency and accuracy of the digital twin model's access and processing of real-time mine data. Train the model to enable it to better process real-time data; optimize the model to improve its accuracy and credibility; deploy the trained and optimized model to the actual environment for application.
[0103] Specifically, in one embodiment of the present invention, step S3 includes:
[0104] S31: Determine the disaster type of the target mine, determine the type of disaster factor according to the disaster type, match the first input data with the type of disaster factor, use the successfully matched first input data as disaster factor data, and classify the disaster factor data according to the type of disaster factor to which it belongs to, so as to form category data;
[0105] S32: Calculate the positive normalization index of the category data using a multi-criteria decision analysis method to obtain a calculated value of the category data;
[0106] S33: selecting a regression analysis model, taking the calculated value as a characteristic value of the regression analysis model, and using the first input data to train and verify the regression analysis model to obtain a disaster factor analysis model;
[0107] S34: During the training process, the correlation between the disaster-causing factors and the disasters, that is, the topological relationship of the disaster-causing factors, is obtained according to the output of the disaster-causing factor analysis model.
[0108] Specifically, step S32 includes:
[0109] S321: extracting features from the category data, including numerical features, category features, and text features, to obtain data features;
[0110] S322: Organize the data features into feature channels, where the feature channels represent feature dimensions in the category data, and each feature channel represents a data feature;
[0111] S324: Perform forward normalization index calculation based on the feature channel on each category data to obtain the maximum eigenvalue in each category data, and use the maximum eigenvalue as the calculation value of the corresponding category data.
[0112] Wherein, step S324 includes:
[0113] Step 1: Select a set of category data and the extracted data features and feature channels;
[0114] Step 2: For each feature channel, calculate the mean and standard deviation of the feature channel in the small batch samples, where the small batch samples are a batch of category data randomly selected from the group of category data;
[0115] Step 3: Normalize the data features of each sample:
[0116]
[0117] Among them, x' is the normalized data feature, x is the original data feature, μ is the mean, and σ is the standard deviation;
[0118] Step 4: Select a normalized data feature;
[0119] Step 5: Based on the normalized data features, learn a scaling parameter γ and an offset parameter β, and perform the following transformation:
[0120] y=γ×x′+β
[0121] Among them, y is the eigenvalue of the output;
[0122] Step 6, repeating steps 4 and 5, transforming all normalized data features of the group of category data, obtaining all eigenvalues of the group of category data, sorting all eigenvalues from large to small, and selecting the first eigenvalue as the maximum eigenvalue of the group of category data;
[0123] Step 7: Repeat steps 1 to 6 to obtain the maximum eigenvalue of all category data, and use the maximum eigenvalue as the calculated value of the corresponding category data.
[0124] Let's take a specific example to illustrate:
[0125] The first input data processed by the above steps is used to calculate the forward normalized index, and a disaster factor analysis model is constructed to enable the digital twin to have the disaster analysis function, including the following steps:
[0126] 1. Select the type of disaster to be analyzed and determine the type of disaster-causing factors. In the mining scenario, possible disaster types include landslides, collapses, etc. Match the disaster-causing factors with the input data to confirm the sample data category related to the disaster;
[0127] Input data obtained through sensors or other means need to be cleaned, calibrated and formatted first to ensure data accuracy. After data processing, the disaster factors corresponding to the disasters that need to be monitored are clearly identified. For example, the disaster factors of floods include rainfall, river water levels, terrain and other information. The disaster factors of specific disaster types can also be identified. The disaster factors are the corresponding data information. The corresponding input data called according to actual needs is the disaster factor.
[0128] Second, calculate the weights of the data categories that match the disasters, using a multi-criteria decision analysis method, where the main calculation step is superimposed weight analysis to obtain the proportion of qualified samples to the total number of samples. According to the information gain of the corresponding eigenvalue in each category, the information gain is normalized to obtain the positive normalization index. The largest eigenvalue is selected as the best dividing feature to complete the calculation of the positive normalization index of a certain category. Repeat the steps to obtain the calculated values of multiple categories of data;
[0129] The purpose of forward normalized index calculation is to unify all valid input data into the same valid data segment, and each input data needs to be calculated.
[0130] The specific formula for calculating the positive normalized index is as follows:
[0131] 1. For each feature channel, calculate the mean μ and standard deviation σ of the channel in the current mini-batch samples.
[0132] 2. Normalize the features of each sample using the following formula:
[0133] x'=(x-μ) / σ
[0134] Among them, x' is the normalized feature, x is the original feature, μ is the mean, and σ is the standard deviation.
[0135] 3. Based on the normalized distribution, learn a scaling parameter γ and an offset parameter β, and perform the following transformation:
[0136] y=γ×x'+β
[0137] Among them, y is the final output feature.
[0138] In this way, through the calculation of the forward normalization exponent, the neural network can reduce the variation range of the input data, making it easier for the model to learn effective feature representations and improving the convergence speed and stability of the model.
[0139] Among them, the scaling parameter γ and the offset parameter β are parameters learned in a neural network, and the learning process is:
[0140] Initialization parameters:
[0141] First, the scaling parameter γ and the offset parameter β need to be initialized. Usually, these parameters are initialized to random values or according to some heuristic method.
[0142] Forward propagation:
[0143] Through the forward propagation process of the neural network, the input data is passed through each layer of the network until the final output is obtained.
[0144] Calculate the loss function:
[0145] The loss function is calculated based on the model's output and the true label to measure the accuracy of the model's prediction.
[0146] Backward Propagation:
[0147] The back-propagation algorithm is used to calculate the gradient of the loss function with respect to the scaling parameter γ and the bias parameter β. These gradients represent the rate of change of the loss function with respect to these parameters.
[0148] Parameter update:
[0149] According to the calculated gradient and optimization algorithm, the values of the scaling parameter γ and the offset parameter β are updated to minimize the loss function.
[0150] Repeat training:
[0151] Repeat the above steps to train the neural network through multiple iterations, and continuously adjust the scaling parameters and offset parameters until the model converges or the conditions for stopping training are met.
[0152] 3. The calculated value is used as the characteristic value of the disaster factor analysis model, and the regression analysis model is selected. The molecular model of the disaster factor is constructed through the characteristics and model to analyze the relationship between the disaster factor and the occurrence of disasters;
[0153] step:
[0154] 1. Clarify objectives and problem definitions: Clarify the types of disasters that need to be monitored and the factors that are related to the formation of disasters.
[0155] 2. Data collection: Collect data from sensors, drones, satellites, historical disaster records, geographic information data, and other data, select data sources based on actual conditions, and ensure data integrity and accuracy.
[0156] 3. Data preprocessing: Clean and standardize the collected data to facilitate centralized use and processing.
[0157] 4. Feature extraction: The effective feature information extracted from the input data is the disaster-causing factor. Each feature information is meaningful and can be extracted and selected through expert opinions and domain knowledge.
[0158] 5. Model selection: The effect of disaster factor analysis can be achieved by using a general regression analysis model.
[0159] 6. Model training and verification: Train the model with the training set and verify the model's effect with the verification set. After obtaining excellent results, maintain and update the model.
[0160] The disaster factor analysis model needs to be continuously iterated and updated, and needs to be adjusted multiple times.
[0161] Fourth, by analyzing the output results of the model, the model is trained and optimized to improve the generalization ability of the model so that the model can input and analyze a variety of disaster data.
[0162] Specifically, in one embodiment of the present invention, step S4 includes:
[0163] S41: taking the disaster type as the evaluation object, obtaining the topological relationship of the disaster-causing factors corresponding to the disaster type, inputting the topological relationship of the disaster-causing factors and the first input data into the Bayesian network, and iteratively training the Bayesian network. During the iteration, the Bayesian network is used to perform probability evaluation on the evaluation object until the objective function converges, thereby obtaining an optimized Bayesian network;
[0164] In step S41, when iteratively training the Bayesian network, the probability score of each evaluation object is calculated using the superposition weight analysis method, and the probability score is used as the occurrence probability of the corresponding evaluation object. The steps of the superposition weight analysis method are as follows:
[0165] Determine the evaluation criteria and evaluation indicators of the evaluation object. The evaluation indicators include qualitative indicators and quantitative indicators. After converting the qualitative indicators into quantitative data, update the evaluation indicators.
[0166] Determine the weight of the evaluation index based on expert opinions, where the weight represents the importance of the evaluation index to the disaster assessment;
[0167] The evaluation indicators are processed by using the standardization method to make the dimensions of different evaluation indicators consistent and obtain the standardized value of each evaluation indicator;
[0168] Multiply the standardized value of each evaluation indicator by the corresponding weight to calculate the weighted score of each evaluation indicator;
[0169] The weighted score is calculated as:
[0170] W i =w i ×r′ i
[0171] Where W i is the weighted score of evaluation index i, w i is the weight of evaluation index i, r' i is the standardized value of the evaluation index i.
[0172] The weighted scores of all evaluation indicators are superimposed to obtain the probability score of the evaluation object.
[0173] The superposition formula is:
[0174]
[0175] Where S is the probability score and n is the number of evaluation indicators.
[0176] S42: According to the topological relationship of the disaster-causing factors in the optimized Bayesian network, feedback is provided to adjust the disaster-causing factor analysis model.
[0177] Specifically, the Bayesian network, also known as the belief network or causal network, is a graphical probability model. It is guided by the relevant theoretical foundations of Bayesian probability and contains a lot of graph theory content. It is a multidisciplinary application theory. The Bayesian network represents causal relationships in the form of a directed acyclic graph (DAG), where nodes represent random variables and edges represent conditional dependencies between these variables. This network structure can effectively represent and infer causal relationships between variables, providing a basis for decision-making.
[0178] Let G = (I, E) represent a directed acyclic graph (DAG), where I represents the set of all nodes in the graph and E represents the set of directed connecting segments, and let X = X i , i∈I is the random variable represented by a node i in the directed acyclic graph. If the joint probability of node X can be expressed as:
[0179]
[0180] Among them, p represents probability, pa(i) represents the "cause" of node i, and X is called a Bayesian network relative to a directed acyclic graph G.
[0181] The overall training process of the Bayesian network is as follows:
[0182] Data preparation: Prepare a data set containing training samples, where each sample contains the value of the node and the corresponding category label.
[0183] Structural learning: Determine the structure of the Bayesian network, that is, the directed connections between nodes. Constraint-based search, heuristic search, and Bayesian information criterion can be used.
[0184] Parameter learning: Learn the conditional probability table of each node, that is, determine the conditional probability distribution of each node given its parent node. For example, maximum likelihood estimation, expectation maximization algorithm, etc.
[0185] Model evaluation: The performance of the trained Bayesian network model is evaluated through methods such as cross-validation.
[0186] Model optimization: Adjust and optimize the model based on the evaluation results to improve its generalization ability and prediction accuracy.
[0187] In this embodiment, when training the Bayesian network, the loss function used is the log-likelihood loss function, which is used to measure the difference between the prediction result of the model and the true label, and is defined as follows:
[0188] Let D = {(x (1) ,y (1) ),(x (2) ,y (2)),...,(x (m) ,y (m) )} is a dataset containing m training samples, where x (i) represents the characteristics of the i-th sample, y (i) represents the category label of the i-th sample. Let f(x) be the prediction result of the Bayesian network model, then the loss function is:
[0189]
[0190] Among them, y (i) represents the true label of the i-th sample, f(x (i) ) is the prediction result of the model for the i-th sample.
[0191] After the trained Bayesian network model is obtained, it contains the topological relationship and probability distribution of the disaster-causing factors. The topological relationship between the nodes in the optimized Bayesian network is checked, including the parent-child node relationship and the conditional probability distribution.
[0192] Feedback adjustment of the analysis model: a. Node adjustment: According to the topological relationship and conditional probability distribution of the nodes in the Bayesian network, the nodes in the analysis model are adjusted to ensure that the dependencies and influencing factors between the nodes are accurately reflected in the model. b. Parameter adjustment: According to the conditional probability distribution in the Bayesian network, the parameters in the analysis model are adjusted to improve the prediction accuracy and generalization ability of the model. c. Model verification: Verify and test the adjusted analysis model to evaluate its performance and effect in disaster prediction and assessment.
[0193] Based on the results of feedback adjustments, the disaster factor analysis model is further optimized to make it more accurate and reliable, thereby improving the ability to predict the probability and possibility of disasters.
[0194] Specifically, in one embodiment of the present invention, step S5 includes:
[0195] The disaster factor analysis model and the Bayesian model are embedded in an adversarial framework, and adversarial training is performed on the disaster factor analysis model and the Bayesian model using the training method of the generative adversarial network to optimize the output data of the disaster factor analysis model and the Bayesian model.
[0196] The output data of the optimized disaster factor analysis model and Bayesian model are used as part of the input data of the initial digital twin model. The disaster factor analysis model and the Bayesian network model are respectively embedded in the initial digital twin model in series to obtain the final digital twin model.
[0197] Adversarial training is a method of training the generator and discriminator models through game theory, in which the generator model attempts to generate realistic data samples, while the discriminator model attempts to distinguish between real data and generated data. In step S5, the disaster factor analysis model and the Bayesian network model are embedded in the adversarial framework, and the two models are adversarially trained using the training method of the generative adversarial network. The specific implementation process is as follows:
[0198] Create a general adversarial framework using ideas from Generative Adversarial Networks (GANs).
[0199] The disaster factor analysis model is used as a generator model to generate data samples similar to real data.
[0200] The Bayesian network model is used as the discriminator model to distinguish the generated data samples from the real data.
[0201] Through adversarial training of the generator and the discriminator, the output data of the disaster factor analysis model and the Bayesian network model are optimized to make them more realistic and accurate.
[0202] The output data of the disaster factor analysis model optimized by adversarial training is used as part of the input data of the digital twin model.
[0203] The output data of the Bayesian network model optimized by adversarial training is used as another part of the input data of the digital twin model.
[0204] The output data of the two models are combined with the original input data to form the complete input of the digital twin model.
[0205] In this embodiment, the output data of the disaster factor analysis model and the Bayesian network model processed by the adversarial framework are used as part of the input data of the digital twin model, and have the characteristics of adversarial embedding. Through adversarial embedding, the robustness and generalization ability of the model are improved, thereby enhancing the digital twin model's ability to predict disasters.
[0206] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for establishing a coal mine digital twin model based on a Bayesian network algorithm, characterized in that: The following steps are involved: S1: Determine the target mine, conduct real-time element monitoring of the target mine, and obtain real-time data; S2: preprocessing the real-time data to obtain first input data, and using the first input data to build an initial digital twin model; S3: Determine the disaster type and corresponding disaster factors of the target mine, calculate the forward normalized index according to the disaster type and disaster factors, and then construct and train a disaster factor analysis model, and use the output of the disaster factor analysis model to obtain the topological relationship of the disaster factors; S4: construct a Bayesian model, use the topological relationship of the disaster-causing factors and the first input data as the training parameters of the Bayesian model, use superposition weight analysis to iteratively train the Bayesian model, and optimize the Bayesian model; S5: Use the adversarial framework to embed the disaster factor analysis model and the Bayesian model into the initial digital twin model in series to obtain the final digital twin model; Step S4 includes: S41: taking the disaster type as the evaluation object, obtaining the topological relationship of the disaster-causing factors corresponding to the disaster type, inputting the topological relationship of the disaster-causing factors and the first input data into the Bayesian network, and iteratively training the Bayesian network. During the iteration, the Bayesian network is used to perform probability evaluation on the evaluation object until the objective function converges, thereby obtaining an optimized Bayesian network; S42: According to the topological relationship of the disaster-causing factors in the optimized Bayesian network, feedback is provided to adjust the disaster-causing factor analysis model; In step S41, when iteratively training the Bayesian network, the probability score of each evaluation object is calculated using the superposition weight analysis method, and the probability score is used as the occurrence probability of the corresponding evaluation object. The steps of the superposition weight analysis method are as follows: Determine the evaluation criteria and evaluation indicators of the evaluation object. The evaluation indicators include qualitative indicators and quantitative indicators. After converting the qualitative indicators into quantitative data, update the evaluation indicators. Determine the weight of the evaluation index based on expert opinions, where the weight represents the importance of the evaluation index to the disaster assessment; The evaluation indicators are processed by using the standardization method to make the dimensions of different evaluation indicators consistent and obtain the standardized value of each evaluation indicator; Multiply the standardized value of each evaluation indicator by the corresponding weight to calculate the weighted score of each evaluation indicator; The weighted scores of all evaluation indicators are superimposed to obtain the probability score of the evaluation object; The weighted score is calculated as: IN i =in i ×r′ i Where W i is the weighted score of evaluation index i, w i is the weight of evaluation index i, r' i is the standardized value of evaluation index i; The superposition formula is: In the formula, S is the probability score, and n is the number of evaluation indicators; Step S5 includes: The disaster factor analysis model and the Bayesian model are embedded in an adversarial framework, and adversarial training is performed on the disaster factor analysis model and the Bayesian model using the training method of the generative adversarial network to optimize the output data of the disaster factor analysis model and the Bayesian model. The output data of the optimized disaster factor analysis model and Bayesian model are used as part of the input data of the initial digital twin model. The disaster factor analysis model and the Bayesian network model are respectively embedded in the initial digital twin model in series to obtain the final digital twin model.
2. The method for establishing a coal mine digital twin model based on a Bayesian network algorithm according to claim 1, characterized in that: Step S1 includes: S11: installing multiple types of sensors in the target mine, and using the sensors to monitor environmental parameters in the target mine in real time; S12: Determine the composition, geological structure and physical and chemical properties of the target mine as geological parameters through field sampling and testing; S13: Using geographic information system technology to obtain topographic data of the target mine; S14: Regularly use drones to perform orthophotography of target mines to obtain drone image data; S15: Acquire satellite remote sensing data of the target mine; S16: Use environmental parameters, geological parameters, topographic data, drone image data and satellite remote sensing data as real-time data of the target mine.
3. The method for establishing a coal mine digital twin model based on a Bayesian network algorithm according to claim 1, characterized in that: Step S2 includes: S21: preprocessing the real-time data to obtain preprocessed real-time data, where the preprocessing includes denoising, cleaning and standardization; S22: performing feature extraction and feature selection on the preprocessed real-time data to obtain feature data, and enhancing the feature data using a convolutional neural network to obtain first input data; S23: Obtain demand analysis and geometric properties of the target mine, establish a three-dimensional model of the target mine according to the demand analysis and the geometric properties, input the first input data into the three-dimensional model for verification and optimization, and obtain an initial digital twin model.
4. The method for establishing a coal mine digital twin model based on a Bayesian network algorithm according to claim 1, characterized in that: Step S3 includes: S31: Determine the disaster type of the target mine, determine the type of disaster factor according to the disaster type, match the first input data with the type of disaster factor, use the successfully matched first input data as disaster factor data, and classify the disaster factor data according to the type of disaster factor to which it belongs to, so as to form category data; S32: Calculate the positive normalization index of the category data using a multi-criteria decision analysis method to obtain a calculated value of the category data; S33: selecting a regression analysis model, taking the calculated value as a characteristic value of the regression analysis model, and using the first input data to train and verify the regression analysis model to obtain a disaster factor analysis model; S34: During the training process, the correlation between the disaster-causing factors and the disasters, that is, the topological relationship of the disaster-causing factors, is obtained according to the output of the disaster-causing factor analysis model.
5. The method for establishing a coal mine digital twin model based on a Bayesian network algorithm as claimed in claim 4, characterized in that: Step S32 includes: S321: extracting features from the category data, including numerical features, category features, and text features, to obtain data features; S322: Organize the data features into feature channels, where the feature channels represent feature dimensions in the category data, and each feature channel represents a data feature; S324: Perform forward normalization index calculation based on the feature channel on each category data to obtain the maximum eigenvalue in each category data, and use the maximum eigenvalue as the calculation value of the corresponding category data.
6. The method for establishing a coal mine digital twin model based on a Bayesian network algorithm as claimed in claim 5, characterized in that: Step S324 includes: Step 1: Select a set of category data and the extracted data features and feature channels; Step 2: For each feature channel, calculate the mean and standard deviation of the feature channel in the small batch samples, where the small batch samples are a batch of category data randomly selected from the group of category data; Step 3: Normalize the data features of each sample: Among them, x' is the normalized data feature, x is the original data feature, μ is the mean, and σ is the standard deviation; Step 4: Select a normalized data feature; Step 5: Based on the normalized data features, learn a scaling parameter γ and an offset parameter β, and perform the following transformation: y=γ×x′+β Among them, y is the eigenvalue of the output; Step 6, repeating steps 4 and 5, transforming all normalized data features of the group of category data, obtaining all eigenvalues of the group of category data, sorting all eigenvalues from large to small, and selecting the first eigenvalue as the maximum eigenvalue of the group of category data; Step 7: Repeat steps 1 to 6 to obtain the maximum eigenvalue of all category data, and use the maximum eigenvalue as the calculated value of the corresponding category data.
Citation Information
Patent Citations
Coal mine rock burst disaster monitoring and early warning system based on digital twinning
CN114658489A
Digital-twin-driven fault prognosis method and system for subsea production system of offshore oil
AU2020102863A4
Road traffic situation simulation method based on digital twinning
CN114519302A
