Coal seam anti-reflection effect prediction method based on multi-source data fusion

Abstract

The invention provides a coal seam anti-reflection effect prediction method based on multi-source data fusion, and belongs to the technical field of machine learning prediction. According to the method, coal seam anti-reflection multi-source data is collected in real time, normalization processing is carried out on the data, then a coding mode is designed for different types of data, and key features are extracted; then performing unified serialization on the multi-source representation data, and aligning a time scale; a self-attention mechanism is adopted to calculate the time dependence weight of single-mode data, and the importance of different data sources is dynamically adjusted; and finally, capturing information complementarity of cross-modal data by adopting a cross attention mechanism. And inputting the fused data into the trained coal seam anti-reflection multi-parameter combined prediction model, outputting four key parameters including permeability, gas yield, pressure variation and fracture density, and performing prediction analysis on the coal seam anti-reflection effect. According to the method, comprehensive evaluation of the coal seam anti-reflection effect is achieved, compared with an existing method, the coal seam anti-reflection process can be described more accurately, and prediction precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning prediction, and in particular relates to a method for predicting coal seam permeability enhancement effect based on multi-source data fusion. Background Art

[0002] Accurate prediction of coal seam permeability enhancement effect is of great significance for coalbed methane development and coal mine safety production. As a clean energy, coalbed methane plays an important role in reducing greenhouse gas emissions and optimizing energy structure. However, coal seams themselves have the characteristics of low permeability. How to improve the permeability of coal seams directly affects the mining efficiency and economic benefits of coalbed methane. Therefore, scientific and accurate prediction of coal seam permeability enhancement effect can not only guide the optimization of permeability enhancement transformation measures, but also improve the recovery rate of coalbed methane resources, reduce development costs, and improve energy utilization. In addition, the prediction of coal seam permeability enhancement effect is also crucial for coal mine safety production. In the process of coal mining, coal seam permeability enhancement transformation may cause changes in geological structure, affect coal seam stability, and even lead to safety hazards such as gas outburst and rock collapse. Through effective prediction of permeability enhancement effect, potential risks can be evaluated before construction, and reasonable permeability enhancement plan can be formulated, thereby reducing the occurrence of safety accidents and ensuring the life safety of coal mine workers. From the perspective of environmental protection, efficient mining of coalbed methane can reduce its disorderly emissions, reduce the impact of greenhouse gases on the atmospheric environment, and reduce the ecological damage caused by coal mine gas accidents. Therefore, accurate prediction of coal seam permeability enhancement effect will not only help promote the efficient development of coalbed methane resources, but also promote the development of the coal mining industry towards a safer, greener and more sustainable direction.

[0003] At present, the prediction of coal seam permeability enhancement effect has become an important research direction for coalbed methane development and coal mine safety production. Traditional methods mainly rely on numerical simulation, experimental analysis and empirical formula calculation. Although these methods can evaluate the coal seam permeability enhancement effect to a certain extent, they often have problems such as complex calculation, poor adaptability and high dependence on field data, which makes it difficult to meet the real-time prediction needs under complex geological conditions. In addition, with the development of sensing technology, various types of sensors have been widely used in coal seam permeability monitoring, such as pressure sensors, temperature sensors, gas flow sensors, microseismic sensors and acoustic sensors, providing rich data support for the prediction of coal seam permeability enhancement effect. However, current research mainly focuses on the analysis of a single or a small number of data sources, lacks deep fusion and intelligent modeling of multi-source data, resulting in a large room for improvement in prediction accuracy and adaptability. In recent years, the progress of artificial intelligence technology has provided new ideas for the prediction of coal seam permeability enhancement effect. Methods based on machine learning and deep learning can automatically extract features from massive sensor data, improve prediction accuracy, and have certain adaptive capabilities. However, such methods still face some challenges, such as fusion modeling of multi-source heterogeneous data, abnormal data processing, missing data completion, and multi-parameter joint prediction. Therefore, how to effectively integrate multiple sensor data and build a highly robust and high-precision prediction model remains an important direction of current research. Summary of the invention

[0004] In view of the above problems, the present invention proposes a method for predicting the coal seam permeability enhancement effect based on multi-source data fusion, which includes the following steps: S1, real-time collection of multi-source data on coal seam permeability enhancement, including pressure inside coal seam pores and fractures , coal seam temperature , Coal Seam Humidity , coal seam gas flow , Coal seam vibration amplitude and vibration frequency ; S2, normalize the data, then design corresponding encoding methods for different types of data, extract key features, and construct a unified learnable representation; S3, uniformly serializes multi-source representation data and aligns time scales; then uses the self-attention mechanism to calculate the time-dependent weight of single-modal data and dynamically adjusts the importance of different data sources; finally, the cross-attention mechanism is used to capture the information complementarity of cross-modal data; S4, input the fused data into the trained coal seam permeability enhancement multi-parameter joint prediction model, which includes a multi-granularity time feature extraction network, a cross-attention mechanism layer and a hierarchical linear regression layer; firstly, a multi-granularity time feature extraction network is used to extract time-dependent features of different granularities of multi-source data; secondly, a cross-attention layer is used to realize the cross-fusion of features of different granularities; finally, the cross-fusion features are input into the hierarchical linear regression layer to predict the output permeability prediction value , Gas production forecast , Prediction of pressure change , crack density prediction value Four key parameters; S5, based on the prediction output curves of the four key parameters of the time series, the coal seam permeability enhancement effect is predicted and analyzed.

[0005] Furthermore, the S2 specifically includes: First, encode the numerical data; for numerical data ( ), design a learnable data segment Token with a dimension of 1024, encode the numerical type data by multiplying the Token and the numerical data, and map the numerical data to a high-dimensional space; Then, temporal position coding is introduced to capture the dynamic characteristics of time series. Sine and cosine position coding are used to introduce explicit time information for different time steps. After combining with temporal position coding, the final coding representation is obtained. Finally, a multi-layer perceptron is used to map the encoded data to a higher-dimensional feature expression. The MLP consists of two fully connected layers, and the activation function is ReLU; the obtained feature representation is It is the unified coding feature of all data sources, and serves as the basic representation of the subsequent data fusion and the input of the multi-parameter joint prediction model for coal seam permeability enhancement.

[0006] Furthermore, the encoding of numerical data requires first regularizing the numerical data, and then performing learnable vector embedding Token multiplication encoding to encode the data of different data types. Randomly initialize a 1024-dimensional learnable Token : ; in, It is the high-dimensional feature after encoding.

[0007] Furthermore, the specific process of S3 is as follows: First, the multi-source representation data is converted into a unified time series, denoted as ,in Indicates the source of data. Represents the time step Moments from sensors The characteristic vector of is the time series length, and the data of all sensors will be in time steps Align to form a multi-source sequence: ; Then, the self-attention mechanism is used to calculate the fusion weights, and the cross-attention mechanism is used for fusion: finally, a linear mapping layer is used to optimize the fusion features and obtain the global correlation information of multi-source data.

[0008] Furthermore, the process between S3 and S4 also includes abnormal data judgment and missing data supplementation of multi-source data joint prediction; first, combined with the encoding representation of sensor data from other sources, on the basis of S3, it is determined whether a certain source data is an abnormal value to ensure the reliability of the input data; then, the abnormal judgment data and naturally missing data are intelligently predicted and supplemented based on multi-source information.

[0009] Furthermore, the specific process of determining abnormal data by joint prediction of multi-source data is as follows: For the target variable to be determined , the historical time window of the current data time step data And the historical data of the remaining 5 variables , based on S2 and S3, the corresponding characterization is obtained ; Then, we use time series modeling and multivariate joint prediction to establish the correlation between time series and data and build a prediction model. , predict the expected value of the target variable , and the observed value Compare and determine whether there is an abnormality; the time series feature modeling uses Transformer to extract time series features and long-term dependencies, and builds a multi-task joint prediction model through a multi-layer perceptron; Calculating prediction error , then the error statistics analysis, the error of normal data Gaussian distribution , using historical data to calculate the mean and standard deviation ; Then, set the dynamic anomaly determination threshold: ; in is a hyperparameter that controls the degree of abnormality. When , it is judged as an abnormal value; According to the degree of abnormality, two abnormality levels are set: mild abnormality, that is, meeting the conditions ; Severe abnormality, that is, meeting the conditions ; If multiple variables are abnormal at the same time, it is necessary to analyze the global abnormal pattern to determine whether it is a systematic error.

[0010] Furthermore, the missing data supplementation process specifically includes: A hybrid prediction method based on time dependency and spatial correlation is used to establish the correlation between time series and spatial relationships and data, and to construct a spatiotemporal-multivariable joint interpolation model. ; The spatiotemporal-multivariable joint interpolation model includes a long short-term memory network and a recurrent neural network. The long short-term memory network is used to analyze the long-term temporal dependency of different variables, and the recurrent neural network is used to analyze the spatial correlation of different variables to achieve high-precision interpolation of variable data. At the same time, a dynamic compensation mechanism is set up, and confidence calculation is first performed. For each predicted value, the corresponding prediction error is calculated; If the abnormal data judgment of the previous multi-source data joint prediction shows that the prediction error of the variable data is lower than the dynamic abnormal judgment threshold, its confidence is adjusted during completion, and then multi-step prediction smoothing is performed.

[0011] Furthermore, in the multi-parameter joint prediction model of coal seam permeability enhancement, the multi-granularity time feature extraction network includes a fully connected layer, a convolutional layer, a gated unit and an LSTM layer; the multi-granularity time feature extraction network is used to capture the fused time series representation The short-term time-dependent characteristics implied in , mid-term time-dependent characteristics and long-term time-dependent characteristics ; The fully connected layer is responsible for extracting The convolutional layer, gated unit and LSTM layer respectively extract the primary features to obtain the deep features , and ; Then, the cross-attention mechanism layer is used to capture the correlation between multi-granular features to obtain cross-fusion features. ; The hierarchical linear regression layer designs four prediction branches with the same architecture but different parameters. Each branch corresponds to a target variable. For each target variable Construct a hierarchical linear regression prediction branch, the basic form of which is as follows: ; in is the cross-fusion feature of the input, is the regression coefficient, is the regression bias, is the interval number index; is the interval number, indicating that different regression coefficients are used for fitting in different data ranges; each branch shares the underlying features but has an independent prediction layer; The prediction results of four key parameters are permeability, gas production, pressure change and fracture density; The multi-parameter joint prediction model for coal seam permeability increase adopts outlier constraints and smoothing strategies when predicting. If the predicted value deviates from the historical trend, it will be corrected.

[0012] Compared with the prior art, the present invention has the following beneficial effects: (1) Adaptive representation construction of multi-type sensor data: Existing methods usually rely on artificial feature engineering or single data source analysis, which makes it difficult to effectively fuse multi-type data. The present invention designs different encoding methods for different types of sensor data (such as numerical data and time series signals) to achieve automatic and efficient data representation construction. By normalizing numerical data and extracting time series data features, the data availability and model generalization ability are improved; (2) Adaptive interactive fusion of multi-source data: Traditional methods often use simple weighting or splicing to fuse multi-source data, ignoring the dynamic relationship between different data sources. The present invention uses an attention mechanism to serialize multi-source data so that the correlation between different sensor data can be fully mined. This method not only enhances the comprehensive expression ability of data, but also has stronger adaptability in complex environments; (3) Abnormal data determination and missing data supplementation mechanism: Existing methods are relatively simple in processing abnormal data and missing data, such as directly eliminating abnormal values ​​or using the mean to fill missing values, which may lead to information loss. The present invention uses a joint prediction mechanism of multi-source data to determine the abnormal values ​​of a single data source and correct the abnormalities through information from other data sources. At the same time, for naturally missing data, the present invention uses a prediction model for intelligent supplementation to improve the integrity and accuracy of the data; (4) Multi-parameter joint prediction to improve prediction accuracy: Traditional methods mostly predict a single parameter, such as only predicting permeability or gas production, while ignoring the coupling relationship between multiple parameters. The present invention constructs a multi-parameter joint prediction model, which comprehensively considers four key parameters: permeability, gas production, pressure change, and fracture density, to achieve a comprehensive evaluation of the coal seam permeability enhancement effect. Compared with single indicator prediction, this method provides richer information, helps to optimize the coal seam permeability enhancement plan, and improves the reliability and practical value of the prediction.

[0013] In summary, the present invention solves the problems of difficulty in multi-source data fusion, difficulty in identifying abnormal data, and single prediction indicators in the prior art by constructing innovative methods such as adaptive characterization, interactive fusion of multi-source data, intelligent judgment and completion of abnormal data, and joint prediction of multiple parameters. It achieves a more accurate, stable, and comprehensive prediction of the coal seam permeability enhancement effect, and provides reliable technical support for coalbed methane development and coal mine safety production. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is the overall flow chart of the coal seam permeability enhancement effect prediction method based on multi-source data fusion.

[0015] Figure 2 Flowchart for multi-source dataset acquisition for coal seam permeability enhancement.

[0016] Figure 3 Coding representation flow chart for multi-source data.

[0017] Figure 4 This is the flowchart of adaptive interactive fusion of multi-source data.

[0018] Figure 5 Flowchart for abnormal data determination for joint prediction of multi-source data.

[0019] Figure 6 This is the flow chart of the multi-parameter joint prediction model for coal seam permeability increase.

[0020] Figure 7 This is a graph showing the total loss change trend during the training process of an embodiment of the present invention.

[0021] Figure 8 This is a graph showing the changing trend of the test set loss value during the training process of an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The invention will be further described below in conjunction with specific embodiments.

[0023] The present invention proposes a method for predicting the permeability enhancement effect of coal seams based on multi-source data fusion. First, a multi-source data set is constructed to collect and integrate data from pressure sensors, temperature sensors, gas flow sensors, microseismic sensors and acoustic sensors. These data cover the key physical parameters in the process of coal seam permeability enhancement and provide comprehensive input information for subsequent modeling. In the data characterization stage, corresponding encoding methods are designed for different types of data. For example, numerical data is normalized and mapped to a high-dimensional feature space, and time series signal data is extracted through convolution or recursive neural networks to effectively characterize different data types. In the data fusion stage, the present invention adopts an adaptive interactive fusion strategy to connect all sensor data into a unified sequence, and introduces an attention mechanism to mine the correlation between different data sources to achieve dynamic weighted fusion of cross-modal information. This method can enhance the expression ability of important features, while suppressing redundant or low-value information, and improve the adaptability of the model to complex permeability enhancement processes. In response to data anomalies and missing problems, the present invention proposes a joint prediction anomaly judgment and data supplementation mechanism. Through cross-comparison of multi-source data, it is judged whether a certain sensor data deviates from the normal range, and the abnormality is further verified in combination with time series pattern analysis. In addition, the attention-based prediction model is used to complete the abnormal data and naturally missing data to ensure the integrity and reliability of the input data. Finally, the present invention constructs a multi-parameter joint prediction model, taking the four key indicators of permeability, gas production, pressure change, and fracture density as target variables, comprehensively considering the mutual influence of various parameters, and realizing accurate prediction of the permeability enhancement effect of coal seams.

[0024] The overall process of this embodiment is as follows Figure 1 As shown: S1, coal seam permeability enhancement multi-source data set collection: deploy corresponding sensors, collect pressure, temperature, gas flow, microseismic, acoustic wave and other sensor data as multi-source data, measure permeability, gas production, pressure change, fracture density and other four indicators as annotations, and construct a multi-source data set; S2, multi-source data encoding representation: design corresponding encoding methods for different types of data, normalize numerical data first, extract key features from time series data, and build a unified learnable representation; S3, adaptive interactive fusion of multi-source data; first, the multi-source data is serialized uniformly and the time scale is aligned; second, the self-attention mechanism is used to calculate the time-dependent weight of single-modal data and dynamically adjust the importance of different data sources; finally, the cross-attention mechanism is used to capture the information complementarity of cross-modal data; S4, abnormal data determination of multi-source data joint prediction: combined with the encoding representation of sensor data from other sources, based on S3, determine whether a certain source data is an abnormal value to ensure the reliability of the input data; S5, missing data supplementation: intelligent prediction and supplementation of abnormal judgment data and naturally missing data based on multi-source information to improve data integrity and consistency; S6, multi-parameter joint prediction model for coal seam permeability enhancement; the model includes a multi-granularity time feature extraction network, a cross-attention mechanism layer and a hierarchical linear regression layer; firstly, a multi-granularity time feature extraction network is used to extract time-dependent features of different granularities of multi-source data; secondly, a cross-attention layer is used to achieve feature fusion of different granularities; finally, the fused features are input into the hierarchical linear regression layer to achieve multi-task prediction of four key parameters: permeability, gas production, pressure change and fracture density; S7, model deployment and testing: deploy the trained prediction model to the coal mine monitoring system to achieve real-time data input, automatic prediction and abnormal warning.

[0025] The specific implementation process of the present invention is described in detail below in conjunction with specific embodiments.

[0026] 1. Coal seam permeability enhancement multi-source data set collection: 1. Deploy the following four types of sensors in the coal seam permeability enhancement operation area to monitor and collect key physical parameters: The pressure sensor is used to measure the pressure changes in the pores and cracks of the coal seam. The unit is MPa, and the sampling frequency is 10 times per second (i.e., 100ms sampling interval), which is recorded as ; The temperature and humidity sensor monitors the coal seam temperature (unit: °C) and humidity (unit: %RH) with a sampling frequency of 10 times per second, which are recorded as , ; Gas flow sensor, which measures the change of coal seam gas flow and air circulation in m³ / h, is used to evaluate the gas migration capacity of coal seam after permeability enhancement. The sampling frequency is 10 times per second, recorded as ; The microseismic sensor records the microseismic events in the coal seam, including the vibration amplitude (unit: m / s²) and frequency (unit: Hz). The sampling frequency is 10 times per second, which is recorded as ; 2. Data preprocessing: First, perform timing alignment. Since the sampling frequencies of various sensors may be different, timestamp synchronization is required to ensure data consistency. 3. Data annotation: Combined with the experimental data of coal seam permeability enhancement and actual manual measurement, the samples are annotated, including four indicators: permeability, gas production, pressure change, and fracture density, which are recorded as , in order to provide supervision signals for subsequent model training and data support for model evaluation; The final data can be expressed as follows: ,in is the time step; The complete process of collecting the multi-source data set for coal seam permeability enhancement is as follows: Figure 2 shown.

[0027] 2. Multi-source data encoding and representation 1. Numerical data encoding: For numeric data ( ), for each type of numerical data, a learnable data segment Token with a dimension of 1024 is designed, and the encoding of the numerical type data is realized by multiplying the Token and the numerical data, and the numerical data is mapped to a high-dimensional space; First, the numerical data is regularized. Since the physical magnitudes of different data are different, the numerical data is first regularized: ; in, represent the mean and standard deviation of the data respectively. ; Then, learnable vector embedding Token multiplication encoding is performed. (such as pressure, temperature, gas flow, etc.) Randomly initialize a 1024-dimensional learnable Token ; ; in, After encoding High-dimensional features of data sources, For the Token embedding vector of each data source; 2. Time series data encoding: Since the present invention performs a sampling operation on the time series data at a frequency of 10 Hz, it is necessary to consider the time series correlation between the data and obtain the pressure inside the coal seam pores and fractures. , coal seam temperature , Coal Seam Humidity , Coalbed gas flow changes , Coal seam vibration amplitude , Coal seam vibration frequency In addition to the above numerical encoding, it is also necessary to introduce temporal position encoding to capture the dynamic characteristics of the time series. The present invention adopts sine and cosine position encoding to introduce explicit time information for different time steps: ; in, is the dimension of the encoded features, is the time step, is the dimension index, is the sinusoidal position encoding, is the cosine position code; Index the dimensions From 1 to 1024, when Calculate when odd When Calculate when it is an even number The calculated sine position encoding value and cosine position encoding value are indexed according to the dimension Arrange to form a temporal position code with a length of 1024 ; Combined with temporal position encoding After that, the final encoding can be obtained as: ; in, is the time step, for forward The value of the time step; 3. Unified coding mapping: In order to uniformly map the coding features of different data sources and make them more expressive, the present invention proposes to use an MLP (multi-layer perceptron) to map the encoded data to a higher-dimensional feature expression. The MLP consists of two fully connected layers, and the activation function is ReLU. The feature expression after mapping is: ; in, and Represent the weight and bias parameters of the first fully connected layer, respectively. and Represent the weight and bias parameters of the second fully connected layer, respectively. is the intermediate encoding feature of all data sources, For the Data sources ago The encoded representation of the data of time steps is obtained after temporal position encoding, which is randomly initialized and the entire fully connected layer is learnable. The final feature representation is It is the unified coding feature of all data sources and can be used as the basic representation for subsequent data fusion and model input; The flowchart of the entire multi-source data encoding representation is as follows Figure 3 shown.

[0028] 3. Adaptive interactive fusion of multi-source data 1. Data serialization: In order to process data from different sources uniformly, the present invention first converts them into a unified time series, denoted as ,in represents the time step, Indicates the data source (sensor type), i.e. ; Represents the time step Moments from sensors The characteristic vector of is the time series length. Finally, all sensor data will be Align to form a multi-source sequence: ; 2. Multi-source data feature fusion: Since the importance of different data sources may change over time, in order to dynamically capture the correlation between different data sources, the present invention uses the self-attention mechanism to calculate the fusion weight; in order to further enhance the information complementarity between different modal data, the present invention uses the cross-attention mechanism for fusion: ; Finally, a linear mapping layer is used to optimize the fusion features: ; It contains global correlation information of multi-source data, providing unified and information-rich feature input for subsequent anomaly detection and prediction tasks; The whole process of adaptive interactive fusion of multi-source data is as follows: Figure 4 shown.

[0029] 4. Abnormal Data Determination for Joint Prediction of Multi-source Data This step aims to utilize the historical information of current data and the joint information of other data sources to perform anomaly detection on various monitoring data inside coal seam pores and fractures through time series modeling and multivariate relationship analysis, screen out unreasonable data points, and ensure the authenticity and reliability of the data.

[0030] Data input and feature extraction: For the target variables that need to be determined ( ), the input data includes the current time step Other 5 sensor data at all times and the historical time window of the current data time step data ; and historical data of the remaining 5 variables According to the steps S2 and S3, the corresponding representation of these input data can be obtained. .

[0031] Time series-multivariate joint anomaly detection model: This step uses time series modeling + multivariate joint prediction to establish the correlation between time series and data and build a prediction model , predict the expected value of the target variable and compare it with the observed value to determine whether there is an anomaly; ; That is, predict the target variable When predicting, it not only relies on the other five variables, but also uses its own historical trends to make predictions; For time series feature modeling, the present invention uses Transformer to extract time series features and long-term dependencies, and builds a multi-task joint prediction model through a multi-layer perceptron:

[0032] Abnormal data determination and processing: Calculating prediction error , then the error statistics analysis, the error of normal data Gaussian distribution , using historical data to calculate the mean and standard deviation . Then, set the dynamic anomaly determination threshold: ; in is a hyperparameter that controls the degree of abnormality. , which represents a 99.7% confidence interval, when When , it is judged as an abnormal value; According to the degree of abnormality, the present invention sets two abnormality levels: mild abnormality, that is, meeting the condition ; Severe abnormality, that is, meeting the conditions If multiple variables are abnormal at the same time, it is necessary to analyze the global abnormal pattern to determine whether it is a systematic error. At the same time, the abnormal data determined are sent to the missing data supplement part and corrected by interpolation prediction; This method combines time trends with multivariate relationships to improve the accuracy of anomaly detection and provide high-quality data support for the prediction of coal seam permeability enhancement effects. Figure 5 shown.

[0033] V. Missing Data Supplementation This step aims to construct a spatiotemporal-multivariate joint interpolation model to supplement the missing data, ensure the integrity of the data, and provide high-quality data input for the subsequent joint prediction of permeability, gas production, pressure change, and fracture density. Missing data mainly includes: abnormal data correction (from abnormal data detection results); natural missing data (data vacancies caused by sensor failure, signal loss, etc.).

[0034] Spatiotemporal-multivariate joint interpolation model: This step uses a hybrid prediction method based on time dependency and spatial correlation to establish the correlation between time series and spatial relationships and data, and build a prediction model ; ; That is, predict the target variable When , it not only depends on the other five variables, but also uses the time dependence and spatial correlation of its own series; The spatiotemporal-multivariable joint interpolation model includes long short-term memory networks and recurrent neural networks. The long short-term memory network is used to analyze the long-term temporal dependencies of different variables, and the recurrent neural network is used to analyze the spatial correlation of different variables to achieve high-precision interpolation of variable data: ; in stands for Long Short-Term Memory Network, Recurrent Neural Network.

[0035] Dynamic uncertainty compensation mechanism: Since the predicted value of missing data has a certain uncertainty, in order to improve the reliability of the completed data, a dynamic compensation mechanism is set. First, the confidence calculation is performed. For each predicted value , calculate the corresponding prediction error: ; If the abnormal data judgment of the previous multi-source data joint prediction shows that the prediction error of the variable data is lower than the dynamic abnormal judgment threshold, its confidence is adjusted during completion: ; in, is the calculated dynamic anomaly threshold; is the confidence compensation factor (dynamically adjusted according to historical data), is the target variable predicted value time series, is the time series of target variable prediction values ​​after confidence adjustment; After that, multi-step prediction smoothing is performed. If the data is missing for a long time, single-step prediction may produce error accumulation, so multi-step prediction smoothing is used: ; in, To smooth the window size, the present invention sets it to 5 steps, namely .

[0036] VI. Multi-parameter Joint Prediction Model of Layer Reflectance Enhancement This step builds a multi-parameter joint prediction model for coal seam permeability enhancement based on the processed multi-source data representation to achieve , Gas production , Pressure change , crack density The simultaneous prediction of the four prediction data reflecting the coal seam permeability enhancement effect is introduced into the prediction result correction and smoothing strategy. The multi-parameter joint prediction model of coal seam permeability enhancement includes a multi-granularity time feature extraction network, a multi-head attention mechanism layer and a hierarchical linear regression layer, which combines time series modeling with multi-variable relationship modeling to accurately depict the dynamic change trend of coal seam permeability enhancement effect.

[0037] Time series feature extraction: After the input data is processed by S1-S5, the fused time series representation is obtained: ; in, For the current moment The fused feature vector contains the time series information of multi-source data; The multi-granularity temporal feature extraction network includes a fully connected layer, a convolutional layer, a gated unit, and an LSTM layer. The short-term time-dependent characteristics implied in , mid-term time-dependent characteristics and long-term time-dependent characteristics ; The fully connected layer is responsible for extracting The convolutional layer, gated unit and LSTM layer respectively extract the primary features to obtain the deep features , and , as follows: ; ); ); in represents the fully connected layer, represents the gate control unit, represents the convolutional layer; Use the cross-attention mechanism layer to capture the correlation between multi-granularity features to obtain cross-features ;

[0038] Stepped linear regression layer: extract cross features To perform joint prediction of four tasks, we specifically designed four prediction branches with the same architecture but different parameters. Here, we use the hierarchical linear regression method to improve the interpretability and stability of the prediction. Each branch corresponds to a target variable. For each target variable ,in is the predicted value of permeability, is the predicted value of gas production, is the predicted value of pressure change, To predict the crack density, a Piecewise Linear Regression (PLR) branch is constructed. Its basic form is as follows: ; is the regression coefficient, is the regression bias, is the interval number index; The prediction results of four key parameters are permeability, gas production, pressure change and fracture density; is an interval number, indicating that different regression coefficients are used for fitting in different data ranges. ; Each branch shares the underlying features but has an independent prediction layer to enhance the pertinence of the prediction and ensure the stability of the prediction.

[0039] Loss function design: Since the four target variables are correlated, the weighted loss of multi-task learning is adopted: ; The loss of each task adopts the mean square error (MSE), and for the loss weight, the present invention adopts ; ; in, For the Data sources in The predicted value at time, For the Data sources in The true value of the moment.

[0040] Correction and output of forecast results: In order to improve the stability and accuracy of the forecast, an outlier constraint and smoothing strategy is adopted. If the forecast value deviates significantly from the historical trend, it will be corrected: ; in, Control the smoothing weight to ensure that the predicted value will not mutate. , For the Data sources in The true value of the moment.

[0041] The short-term trend is also smoothed using sliding window averaging. This strategy can reduce high-frequency noise and improve the continuity of the forecast curve. ; in, For the first Data sources in The predicted value at the moment, and finally the prediction result output is obtained to generate the future Prediction results at this moment: ,in For the future The penetration rate forecast time series at the moment, For the future The gas production forecast time series at the moment, For the future The pressure change prediction time series at each moment, For the future The prediction results are used for coal seam transformation optimization decision-making, such as evaluating the permeability improvement trend, analyzing the change of coal seam gas production, judging the impact of pressure adjustment on coal seam permeability enhancement, and predicting the evolution trend of fracture density. Specifically: Permeability improvement trend: Permeability is a key indicator of gas flow in coal seams and directly affects the efficiency of coal seam gas extraction. By predicting the permeability improvement trend, we can understand whether the coal seam transformation has effectively improved gas mobility, and thus determine whether the coal seam permeability enhancement effect has achieved the expected goal. If the permeability is low, it is necessary to further optimize the fracture structure of the coal seam or take other measures to increase the gas channel; Gas production: Coal seam gas production is a direct measure of the effect of coal seam permeability enhancement. An increase in gas production usually means that the permeability and permeability of the coal seam have been improved. By analyzing the changes in gas production, the degree of permeability enhancement of the coal seam can be evaluated, and then it can be determined whether it is necessary to adjust the mining strategy, transformation plan or pressure control strategy; The impact of pressure change on coal seam permeability enhancement: Pressure change has a significant impact on the permeability and gas release capacity of coal seams. Appropriate pressure reduction can cause coal seams to release more gas, but excessive pressure changes may cause coal seam damage or unstable gas flow. By predicting the relationship between pressure change and coal seam permeability enhancement, decision makers can reasonably adjust the fracturing pressure or gas injection pressure to optimize the coal seam permeability enhancement effect; Evolution trend of fracture density: Fracture density is one of the physical properties of coal seams, which directly affects the permeability of coal seams and the connectivity of fracture channels. Too low fracture density will make it difficult for coal seam gas to flow, affecting the permeability enhancement effect; while too high fracture density may cause coal seam instability and even cause collapse or gas leakage. By predicting the evolution trend of fracture density, decision makers can better assess the health of fractures, take timely measures to repair or adjust the fracture structure of coal seams, and ensure the long-term effect of coal seam transformation.

[0042] In general, the joint prediction of these four parameters provides a comprehensive evaluation framework for coal seam permeability enhancement effects, helping decision makers to accurately adjust transformation strategies, optimize coal seam permeability, maximize coal seam gas extraction benefits, and ensure the long-term stability and safety of coal seams.

[0043] Through the Transformer structure and multi-task learning method, the joint prediction of the four core parameters of coal seam permeability enhancement is achieved, and the prediction results are optimized by using anomaly detection and smoothing strategies to ensure the accuracy and stability of the prediction, providing data support for coal seam permeability enhancement optimization. The data processing flow chart of the entire coal seam permeability enhancement multi-parameter joint prediction model is shown in the figure. Figure 6 shown.

[0044] The trained multi-parameter joint prediction model for coal seam permeability enhancement is deployed to the coal mine monitoring system to achieve real-time data input, automatic prediction and abnormal warning. First, the model needs to be packaged and optimized to adapt to the actual industrial deployment environment. The model is converted to ONNX and TensorRT formats to improve reasoning efficiency, and the prediction results are used for real-time visualization and abnormal warning. When the predicted value deviates from the set threshold, an alarm is automatically triggered, combined with the anomaly detection module reused by S4-S5 to reduce false alarms and improve detection accuracy.

[0045] VII. Experimental results prove This experiment is based on the multi-source sensor monitoring system actually deployed on-site in the coal mine. It continuously collects coal seam permeability-related data for 15 days, including six multi-source time series data such as pressure (P), temperature and humidity (T, W), gas flow (F), vibration amplitude (M), and vibration frequency (f), as well as four target parameters such as permeability (K), gas production (Q), pressure change (ΔP), and fracture density (D). The sampling frequency of the data acquisition system is set to 10Hz, and the ratio of training to test data is 9:1.

[0046] During the training process, the total loss value changes as follows Figure 7 As shown, accordingly, after each round of training, the test is performed, and the change process of the corresponding test loss value is as follows Figure 8 As shown. Figure 7 and Figure 8 It can be seen that during the entire training process, the value of the loss function continues to decrease, the model continues to converge, and finally tends to be stable.

[0047] In terms of model performance evaluation, the model is verified from three perspectives: multi-parameter prediction, abnormal data determination, and missing data completion. For multi-parameter prediction tasks, the following three error evaluation indicators are used for analysis: Mean Absolute Error (MAE): measures the average absolute error between the predicted value and the true value. The smaller the value, the more accurate the prediction result; Mean Squared Error (MSE): calculates the average square error between the predicted value and the true value, which can amplify the impact of large errors on the overall error to evaluate the stability of the model; Root Mean Squared Error (RMSE): is the square root of MSE, which measures the overall level of prediction error. The unit is consistent with the original data, which more intuitively reflects the prediction accuracy; the experimental results are shown in Table 1.

[0048] Table 1 Multi-parameter prediction experimental results

[0049] For the abnormal data judgment task, accuracy is used as the core evaluation indicator, and the experimental results are shown in Table 2.

[0050] Table 2 Experimental results of abnormal data judgment accuracy

[0051] For the missing data completion task, MAE and MSE are used to evaluate the completion effect of the model in the case of missing data, and the experimental results are shown in Table 3. Lower MAE and MSE indicate that the data completed by the model is close to the true value, which improves data integrity and ensures the reliability of the prediction results.

[0052] Table 3 Missing data completion experimental results

[0053] The experimental results show that the method of the present invention achieves low MAE, MSE and RMSE values ​​in the four core parameter prediction tasks, proving that the model can accurately predict the permeability enhancement effect of coal seams; in the abnormal data judgment task, the model has a high accuracy rate and effectively avoids abnormal data interference; in the missing data completion task, the MAE and MSE errors are small, and the completed data is highly consistent with the real data. In summary, the experiment verifies the effectiveness and practical application value of the present invention in the task of predicting the permeability enhancement effect of coal seams.

[0054] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0055] Although the above describes the specific implementation methods of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting coal seam permeability enhancement effect based on multi-source data fusion, characterized in that: The following steps are involved: S1, real-time collection of multi-source data on coal seam permeability enhancement, including pressure inside coal seam pores and fractures , coal seam temperature , Coal Seam Humidity , coal seam gas flow , Coal seam vibration amplitude and vibration frequency ; S2, normalize the data, then design corresponding encoding methods for different types of data, extract key features, and construct a unified learnable representation; S3, uniformly serializes multi-source representation data and aligns time scales; Then, the self-attention mechanism is used to calculate the time-dependent weight of single-modal data and dynamically adjust the importance of different data sources; finally, the cross-attention mechanism is used to capture the information complementarity of cross-modal data; S4, input the fused data into the trained coal seam permeability enhancement multi-parameter joint prediction model, which includes a multi-granularity time feature extraction network, a cross-attention mechanism layer and a hierarchical linear regression layer; firstly, a multi-granularity time feature extraction network is used to extract time-dependent features of different granularities of multi-source data; Secondly, the cross-attention layer is used to realize the cross-fusion of features of different granularities; finally, the cross-fusion features are input into the hierarchical linear regression layer to predict the output permeability prediction value , Gas production forecast , Prediction of pressure change , crack density prediction value Four key parameters; S5, based on the prediction output curves of the four key parameters of the time series, the coal seam permeability enhancement effect is predicted and analyzed.

2. A method for predicting coal seam permeability enhancement effect based on multi-source data fusion according to claim 1, characterized in that: The S2 specifically includes: First, encode the numerical data; for numerical data ( ), for each type of data, a learnable data fragment Token with a dimension of 1024 is designed, and the encoding of the numerical type data is realized by multiplying the Token and the numerical data, and the numerical data is mapped to a high-dimensional space; Then, temporal position coding is introduced to capture the dynamic characteristics of time series. Sine and cosine position coding are used to introduce explicit time information for different time steps. After combining with temporal position coding, the final coding representation is obtained. Finally, a multi-layer perceptron is used to map the encoded data to a higher-dimensional feature expression. The MLP consists of two fully connected layers, and the activation function is ReLU; the obtained feature representation is It is the unified coding feature of all data sources, and serves as the basic representation of the subsequent data fusion and the input of the multi-parameter joint prediction model of coal seam permeability enhancement.

3. A method for predicting coal seam permeability enhancement effect based on multi-source data fusion as claimed in claim 2, characterized in that: The encoding of numerical data requires first regularizing the numerical data, and then performing learnable vector embedding Token multiplication encoding to encode the data of different data types. Randomly initialize a 1024-dimensional learnable Token : ； in, It is the high-dimensional feature after encoding.

4. The method for predicting coal seam permeability enhancement effect based on multi-source data fusion according to claim 1, characterized in that: The specific process of S3 is as follows: First, the multi-source representation data is converted into a unified time series, denoted as ,in Indicates the source of data. Represents the time step Moments from sensors The characteristic vector of is the time series length, and the data of all sensors will be in time steps Align to form a multi-source sequence: ； Then, the self-attention mechanism is used to calculate the fusion weights, and the cross-attention mechanism is used for fusion: finally, a linear mapping layer is used to optimize the fusion features and obtain the global correlation information of multi-source data.

5. The method for predicting coal seam permeability enhancement effect based on multi-source data fusion according to claim 1, characterized in that: The process between S3 and S4 also includes abnormal data determination and missing data supplementation for joint prediction of multi-source data. First, based on S3, the sensor data encoding representation of other sources is combined to determine whether a certain source data is an abnormal value to ensure the reliability of the input data. Then, based on multi-source information, intelligent prediction and completion are performed on abnormal judgment data and naturally missing data.

6. A method for predicting coal seam permeability enhancement effect based on multi-source data fusion as claimed in claim 5, characterized in that: The specific process of abnormal data determination in the multi-source data joint prediction is as follows: For the target variable to be determined , the historical time window of the current data time step data And the historical data of the remaining 5 variables , based on S2 and S3, the corresponding characterization is obtained ; Then, we use time series modeling and multivariate joint prediction to establish the correlation between time series and data and build a prediction model. , predict the expected value of the target variable , and the observed value Compare and determine whether there is an abnormality; the time series modeling uses Transformer to extract time series features and long-term dependencies, and builds a multi-task joint prediction model through a multi-layer perceptron; Calculating prediction error , then the error statistics analysis, the error of normal data Gaussian distribution , using historical data to calculate the mean and standard deviation ; Then, set the dynamic anomaly determination threshold: ； in is a hyperparameter that controls the degree of abnormality. When , it is judged as an abnormal value; According to the degree of abnormality, two abnormality levels are set: mild abnormality, that is, meeting the conditions ; Severe abnormality, that is, meeting the conditions ; If multiple variables are abnormal at the same time, it is necessary to analyze the global abnormal pattern to determine whether it is a systematic error.

7. The method for predicting coal seam permeability enhancement effect based on multi-source data fusion according to claim 5, characterized in that: The missing data supplementation process specifically includes: A hybrid prediction method based on time dependency and spatial correlation is used to establish the correlation between time series and spatial relationships and data, and to construct a spatiotemporal-multivariable joint interpolation model. ; The spatiotemporal-multivariate joint interpolation model includes long short-term memory networks and recurrent neural networks. Long short-term memory networks are used to analyze the long-term temporal dependencies of different variables, and recurrent neural networks are used to analyze the spatial correlations of different variables. At the same time, a dynamic compensation mechanism is set up, and confidence calculation is first performed. For each predicted value, the corresponding prediction error is calculated; If the previous anomaly detection indicates that the reliability of the variable data is below the threshold, its confidence is adjusted during completion, and then multi-step forecast smoothing is performed.

8. The method for predicting coal seam permeability enhancement effect based on multi-source data fusion according to claim 1, characterized in that: In the multi-parameter joint prediction model for coal seam permeability enhancement, the multi-granularity time feature extraction network includes a fully connected layer, a convolutional layer, a gated unit and an LSTM layer; the multi-granularity time feature extraction network is used to capture the fused time series representation The short-term time-dependent characteristics implied in , mid-term time-dependent characteristics and long-term time-dependent characteristics ; The fully connected layer is responsible for extracting The convolutional layer, gated unit and LSTM layer respectively extract the primary features to obtain the deep features , and ; Then, the cross-attention mechanism layer is used to capture the correlation between multi-granular features to obtain cross-fusion features. ; The hierarchical linear regression layer designs four prediction branches with the same architecture but different parameters. Each branch corresponds to a target variable. For each target variable Construct a hierarchical linear regression prediction branch, the basic form of which is as follows: ； in is the cross-fusion feature of the input, is the regression coefficient, is the regression bias, is the interval number index; is the interval number, indicating that different regression coefficients are used for fitting in different data ranges; each branch shares the underlying features but has an independent prediction layer; The prediction results of four key parameters are permeability, gas production, pressure change and fracture density; The multi-parameter joint prediction model for coal seam permeability increase adopts outlier constraints and smoothing strategies when predicting. If the predicted value deviates from the historical trend, it will be corrected.

Images