Coal and gas outburst disaster prediction method and device based on multi-physics field coupling

Abstract

The present application relates to the technical field of disaster prediction, and discloses a method and device for predicting coal and gas outburst disasters based on multi-physical field coupling. The method comprises: performing data enhancement and normalization processing on the training sample set to obtain a model training data set, performing pattern recognition and parameter optimization on the model training data set to obtain an outburst disaster prediction model; performing probability distribution calculation on the output data of the outburst disaster prediction model to obtain a confidence interval, performing multi-dimensional cross-validation on the confidence interval to obtain a prediction reliability assessment result; performing threshold grading and risk level division on the prediction reliability assessment result to obtain a warning level determination result, performing spatiotemporal correlation analysis on the warning level determination result to obtain target warning decision information. The present application establishes a multi-physical field coupling prediction model by analyzing the coupling relationship between multi-physical field parameters, thereby achieving accurate prediction of coal and gas outburst disasters, thereby improving the reliability and accuracy of the prediction.

Description

Technical Field

[0001] The present application relates to the field of disaster prediction, and in particular to a method and device for predicting coal and gas outburst disasters based on multi-physical field coupling. Background Art

[0002] Coal and gas outbursts are one of the most serious dynamic hazards in coal mining operations, posing a significant threat to coal mine safety. Currently, coal and gas outburst predictions are primarily based on analysis of single-physics field monitoring data, such as stress, gas pressure, and temperature. With the advancement of monitoring technology, a variety of sensors have been applied to coal mine safety monitoring, including stress sensors, gas concentration sensors, temperature sensors, and acoustic emission sensors. These sensors can collect multi-physics field parameter data in real time. However, existing prediction methods primarily focus on the variation characteristics of single-physics field parameters and fail to fully consider the coupling relationships between multiple physical fields and their combined impact on the evolution of coal and gas outburst hazards.

[0003] The existing technology has the following shortcomings: First, single physical field monitoring is difficult to fully reflect the incubation process of coal and gas outbursts, resulting in low reliability of prediction results; second, although multi-physical field monitoring data contains rich disaster precursor information, due to the lack of effective multi-physical field coupling analysis methods, it is difficult to extract key feature information from massive monitoring data; third, existing prediction methods generally have problems such as insufficient data preprocessing, incomplete feature extraction, and insufficient generalization ability of prediction models, which affect the accuracy and practicality of the prediction. Summary of the Invention

[0004] The present application provides a method and device for predicting coal and gas outburst disasters based on multi-physical field coupling, which is used to establish a multi-physical field coupling prediction model by analyzing the coupling relationship between multi-physical field parameters, thereby achieving accurate prediction of coal and gas outburst disasters, thereby improving the reliability and accuracy of the prediction.

[0005] In a first aspect, the present application provides a method for predicting coal and gas outburst disasters based on multi-physics field coupling, the method comprising: collecting and normalizing the original data of stress field, gas field, temperature field and acoustic emission field at the coal mine site to obtain a multi-physics field initial data set, and performing outlier detection and data completion processing on the multi-physics field initial data set to obtain a preprocessed multi-physics field parameter data set;

[0006] Performing physical field component extraction and characteristic decomposition on the preprocessed multi-physical field parameter data set to obtain a basic parameter matrix of each physical field, and performing cross-correlation analysis and coupling factor calculation on the basic parameter matrix to obtain a physical field coupling relationship matrix;

[0007] Extracting time domain features and frequency domain features from the physical field coupling relationship matrix to obtain a multidimensional feature vector, and performing redundant feature elimination and feature reconstruction on the multidimensional feature vector to obtain an optimized feature parameter set;

[0008] Performing time series segmentation and sample labeling on the optimized feature parameter set to obtain a training sample set, performing data enhancement and normalization processing on the training sample set to obtain a model training data set, and performing pattern recognition and parameter optimization on the model training data set to obtain a prominent disaster prediction model;

[0009] Performing probability distribution calculation on the output data of the prominent disaster prediction model to obtain a confidence interval of the prediction result, and performing multi-dimensional cross-validation on the confidence interval to obtain a prediction reliability assessment result;

[0010] The prediction reliability assessment result is subjected to threshold classification and risk level division to obtain a warning level determination result, and the warning level determination result is subjected to spatiotemporal correlation analysis to obtain target warning decision information.

[0011] In a second aspect, the present application provides a coal and gas outburst disaster prediction device based on multi-physics field coupling, the coal and gas outburst disaster prediction device based on multi-physics field coupling comprising:

[0012] An acquisition module is used to collect and standardize the raw data of the stress field, gas field, temperature field, and acoustic emission field at the coal mine site to obtain an initial multi-physics field data set, and perform outlier detection and data completion processing on the initial multi-physics field data set to obtain a pre-processed multi-physics field parameter data set;

[0013] A decomposition module is used to perform physical field component extraction and feature decomposition on the preprocessed multi-physical field parameter data set to obtain a basic parameter matrix of each physical field, and to perform cross-correlation analysis and coupling factor calculation on the basic parameter matrix to obtain a physical field coupling relationship matrix;

[0014] An extraction module is used to extract time domain features and frequency domain features of the physical field coupling relationship matrix to obtain a multidimensional feature vector, and to eliminate redundant features and reconstruct features of the multidimensional feature vector to obtain an optimized feature parameter set;

[0015] a processing module, configured to perform time series segmentation and sample labeling on the optimized feature parameter set to obtain a training sample set, perform data enhancement and normalization on the training sample set to obtain a model training data set, perform pattern recognition and parameter optimization on the model training data set to obtain a prominent disaster prediction model;

[0016] A calculation module is used to perform probability distribution calculation on the output data of the sudden disaster prediction model to obtain a confidence interval of the prediction result, and to perform multi-dimensional cross-validation on the confidence interval to obtain a prediction reliability assessment result;

[0017] The classification module is used to perform threshold classification and risk level classification on the prediction reliability assessment result to obtain the warning level determination result, and perform spatiotemporal correlation analysis on the warning level determination result to obtain target warning decision information.

[0018] In the technical solution provided by this application, by collecting and preprocessing multi-physics field data of stress field, gas field, temperature field and acoustic emission field at the coal mine site, the noise and outliers in the original data are effectively eliminated, and the data quality is improved. Through physical field component extraction and feature decomposition, the basic characteristics of each physical field are deeply analyzed, and through cross-correlation analysis and coupling factor calculation, the coupling relationship between multiple physical fields is fully revealed, providing more comprehensive feature information for the prediction of prominent disasters. The time domain features and frequency domain features of the physical field coupling relationship matrix are extracted, and the redundant feature elimination and feature reconstruction technology are combined to obtain an optimized feature parameter set, which improves the representativeness and effectiveness of the features. Through time series segmentation and sample labeling, a standardized training sample set is established, and data enhancement and normalization processing methods are used to enhance the representativeness of the samples and the generalization ability of the model. The prominent disaster prediction model based on deep neural network realizes the accurate prediction of prominent risks. The reliability of the prediction results is evaluated through probability distribution calculation and multi-dimensional cross-validation to ensure the credibility of the prediction results. Finally, through threshold classification and risk level categorization, combined with spatiotemporal correlation analysis, scientific early warning decision-making information was generated, providing important decision-making support for coal mine safety production. This method overcomes the incomplete information provided by traditional single-physical field monitoring methods, improves the accuracy and reliability of predictions, and achieves precise early warning of coal and gas outburst disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 This is a schematic diagram of an embodiment of a method for predicting coal and gas outburst disasters based on multi-physics field coupling in an embodiment of the present application;

[0021] Figure 2 This is a schematic diagram of an embodiment of a coal and gas outburst disaster prediction device based on multi-physical field coupling in an embodiment of the present application. DETAILED DESCRIPTION

[0022] The embodiments of the present application provide a method and apparatus for predicting coal and gas outburst disasters based on multi-physics field coupling. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or apparatus.

[0023] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the coal and gas outburst disaster prediction method based on multi-physics field coupling includes:

[0024] Step S101: collecting and normalizing the original data of the stress field, gas field, temperature field, and acoustic emission field at the coal mine site to obtain a multi-physics field initial data set, and performing outlier detection and data completion processing on the multi-physics field initial data set to obtain a pre-processed multi-physics field parameter data set;

[0025] Step S102: performing physical field component extraction and eigendecomposition on the preprocessed multi-physics field parameter data set to obtain a basic parameter matrix of each physical field, and performing cross-correlation analysis and coupling factor calculation on the basic parameter matrix to obtain a physical field coupling relationship matrix;

[0026] Step S103: extracting time domain features and frequency domain features from the physical field coupling relationship matrix to obtain a multidimensional feature vector, and performing redundant feature elimination and feature reconstruction on the multidimensional feature vector to obtain an optimized feature parameter set;

[0027] Step S104: performing time series segmentation and sample labeling on the optimized feature parameter set to obtain a training sample set, performing data enhancement and normalization on the training sample set to obtain a model training data set, performing pattern recognition and parameter optimization on the model training data set to obtain a prominent disaster prediction model;

[0028] Step S105: Calculate the probability distribution of the output data of the prominent disaster prediction model to obtain the confidence interval of the prediction result, and perform multi-dimensional cross-validation on the confidence interval to obtain the prediction reliability assessment result;

[0029] Step S106: threshold classification and risk level classification are performed on the prediction reliability assessment results to obtain warning level determination results, and spatiotemporal correlation analysis is performed on the warning level determination results to obtain target warning decision information.

[0030] It is understandable that the execution subject of this application can be a coal and gas outburst disaster prediction device based on multi-physics field coupling, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0031] Specifically, multi-physics field data is collected through a sensor network deployed on the coal mine working face. Specifically, stress sensors collect coal stress change data, gas sensors collect gas concentration and pressure data, temperature sensors collect coal temperature change data, and acoustic emission sensors collect coal fracture acoustic emission signal data. The collected raw data are standardized to eliminate the dimensional differences of different physical quantities. Abnormal data are processed by peak normalization and median filling, and the sliding window technology is used to segment the data, with the window size set to 1 hour. The interquartile range method is used to detect outliers, and data points that are out of the normal range are marked. The marked abnormal data are supplemented by linear interpolation, and then the noise components in the data are removed by wavelet transform. All physical field data are aligned to a unified time series, and equal interval resampling is used to ensure the temporal consistency of the data.

[0032] Physical field components were extracted from the preprocessed data to obtain characteristic matrices for the stress field, gas field, temperature field, and acoustic emission field. Principal component analysis was used to extract the main characteristic components of each physical field, calculate the eigenvalues ​​and eigenvectors, and select the principal components with a cumulative contribution rate of 85% as key features. Cross-correlation analysis was performed on the extracted features, and the coupling coefficients between the physical fields were calculated to construct a coupling relationship matrix. Time and frequency domain features were extracted from the coupling relationship matrix. Time domain features were obtained through wavelet transform, and frequency domain features were obtained through Fourier transform. The extracted features were then dimensionally concatenated to form a multidimensional feature vector. Correlation analysis and hierarchical clustering were used to eliminate redundant features, resulting in an optimized feature parameter set.

[0033] Training samples were constructed based on an optimized feature parameter set, and time series segmentation was performed using a sliding window with a window size set to 24 hours and a step size set to 1 hour. The segmented samples were labeled as "highly dangerous" and "highly safe." Data rotation and noise injection were used for data enhancement to improve sample representativeness. A prediction model was constructed using a deep neural network, and model performance was improved through parameter optimization. The reliability of the prediction results was evaluated, and confidence intervals were calculated using kernel density estimation. K-fold cross-validation (K=5) was used to assess model stability. The warning level threshold was determined based on the evaluation results, and the warning levels were divided into three levels: "high risk," "medium risk," and "low risk." The spatial distribution characteristics and evolution trends of dangerous areas were determined through spatiotemporal correlation analysis.

[0034] For example, gas pressure data collected at a coal mine working face showed that the normal range was between 0.74 and 2.35 MPa. Outlier detection revealed an outlier of 3.86 MPa. After data completion, the outlier was corrected to 2.28 MPa. Simultaneously, stress monitoring data showed that vertical stress increased from 18.5 MPa to 26.7 MPa, temperature rose from 24.3°C to 28.9°C, and the frequency of acoustic emission events increased from 12 to 45 per hour. Multi-physics field coupling analysis revealed that these abnormal changes in parameters were significantly correlated, with a coupling coefficient of 0.87. The predictive model issued a "high-risk" warning, indicating a significant risk of gas outbursts in the area.

[0035] In the embodiments of the present application, multi-physics field data collection and preprocessing of stress, gas, temperature, and acoustic emission fields at a coal mine site effectively eliminates noise and outliers in the raw data, improving data quality. Through physical field component extraction and feature decomposition, the basic characteristics of each physical field are deeply analyzed. Cross-correlation analysis and coupling factor calculation comprehensively reveal the coupling relationships between multiple physical fields, providing more comprehensive feature information for outburst disaster prediction. Time-domain and frequency-domain feature extraction are performed on the physical field coupling relationship matrix. Combined with redundant feature elimination and feature reconstruction techniques, an optimized feature parameter set is obtained, improving the representativeness and effectiveness of the features. Through time series segmentation and sample labeling, a standardized training sample set is established. Data augmentation and normalization methods are used to enhance the representativeness of the samples and the generalization ability of the model. A deep neural network-based outburst disaster prediction model achieves accurate prediction of outburst risks. The prediction results are evaluated for reliability through probability distribution calculation and multi-dimensional cross-validation, ensuring their credibility. Finally, through threshold classification and risk level division, combined with spatiotemporal correlation analysis, scientific early warning decision information is generated, providing important decision support for coal mine safety production. This method overcomes the defect of incomplete information in traditional single physical field monitoring methods, improves the accuracy and reliability of prediction, and realizes accurate early warning of coal and gas outburst disasters.

[0036] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0037] (1) Collect and standardize the original data of stress field, gas field, temperature field and acoustic emission field at the coal mine site to obtain the initial data set of multi-physics field

[0038] (2) Perform peak normalization on the initial data set of the multi-physics field to obtain a normalized data set, and perform median filling on the normalized data set to obtain the supplemented initial data;

[0039] (3) Perform sliding window segmentation processing on the supplemented initial data to obtain a data segment sequence, and perform interquartile range anomaly detection on the data segment sequence to obtain a labeled data sequence;

[0040] (4) Perform linear interpolation on the outliers in the marked data sequence to obtain an interpolated data sequence, and perform wavelet denoising on the interpolated data sequence to obtain a denoised data sequence;

[0041] (5) Performing time window synchronization processing on the denoised data sequence to obtain an aligned data sequence, and performing data resampling processing on the aligned data sequence to obtain an equally spaced data sequence;

[0042] (6) Perform Z-score normalization on the equally spaced data sequence to obtain a standardized data sequence, and perform batch verification on the standardized data sequence to obtain the initial data set of the multi-physics field;

[0043] (7) Perform 3σ criterion anomaly detection on the initial data set of multi-physics fields to obtain an abnormal point mark sequence, and perform spline interpolation processing on the abnormal point mark sequence to obtain a data completion sequence;

[0044] (8) Perform moving average filtering on the data completion sequence to obtain a smoothed data sequence, and perform data quality assessment on the smoothed data sequence to obtain a data quality index;

[0045] (9) Perform threshold screening on the data quality indicators to obtain a qualified data sequence, and perform outlier detection on the qualified data sequence to obtain a test result sequence;

[0046] (10) Performing data verification processing on the detection result sequence to obtain a verified data sequence, and performing data format unification processing on the verified data sequence to obtain a preprocessed multi-physical field parameter data set.

[0047] Specifically, multiple types of sensors are deployed at the coal mine site to collect data from multiple physical fields. Stress sensors collect coal stress data, including vertical and horizontal stresses; gas sensors collect data on gas concentration and pressure; temperature sensors collect data on coal temperature changes; and acoustic emission sensors collect data on acoustic emission signals generated when coal breaks. The data collection frequency is set to once per second, and continuous data is collected for 24 hours to form the raw data set.

[0048] First, the raw data was peak-normalized. Missing values ​​were supplemented using the median filling method. The supplemented data was segmented using a sliding window technique with a window size of 1 hour and a sliding step of 10 minutes to obtain a sequence of data segments.

[0049] The interquartile range method is used to detect anomalies for each data segment. The calculation formula is:

[0050] ;

[0051] ;

[0052] ;

[0053] in is the lower quartile, is the upper quartile, is the interquartile range, and are the upper and lower thresholds respectively. Data points outside the threshold range are marked as outliers.

[0054] Marked outliers were corrected using linear interpolation, followed by wavelet decomposition for data noise reduction. All physical field data were aligned to a unified time series using time window synchronization, and resampling was performed at equal intervals to ensure temporal consistency. The resampled data were Z-score normalized to eliminate dimensional differences between different physical quantities. Anomaly detection was then performed using the 3σ criterion, and spline interpolation was performed on data points outside the normal range. Finally, the processed data were subjected to a moving average filter with a window size set to 5 data points to obtain a smoothed data sequence. Data quality was assessed by calculating indicators such as data integrity, consistency, and accuracy, and thresholds were set to filter out qualified data. The filtered data were finally standardized to form a preprocessed multi-physics parameter dataset.

[0055] For example, in the gas pressure data collected from a coal mine working face, the original data range is 0.5-3.8 MPa. After peak normalization, the data is mapped to the range of 0-1. An outlier value of 3.8 MPa (normalized value 1.0) is detected by the interquartile range method, where =0.8MPa (normalized value 0.21), =2.1MPa (normalized value 0.55), =1.3MPa, calculated =4.05MPa, = -1.15MPa. Since 3.8MPa is less than the upper threshold of 4.05MPa, this data point is retained. After subsequent data processing steps, a standardized and smoothed data series is obtained.

[0056] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0057] (1) The preprocessed multi-physics field parameter data set is divided into physical field types to obtain stress field data subsets, gas field data subsets, temperature field data subsets, and acoustic emission field data subsets. Each data subset is then sorted in time series to obtain an ordered data sequence;

[0058] (2) Perform principal component decomposition on the ordered data sequence to obtain the main characteristic components, and calculate the variance contribution rate of the main characteristic components to obtain the characteristic weight sequence;

[0059] (3) Perform threshold screening on the feature weight sequence to obtain the key feature sequence, and then perform physical quantity unit unification on the key feature sequence to obtain the standardized feature sequence;

[0060] (4) Perform matrix reconstruction on the standardized characteristic sequence to obtain the basic parameter matrix of each physical field, and perform data integrity check on the basic parameter matrix to obtain the verified parameter matrix;

[0061] (5) Calculate the Pearson correlation coefficient of the verified parameter matrix to obtain the correlation coefficient matrix, and perform a significance test on the correlation coefficient matrix to obtain the effective correlation items;

[0062] (6) Calculate the coupling degree of the effective correlation items to obtain a coupling factor sequence, and normalize the coupling factor sequence to obtain a standardized coupling factor;

[0063] (7) Perform matrix mapping on the standardized coupling factor to obtain the initial coupling relationship matrix, and perform symmetry test on the initial coupling relationship matrix to obtain the physical field coupling relationship matrix.

[0064] Specifically, the preprocessed multi-physics parameter dataset was first classified by physical field type, resulting in four subsets: a stress field data subset (containing vertical stress, horizontal stress, and other data), a gas field data subset (containing gas concentration, pressure, and other data), a temperature field data subset (containing temperature change data), and an acoustic emission field data subset (containing acoustic emission signal data). Each data subset was sorted by timestamp to generate ordered time series data. After obtaining the ordered data series, principal component analysis (PCA) was used to perform eigendecomposition. The eigenvalues ​​and eigenvectors of the data covariance matrix were calculated to obtain the primary eigencomponents. For each eigencomponent, its variance contribution was calculated—the ratio of the eigenvalue corresponding to that component to the sum of all eigenvalues—to obtain a feature weight sequence. A variance contribution threshold of 85% was set to screen out key features whose cumulative contribution reached the threshold, forming a key feature sequence. The physical quantities in the key feature sequence were normalized to eliminate dimensional differences between the different quantities.

[0065] The standardized feature sequence is reconstructed into a matrix, and the key features of each physical field are organized into a basic parameter matrix according to the time series. The quality of the matrix data is ensured by checking the integrity, continuity, and consistency of the data, and a verified parameter matrix is ​​obtained. The Pearson correlation coefficient is calculated for the verified parameter matrix to analyze the correlation between different physical field parameters. A significance test (confidence level set at 95%) is performed to screen out parameter pairs with significant correlations to form valid correlation terms. The coupling degree is calculated for the valid correlation terms to characterize the interaction strength between different physical field parameters. The calculated coupling factors are normalized to unify their value range to between 0 and 1. The initial coupling relationship matrix is ​​constructed for the standardized coupling factors, and the symmetry of the matrix is ​​ensured through a symmetry test to obtain the physical field coupling relationship matrix.

[0066] It should be noted that when constructing the physical field coupling relationship matrix, for each pair of physical field parameters (such as stress field and gas field, temperature field and acoustic emission field, etc.), not only the overall correlation coefficient and coupling factor are calculated, but also the dynamic change process in the time series needs to be recorded. First, the coupling relationship matrix is ​​expanded and constructed. For each pair of physical field parameters, its dynamic coupling coefficient needs to be calculated within a continuous time window, and the covariance and standard deviation are used to measure the interaction strength of the two physical fields within the time window. The value of each physical field parameter represents the measured value of the physical field at different time points. Then a three-dimensional coupling relationship matrix is ​​constructed, which contains the physical field type dimension and the time series dimension. Each element in the matrix represents the coupling coefficient of the corresponding physical field pair at a specific time point. In this way, the coupling relationship matrix contains information of the time dimension. For time domain feature extraction, the three-dimensional coupling relationship matrix is ​​subjected to wavelet transform in the time dimension.

[0067] For example, in coal mine face monitoring data processing, after physical field segmentation, the collected data is divided into stress field data including vertical stress (15-25 MPa) and horizontal stress (12-20 MPa). Gas field data includes gas pressure (0.74-2.35 MPa) and concentration (0.5%-1.8%). Temperature field data ranges from 24-29°C. Acoustic emission field data includes acoustic emission event frequency (10-50 events / hour). Principal component analysis (PCA) revealed that the first principal component (vertical stress variation) contributes 76% of the variance in the stress field, while the second principal component (horizontal stress variation) contributes 21%. Their cumulative contribution reaches 97%, so these two principal components are retained as key features. Correlation analysis revealed a significant correlation coefficient of 0.82 between vertical stress and gas pressure. Coupling calculations yielded a coupling factor of 0.78 between the stress and gas fields, indicating a strong coupling relationship between the two physical fields.

[0068] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0069] (1) Perform wavelet transform on the physical field coupling relationship matrix to obtain a time domain coefficient sequence, and perform statistical feature calculation on the time domain coefficient sequence to obtain a time domain feature set;

[0070] (2) Perform Fourier transform on the time domain feature set to obtain the frequency domain feature sequence, and calculate the frequency band energy of the frequency domain feature sequence to obtain the frequency band energy distribution;

[0071] (3) Extract the characteristic points of the frequency band energy distribution to obtain a characteristic point sequence, and perform dimension splicing on the characteristic point sequence to obtain a multi-dimensional feature vector;

[0072] (4) Perform correlation analysis on the multidimensional feature vectors to obtain a feature correlation matrix, and perform hierarchical clustering on the feature correlation matrix to obtain a feature cluster sequence;

[0073] (5) Calculate the information gain of the feature cluster sequence to obtain the feature importance sequence, and perform threshold screening on the feature importance sequence to obtain the candidate feature set;

[0074] (6) Perform orthogonal transformation on the candidate feature set to obtain feature basis vectors, and perform linear combination on the feature basis vectors to obtain the optimized feature parameter set.

[0075] Specifically, the wavelet transform can provide local characteristics of a signal at different time scales. By setting appropriate scale and translation parameters, the time series is decomposed using wavelet basis functions. The statistical characteristics of the wavelet coefficients, including mean, standard deviation, energy, and entropy, are then calculated. The mean reflects the overall level of coupling strength, the standard deviation indicates the degree of coupling strength fluctuation, the energy reflects signal strength, and the entropy indicates signal complexity. For frequency domain feature extraction, the wavelet transform coefficients are Fourier transformed to convert the time domain signal into the frequency domain. In the frequency domain, the frequency range is divided into several bands, and the energy distribution within each band is calculated. By analyzing the energy distribution characteristics in different frequency bands, the periodic variation of the coupling relationship can be determined. Finally, feature fusion is performed to combine the extracted time domain and frequency domain features to form a feature vector. When optimizing features, correlations between features are first calculated to remove redundant features. Dimensionality reduction is then performed using principal component analysis, and features whose cumulative contribution reaches a predetermined threshold are selected to form the final optimized feature set. By analyzing the time-varying characteristics of the coupling matrix, we not only preserve the coupling relationship between the physical fields but also capture the dynamic characteristics of this coupling relationship over time. For each physical field pair, the resulting eigenvector contains not only the coupling strength information between the physical fields but also the dynamic characteristics of the coupling relationship.

[0076] The physical field coupling relationship matrix is ​​processed by wavelet transform, and the time domain features are extracted using a multi-resolution analysis method. Wavelet transform is a time-frequency localization analysis method that can provide local feature information of the signal in both the time domain and the frequency domain. The calculation formula of wavelet transform is:

[0077] ;

[0078] in, is the wavelet coefficient, is the scale parameter (controls the expansion and contraction of the wavelet), b is the translation parameter (controls the translation of the wavelet), is the input signal, is the wavelet basis function. Calculate the statistical characteristics of the obtained wavelet coefficients:

[0079] ;

[0080] in, is the k-order statistical moment, n is the number of samples, is the wavelet coefficient, when k=1 it is the mean, when k=2 it is the variance, when k=3 it is the skewness, and when k=4 it is the kurtosis.

[0081] Perform Fourier transform on the time domain feature set to convert the time domain signal into frequency domain representation:

[0082] ;

[0083] in, is the frequency domain feature sequence, is the time domain signal, is the angular frequency, is an imaginary unit. Calculate the energy distribution in different frequency bands:

[0084] ;

[0085] in, is the energy of the ith frequency band, and The upper and lower limits of the frequency band.

[0086] It should be noted that in the embodiment of the present application, the data after wavelet transformation is the result of decomposing the original signal at different scales. For each scale level, a coefficient sequence that changes with time can be obtained. These coefficient sequences are essentially still time domain signals, but they represent the performance of the original signal at different scales (which can be understood as different frequency bands). When these coefficient sequences are Fourier transformed again, the frequency characteristics of the signal at each scale are actually analyzed. Because any time domain signal can be Fourier transformed to obtain its frequency characteristics. In this way, more detailed analysis results can be obtained: first, the wavelet transform helps to decompose the signal into different scale spaces, which is equivalent to frequency band division; then, the signal in each frequency band is Fourier transformed again to obtain more accurate frequency structure information within the frequency band. Such an analysis method provides a multi-level signal feature extraction approach, which not only retains the ability of the wavelet transform to characterize local features, but also obtains the periodic characteristics of the frequency domain through the Fourier transform.

[0087] Extract the characteristic points of the frequency band energy distribution, including the energy peak, mean and variance of each frequency band. All features are concatenated into vector form according to the dimension:

[0088] ;

[0089] in, is a multidimensional feature vector, is the i-th feature component. Calculate the correlation coefficient matrix between features:

[0090] ;

[0091] in, Features i and j The correlation coefficient of is the covariance, is the variance.

[0092] Among them, the feature extraction and analysis are based on the idea of ​​multi-level feature mining: the first stage of feature extraction is based on the original data of the physical field, and the coupling relationship matrix is ​​obtained by analyzing the interaction between different physical fields. This matrix reflects the intensity of the mutual influence between the physical fields; the second stage is to conduct a deep time-frequency feature analysis on this coupling relationship matrix, obtain the time domain features through wavelet transform, and then obtain the frequency domain features through Fourier transform, and extract key feature points from it. These feature points reflect the change pattern of the coupling relationship over time and frequency. This step-by-step analysis method makes it possible to understand the coupling mechanism between physical fields from different dimensions: first, understand the physical fields with coupling, and then further analyze how the coupling relationship evolves over time and manifests at different frequencies. Finally, through correlation analysis and clustering of these time-frequency features, the most representative feature combination is screened out to form an optimized feature parameter set.

[0093] The features are grouped using a hierarchical clustering method and the distances between the features are calculated:

[0094] ;

[0095] in, is the Euclidean distance between features p and q. Information gain is calculated for the feature cluster sequence to perform feature selection. Finally, feature basis vectors are obtained through orthogonal transformation and linearly combined to obtain the optimized feature parameter set.

[0096] For example, in coal mine working face monitoring, collected gas pressure data (sampling frequency 100 Hz) was analyzed. First, wavelet transforms were used to obtain wavelet coefficients at different scales. At scale 1, the mean was calculated to be 0.85 MPa, the variance was 0.12, the skewness was 0.34, and the kurtosis was 2.87. After Fourier transforms, the frequency range was divided into low-frequency bands (0-10 Hz), mid-frequency bands (10-50 Hz), and high-frequency bands (50-100 Hz). The energy contributions of each band were calculated to be 65%, 25%, and 10%, respectively. For each frequency band, the peak energy values ​​(0.82, 0.45, and 0.21, respectively), mean values ​​(0.56, 0.32, and 0.15), and variances (0.08, 0.05, and 0.03) were extracted to form an 18-dimensional feature vector. Through correlation analysis and hierarchical clustering, features with correlation coefficients greater than 0.85 were merged to obtain eight key feature parameters with strong independence, which formed the optimized feature parameter set.

[0097] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0098] (1) Perform sliding window processing on the optimized feature parameter set to obtain a time series segment sequence, and calculate the overlap of the time series segment sequence to obtain a set of segmentation points;

[0099] (2) Perform boundary optimization on the segmentation point set to obtain a time series segment sequence, and perform feature statistics on the time series segment sequence to obtain a sample feature matrix;

[0100] (3) Perform label assignment processing on the sample feature matrix to obtain the initial training samples, and perform balanced sampling processing on the initial training samples to obtain the training sample set;

[0101] (4) Perform data rotation processing on the training sample set to obtain an amplified sample sequence, and perform noise injection processing on the amplified sample sequence to obtain an enhanced data set;

[0102] (5) Perform maximum and minimum value normalization on the enhanced data set to obtain a standardized data sequence, and perform data verification on the standardized data sequence to obtain a model training data set;

[0103] (6) The parameter space of the model training data set is divided to obtain a parameter grid sequence, and the parameter grid sequence is processed by deep neural network to obtain a prominent disaster prediction model.

[0104] Specifically, the processing of the optimized feature parameter set first uses the sliding window technology to perform time series segmentation. The window size is set to 24 hours and the sliding step is 1 hour. The segmentation point is determined by calculating the data overlap between adjacent windows. The overlap calculation is based on the number of data points jointly covered by adjacent windows divided by the window size. The overlap threshold is set to 50%. When the overlap is lower than the threshold, it is determined as a segmentation point. The boundary of the obtained segmentation point set is optimized, and the boundary position is adjusted by detecting the degree of mutation of the data before and after the segmentation point. Statistical features are calculated for each time series segment, including statistical quantities such as mean, standard deviation, skewness, kurtosis, and the rate of change and trend characteristics of physical field parameters, and these features are organized into a sample feature matrix. Each row of the matrix represents the eigenvector of a time window, and each column represents a feature.

[0105] Based on historical records of significant events and expert annotations, labels are assigned to the sample feature matrix. These labels are categorized as "significant danger" and "significant safety." Samples within 24 hours of a significant event are labeled "significant danger," while all other samples are labeled "significant safety." Because "significant danger" samples are relatively rare in real-world data, SMOTE (Synthetic Minority Oversampling Technique) is used for balanced sampling. This method generates new samples by interpolating between minority class samples to achieve a balanced sample size between the two classes.

[0106] Data augmentation was performed on the balanced training sample set. First, a data rotation method was used to transform the eigenvectors by a small angle to generate new samples. The rotation angle range was set from -15° to 15°, with a step size of 5°. Noise was then injected into the rotated samples, adding Gaussian noise with a mean of 0 and a standard deviation of 0.01 to enhance the model's noise tolerance. The augmented dataset was normalized to its maximum and minimum values, mapping all eigenvalues ​​to the interval [0, 1] to eliminate dimensional differences between different features. Data validation was performed to ensure that the normalized data was reasonably distributed and free of outliers and missing values.

[0107] Finally, a deep neural network model was constructed using a multilayer perceptron architecture, consisting of an input layer (feature dimension), three hidden layers (128, 64, and 32 nodes, respectively), and an output layer (two nodes, corresponding to the two labels). The hidden layers used the ReLU activation function, and the output layer used the Softmax function. A grid search method was used to optimize model parameters, including hyperparameters such as the learning rate (range: 0.0001-0.1), batch size (range: 32-256), and dropout rate (range: 0.1-0.5).

[0108] For example, for outburst prediction at a coal mine working face, multi-physics monitoring data was collected over 30 consecutive days. Using sliding window processing, the data was segmented into 720 time windows (24 hours per window, with a step size of 1 hour). Fifteen segmentation points were calculated, with the lowest overlap occurring on the eighth day, at 45%. Twenty statistical features were extracted for each time window, forming a 720×20 feature matrix. Based on historical data, two outburst events occurred during this period. Forty-eight samples within the 24 hours prior to the event were labeled "outburst hazard," and the remaining 672 samples were labeled "outburst safe." The SMOTE method was used to generate 624 artificial "outburst hazard" samples, bringing the number of samples in each category to 672. After data augmentation, the total number of training samples reached 4032. The model was trained using a batch size of 128, a learning rate of 0.001, and a dropout rate of 0.3. The final model achieved 90% prediction accuracy on the test set.

[0109] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0110] (1) Perform kernel density estimation on the output data of the catastrophic disaster prediction model to obtain the probability density function, and then perform integral calculation on the probability density function to obtain the cumulative distribution sequence;

[0111] (2) Calculate the confidence level of the cumulative distribution sequence to obtain the confidence boundary sequence, and perform interval estimation on the confidence boundary sequence to obtain the confidence interval of the prediction result;

[0112] (3) Perform K-fold cross validation on the confidence interval of the prediction result to obtain a validation index sequence, and perform leave-one-out validation on the validation index sequence to obtain a validation result set;

[0113] (4) Perform stability analysis on the verification result set to obtain a stability index sequence, and perform consistency test on the stability index sequence to obtain a test result matrix;

[0114] (5) Perform comprehensive scoring on the test result matrix to obtain an evaluation score sequence, and perform weight distribution on the evaluation score sequence to obtain the prediction reliability evaluation result.

[0115] Specifically, the output data of the prominent disaster prediction model is first processed by kernel density estimation, and the probability density function of the predicted value is obtained by non-parametric estimation method. Kernel density estimation uses the Gaussian kernel function as the kernel function, and the bandwidth parameter is determined by cross-validation. The kernel density estimation calculation formula is:

[0116] ;

[0117] in, is the probability density function, is the predicted output value, n is the number of samples, B is the bandwidth parameter, and K is the Gaussian kernel function. The cumulative distribution function is obtained by numerically integrating the probability density function:

[0118]

[0119] in, is the cumulative distribution function, is the probability density function.

[0120] A confidence level calculation was performed on the cumulative distribution sequence, setting the confidence level to 95%. The upper and lower bounds were determined based on confidence interval theory. K-fold cross-validation was performed using a setting of K=5, randomly dividing the dataset into five subsets. Four subsets were used for model training at a time, and the remaining subset was used for validation. This cycle was repeated five times to obtain a complete set of validation metrics. Validation metrics included accuracy, precision, recall, and F1 score. Through leave-one-out cross-validation, each sample was used as a validation sample in turn, while the remaining samples were used for training, resulting in a more detailed set of validation results. A stability analysis was performed on the validation results to calculate the fluctuation of the prediction results across different validation processes. Stability metrics include statistical measures such as standard deviation and coefficient of variation. Cohen's Kappa coefficient was used as a consistency test to assess the consistency of prediction results across different validation processes. During the comprehensive scoring phase, metrics such as accuracy, stability, and consistency were weighted according to their importance to obtain the final prediction reliability assessment results.

[0121] For example, in the prediction of outbursts at a coal mine working face, 30 consecutive days of monitoring data were processed. The distribution of predicted values ​​obtained by kernel density estimation showed that 85% of the predicted values ​​fell within the range of 0.2-0.8, with a median of 0.45. At a 95% confidence level, the confidence interval was [0.15, 0.85]. Through 5-fold cross-validation, the average accuracy was 92.5%, with a standard deviation of 2.3%. The precision for each validation fold was 90.2%, 91.5%, 89.8%, 92.1%, and 91.8%, respectively, and the recall was 88.5%, 89.2%, 87.9%, 90.3%, and 89.6%, respectively. Leave-one-out validation results showed that 97.2% of the sample predictions were consistent with the original labels. Stability analysis showed that the coefficient of variation of the predictions across different validation processes was 0.068, and the Kappa coefficient was 0.856, indicating high stability and consistency of the predictions. Finally, the accuracy, stability, and consistency indicators were weighted according to a weight ratio of 0.4:0.3:0.3, and a comprehensive evaluation score of 0.895 was obtained, indicating that the prediction model has high reliability.

[0122] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0123] (1) Calculate the distribution characteristics of the prediction reliability assessment results to obtain a characteristic statistical sequence, and perform cluster analysis on the characteristic statistical sequence to obtain a graded classification sequence;

[0124] (2) Calculate the threshold interval of the level classification sequence to obtain the risk level boundary, and perform fuzzy classification processing on the risk level boundary to obtain the warning level judgment result;

[0125] (3) Perform time series analysis on the warning level determination results to obtain a time evolution sequence, and perform spatial interpolation on the time evolution sequence to obtain a spatial distribution matrix;

[0126] (4) Perform regional correlation analysis on the spatial distribution matrix to obtain a correlation feature sequence, and perform trend prediction on the correlation feature sequence to obtain an evolution trend sequence;

[0127] (5) Perform risk propagation analysis on the evolution trend sequence to obtain the propagation path sequence, and perform decision optimization on the propagation path sequence to obtain target warning decision information.

[0128] Specifically, the prediction reliability assessment results are first subjected to distributional calculations. By calculating statistical characteristics such as the distribution density, skewness, and dispersion of the assessment values, a characteristic statistical sequence is generated. The characteristic statistical sequence is analyzed using the K-means clustering algorithm, with the number of clusters K set to 3. Data points are classified into three levels: high risk, medium risk, and low risk, forming a grading sequence. Cluster centers are selected based on the natural grouping characteristics of the data distribution, and classification is performed by calculating the Euclidean distance from the sample points to the cluster center. Threshold intervals are calculated for the grading sequence, and risk level boundaries are determined using a method based on probability density functions. By analyzing the probability distribution characteristics of the data at each level and incorporating expert experience, threshold intervals are set: the high-risk threshold is set at the 90th percentile, and the low-risk threshold is set at the 10th percentile. Fuzzy classification methods are used to process boundary values, and membership functions are introduced to describe the degree to which data points belong to different risk levels, resulting in more reasonable warning level determination results.

[0129] Time series analysis was conducted on the warning level determination results, using a time series decomposition method to extract trend, cyclic, and random terms. Autocorrelation and cross-correlation analyses were used to identify the changing patterns of the time series, resulting in a time evolution sequence. Kriging spatial interpolation was used to extend the warning level information from discrete monitoring points to the entire monitoring area, generating a continuous spatial distribution matrix. The interpolation process accounted for the correlation of spatial locations and the weights of each monitoring point. Regional correlation analysis was conducted on the spatial distribution matrix to calculate the risk propagation relationship between different regions. The spatial autocorrelation index was calculated to assess the degree of spatial risk concentration and identify the spatial distribution characteristics of high-risk areas. Based on the changing patterns of risk levels in adjacent regions, a correlation feature sequence was extracted, and the risk evolution trend was predicted using a time series forecasting method. Forecasting was performed using a sliding time window method with a window size of 24 hours and a prediction step size of 1 hour.

[0130] Risk propagation analysis is conducted on the evolutionary trend sequence, and the risk diffusion path and speed are calculated based on a risk propagation network model. By analyzing the propagation characteristics of historically significant events, a risk propagation equation is established to calculate the probability and intensity of risk spreading from high-risk areas to surrounding areas. Finally, combining risk levels, propagation paths, and evolutionary trends, regional early warning decision-making information is generated.

[0131] For example, in a coal mine face outburst prediction study, the reliability assessment results of the predictions were analyzed over seven consecutive days. K-means clustering was used to categorize the assessment values ​​into three categories: high risk (assessment value > 0.85), medium risk (0.45-0.85), and low risk (< 0.45). Of the 15 monitoring points located within the face, three had assessment values ​​exceeding 0.85. Spatial interpolation analysis revealed that these high-risk points were primarily concentrated near the fault zone within the face. Time series analysis revealed a gradual increase in risk levels in high-risk areas, with a 24-hour risk growth rate of 15%. Risk propagation analysis indicated that the risk propagated from the fault zone to its flanks at a rate of 2 meters per hour, and the impact was expected to expand to adjacent areas within 48 hours. Based on these analysis results, regional warning decision information was generated: a red alert was issued for areas near the fault zone, requiring immediate production halts and evacuations; a yellow alert was issued for adjacent areas, intensifying monitoring; and normal monitoring was maintained in other areas.

[0132] The above describes the coal and gas outburst disaster prediction method based on multi-physics field coupling in the embodiment of the present application. The following describes the coal and gas outburst disaster prediction device based on multi-physics field coupling in the embodiment of the present application. Figure 2 In one embodiment of the present application, a device for predicting coal and gas outburst disasters based on multi-physics field coupling includes:

[0133] The acquisition module 201 is used to collect and standardize the original data of the stress field, gas field, temperature field and acoustic emission field at the coal mine site to obtain a multi-physics field initial data set, and perform outlier detection and data filling processing on the multi-physics field initial data set to obtain a pre-processed multi-physics field parameter data set;

[0134] A decomposition module 202 is configured to perform physical field component extraction and eigendecomposition on the preprocessed multi-physics field parameter data set to obtain a basic parameter matrix for each physical field, and perform cross-correlation analysis and coupling factor calculation on the basic parameter matrix to obtain a physical field coupling relationship matrix;

[0135] An extraction module 203 is configured to extract time domain features and frequency domain features from the physical field coupling relationship matrix to obtain a multidimensional feature vector, and perform redundant feature elimination and feature reconstruction on the multidimensional feature vector to obtain an optimized feature parameter set;

[0136] The processing module 204 is used to perform time series segmentation and sample labeling on the optimized feature parameter set to obtain a training sample set, perform data enhancement and normalization on the training sample set to obtain a model training data set, and perform pattern recognition and parameter optimization on the model training data set to obtain a prominent disaster prediction model;

[0137] The calculation module 205 is used to perform probability distribution calculation on the output data of the sudden disaster prediction model to obtain a confidence interval of the prediction result, and perform multi-dimensional cross-validation on the confidence interval to obtain a prediction reliability assessment result;

[0138] The classification module 206 is used to perform threshold classification and risk level classification on the prediction reliability evaluation result to obtain a warning level determination result, and perform spatiotemporal correlation analysis on the warning level determination result to obtain target warning decision information.

[0139] Through the collaborative efforts of these components, multi-physics field data collection and preprocessing of stress, gas, temperature, and acoustic emission fields at coal mine sites effectively eliminates noise and outliers from the raw data, improving data quality. Physical field component extraction and feature decomposition provide in-depth analysis of the fundamental characteristics of each physical field. Cross-correlation analysis and coupling factor calculation comprehensively reveal the coupling relationships between multiple physical fields, providing more comprehensive feature information for outburst disaster prediction. Time-domain and frequency-domain feature extraction of the physical field coupling relationship matrix, combined with redundant feature elimination and feature reconstruction techniques, yields an optimized feature parameter set, enhancing both feature representativeness and effectiveness. Time series segmentation and sample labeling establish a standardized training sample set. Data augmentation and normalization methods enhance sample representativeness and model generalization. A deep neural network-based outburst disaster prediction model achieves accurate predictions of outburst risks. The reliability of the prediction results is assessed through probability distribution calculation and multi-dimensional cross-validation, ensuring their credibility. Finally, through threshold classification and risk level categorization, combined with spatiotemporal correlation analysis, scientific early warning decision-making information was generated, providing important decision-making support for coal mine safety production. This method overcomes the incomplete information provided by traditional single-physical field monitoring methods, improves the accuracy and reliability of predictions, and achieves precise early warning of coal and gas outburst disasters.

[0140] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0141] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A coal and gas outburst disaster prediction method based on multi-physics field coupling, characterized in that: The coal and gas outburst disaster prediction method based on multi-physics field coupling includes: The original data of the stress field, gas field, temperature field and acoustic emission field at the coal mine site are collected and normalized to obtain a multi-physics field initial data set, and outlier detection and data completion processing are performed on the multi-physics field initial data set to obtain a preprocessed multi-physics field parameter data set; Performing physical field component extraction and characteristic decomposition on the preprocessed multi-physical field parameter data set to obtain a basic parameter matrix of each physical field, and performing cross-correlation analysis and coupling factor calculation on the basic parameter matrix to obtain a physical field coupling relationship matrix; Extracting time domain features and frequency domain features from the physical field coupling relationship matrix to obtain a multidimensional feature vector, and performing redundant feature elimination and feature reconstruction on the multidimensional feature vector to obtain an optimized feature parameter set; Performing time series segmentation and sample labeling on the optimized feature parameter set to obtain a training sample set, performing data enhancement and normalization processing on the training sample set to obtain a model training data set, and performing pattern recognition and parameter optimization on the model training data set to obtain a prominent disaster prediction model; Performing probability distribution calculation on the output data of the prominent disaster prediction model to obtain a confidence interval of the prediction result, and performing multi-dimensional cross-validation on the confidence interval to obtain a prediction reliability assessment result; The prediction reliability assessment result is subjected to threshold classification and risk level division to obtain a warning level determination result, and the warning level determination result is subjected to spatiotemporal correlation analysis to obtain target warning decision information.

2. The method for predicting coal and gas outburst disasters based on multi-physics field coupling according to claim 1 is characterized in that: The original data of the stress field, gas field, temperature field and acoustic emission field at the coal mine site are collected and normalized to obtain a multi-physics field initial data set, and outlier detection and data completion processing are performed on the multi-physics field initial data set to obtain a pre-processed multi-physics field parameter data set, including: The original data of stress field, gas field, temperature field and acoustic emission field at the coal mine site are collected and standardized to obtain the initial data set of multi-physics field Performing peak normalization processing on the multi-physics field initial data set to obtain a normalized data set, and performing median filling on the normalized data set to obtain supplemented initial data; Performing sliding window segmentation processing on the supplemented initial data to obtain a data segment sequence, and performing interquartile range anomaly detection on the data segment sequence to obtain a labeled data sequence; Performing linear interpolation processing on the outliers in the marked data sequence to obtain an interpolated data sequence, and performing wavelet denoising processing on the interpolated data sequence to obtain a reduced-noise data sequence; Performing time window synchronization processing on the noise reduction data sequence to obtain an aligned data sequence, and performing data resampling processing on the aligned data sequence to obtain an equally spaced data sequence; Performing Z-score normalization processing on the equally spaced data sequence to obtain a normalized data sequence, and performing batch verification processing on the normalized data sequence to obtain a multi-physics field initial data set; Performing 3σ criterion anomaly detection on the multi-physics field initial data set to obtain an abnormal point marker sequence, and performing spline interpolation processing on the abnormal point marker sequence to obtain a data completion sequence; Performing moving average filtering on the data padding sequence to obtain a smoothed data sequence, and performing data quality assessment on the smoothed data sequence to obtain a data quality index; Performing threshold screening processing on the data quality indicators to obtain a qualified data sequence, and performing outlier detection on the qualified data sequence to obtain a detection result sequence; The detection result sequence is subjected to data verification processing to obtain a verified data sequence, and the verified data sequence is subjected to data format unification processing to obtain a preprocessed multi-physics field parameter data set.

3. The method for predicting coal and gas outburst disasters based on multi-physics field coupling according to claim 1 is characterized in that: The physical field component extraction and characteristic decomposition are performed on the preprocessed multi-physical field parameter data set to obtain a basic parameter matrix of each physical field, and the cross-correlation analysis and coupling factor calculation are performed on the basic parameter matrix to obtain a physical field coupling relationship matrix, including: Performing physical field type classification processing on the preprocessed multi-physics field parameter data set to obtain a stress field data subset, a gas field data subset, a temperature field data subset, and an acoustic emission field data subset, and performing time series sorting processing on each data subset to obtain an ordered data sequence; Performing principal component decomposition on the ordered data sequence to obtain main feature components, and calculating variance contribution rates on the main feature components to obtain a feature weight sequence; Performing threshold screening processing on the feature weight sequence to obtain a key feature sequence, and performing physical quantity unit unification processing on the key feature sequence to obtain a standardized feature sequence; Performing matrix reconstruction processing on the standardized characteristic sequence to obtain a basic parameter matrix of each physical field, and performing data integrity verification on the basic parameter matrix to obtain a verified parameter matrix; Calculating the Pearson correlation coefficient on the verified parameter matrix to obtain a correlation coefficient matrix, and performing a significance test on the correlation coefficient matrix to obtain effective correlation items; performing coupling calculation on the effective correlation items to obtain a coupling factor sequence, and performing normalization processing on the coupling factor sequence to obtain a standardized coupling factor; Matrix mapping processing is performed on the standardized coupling factors to obtain an initial coupling relationship matrix, and symmetry testing is performed on the initial coupling relationship matrix to obtain a physical field coupling relationship matrix.

4. The method for predicting coal and gas outburst disasters based on multi-physics field coupling according to claim 1, characterized in that: The extracting time domain features and frequency domain features of the physical field coupling relationship matrix to obtain a multidimensional feature vector, and performing redundant feature elimination and feature reconstruction on the multidimensional feature vector to obtain an optimized feature parameter set includes: Performing wavelet transform processing on the physical field coupling relationship matrix to obtain a time domain coefficient sequence, and performing statistical feature calculation on the time domain coefficient sequence to obtain a time domain feature set; Performing Fourier transform processing on the time domain feature set to obtain a frequency domain feature sequence, and performing frequency band energy calculation on the frequency domain feature sequence to obtain a frequency band energy distribution; Performing feature point extraction processing on the frequency band energy distribution to obtain a feature point sequence, and performing dimension splicing processing on the feature point sequence to obtain a multi-dimensional feature vector; Performing correlation analysis on the multidimensional feature vector to obtain a feature correlation matrix, and performing hierarchical clustering on the feature correlation matrix to obtain a feature cluster sequence; Performing information gain calculation on the feature cluster sequence to obtain a feature importance sequence, and performing threshold screening processing on the feature importance sequence to obtain a candidate feature set; An orthogonal transformation process is performed on the candidate feature set to obtain feature basis vectors, and a linear combination process is performed on the feature basis vectors to obtain an optimized feature parameter set.

5. The method for predicting coal and gas outburst disasters based on multi-physics field coupling according to claim 1, characterized in that: The optimized feature parameter set is subjected to time series segmentation and sample labeling to obtain a training sample set, and the training sample set is subjected to data enhancement and normalization processing to obtain a model training data set, and the model training data set is subjected to pattern recognition and parameter optimization to obtain a prominent disaster prediction model, including: Performing sliding window processing on the optimized feature parameter set to obtain a time series segment sequence, and performing overlap calculation on the time series segment sequence to obtain a segmentation point set; Performing boundary optimization processing on the segmentation point set to obtain a time series segment sequence, and performing feature statistical processing on the time series segment sequence to obtain a sample feature matrix; Performing label assignment processing on the sample feature matrix to obtain initial training samples, and performing balanced sampling processing on the initial training samples to obtain a training sample set; Performing data rotation processing on the training sample set to obtain an amplified sample sequence, and performing noise injection processing on the amplified sample sequence to obtain an enhanced data set; Performing maximum and minimum value normalization processing on the enhanced data set to obtain a standardized data sequence, and performing data verification processing on the standardized data sequence to obtain a model training data set; The model training data set is processed by parameter space division to obtain a parameter grid sequence, and the parameter grid sequence is processed by deep neural network to obtain a prominent disaster prediction model.

6. The method for predicting coal and gas outburst disasters based on multi-physics field coupling according to claim 1, characterized in that: The probability distribution calculation is performed on the output data of the prominent disaster prediction model to obtain the confidence interval of the prediction result, and the confidence interval is cross-validated in multiple dimensions to obtain the prediction reliability evaluation result, including: Performing kernel density estimation processing on the output data of the sudden disaster prediction model to obtain a probability density function, and performing integral calculation on the probability density function to obtain a cumulative distribution sequence; Calculating the confidence level of the cumulative distribution sequence to obtain a confidence boundary sequence, and performing interval estimation processing on the confidence boundary sequence to obtain a confidence interval of the prediction result; Performing K-fold cross validation on the confidence interval of the prediction result to obtain a validation index sequence, and performing leave-one-out validation on the validation index sequence to obtain a validation result set; Performing stability analysis on the verification result set to obtain a stability index sequence, and performing consistency check on the stability index sequence to obtain a check result matrix; A comprehensive scoring process is performed on the test result matrix to obtain an evaluation score sequence, and a weight distribution process is performed on the evaluation score sequence to obtain a prediction reliability evaluation result.

7. The method for predicting coal and gas outburst disasters based on multi-physics field coupling according to claim 1, characterized in that: The threshold classification and risk level classification of the prediction reliability assessment result are performed to obtain a warning level determination result, and the spatiotemporal correlation analysis of the warning level determination result is performed to obtain target warning decision information, including: Calculating distribution characteristics of the prediction reliability evaluation results to obtain a characteristic statistical sequence, and performing cluster analysis on the characteristic statistical sequence to obtain a graded classification sequence; Performing threshold interval calculation on the level division sequence to obtain risk level boundaries, and performing fuzzy classification processing on the risk level boundaries to obtain warning level determination results; Performing time series analysis on the warning level determination result to obtain a time evolution sequence, and performing spatial interpolation processing on the time evolution sequence to obtain a spatial distribution matrix; Performing regional correlation analysis on the spatial distribution matrix to obtain a correlation feature sequence, and performing trend prediction on the correlation feature sequence to obtain an evolution trend sequence; The evolution trend sequence is subjected to risk propagation analysis processing to obtain a propagation path sequence, and the propagation path sequence is subjected to decision optimization processing to obtain target early warning decision information.

8. A coal and gas outburst disaster prediction device based on multi-physics field coupling, used to implement the coal and gas outburst disaster prediction method based on multi-physics field coupling as described in any one of claims 1 to 7, characterized in that: The coal and gas outburst disaster prediction device based on multi-physics field coupling includes: An acquisition module is used to collect and standardize the raw data of the stress field, gas field, temperature field, and acoustic emission field at the coal mine site to obtain an initial multi-physics field data set, and perform outlier detection and data completion processing on the initial multi-physics field data set to obtain a pre-processed multi-physics field parameter data set; A decomposition module is used to perform physical field component extraction and feature decomposition on the preprocessed multi-physical field parameter data set to obtain a basic parameter matrix of each physical field, and to perform cross-correlation analysis and coupling factor calculation on the basic parameter matrix to obtain a physical field coupling relationship matrix; An extraction module is used to extract time domain features and frequency domain features of the physical field coupling relationship matrix to obtain a multidimensional feature vector, and to eliminate redundant features and reconstruct features of the multidimensional feature vector to obtain an optimized feature parameter set; a processing module, configured to perform time series segmentation and sample labeling on the optimized feature parameter set to obtain a training sample set, perform data enhancement and normalization on the training sample set to obtain a model training data set, perform pattern recognition and parameter optimization on the model training data set to obtain a prominent disaster prediction model; A calculation module is used to perform probability distribution calculation on the output data of the sudden disaster prediction model to obtain a confidence interval of the prediction result, and to perform multi-dimensional cross-validation on the confidence interval to obtain a prediction reliability assessment result; The classification module is used to perform threshold classification and risk level classification on the prediction reliability assessment result to obtain the warning level determination result, and perform spatiotemporal correlation analysis on the warning level determination result to obtain target warning decision information.

Images