A coal and gas outburst early warning method and system based on multi-source information fusion

Abstract

The application discloses a coal and gas outburst early warning method and system based on multi-source information fusion, relates to the technical field of coal mine safety monitoring, and comprises the following steps: collecting multi-source monitoring data; filtering noise and reconstructing AE signals and EMR signals; combining FBG strain signals and CH4 concentration signals, carrying out time domain, frequency domain and time-frequency domain characteristic analysis on the multi-source monitoring data, extracting precursor information, and constructing a multi-dimensional characteristic expression; inputting the obtained two-dimensional time-frequency images of the multi-source monitoring data, corresponding time domain characteristic parameters and frequency domain characteristic parameters into a single-source precursor characteristic intelligent identification model, identifying the outburst precursor state corresponding to each single monitoring source; and adopting a multi-source information fusion method based on dynamic weighted Bayesian inference as a multi-source information fusion decision core model to output the identification results of different monitoring sources. The application can realize stable and reliable early warning in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety monitoring technology, and in particular to a method and system for early warning of coal and gas outbursts based on multi-source information fusion. Background Technology

[0002] Coal and gas outbursts are among the most common and dangerous dynamic hazards in underground coal mining. Their occurrence mechanism is complex, typically involving the coupled effects of multiple factors such as ground stress, coal seam structural characteristics, gas occurrence state, and mining disturbance. They are characterized by their sudden onset and high destructiveness, seriously threatening the safety of underground personnel and continuous mine production. Current early warning systems for coal and gas outburst risks mainly rely on two types of technologies: static prediction and dynamic monitoring. Static prediction methods are based on indicators such as regional geology, coal seam gas pressure, initial gas emission velocity, coal seam firmness coefficient, and drill cuttings volume. At the mining face, sampling methods such as borehole testing are often used, essentially a "point-based, phased" detection mode. While these methods have some operability, they are difficult to implement for continuous, real-time monitoring during face excavation, insufficiently reflect dynamic precursor information during the outburst's development and evolution, and are easily affected by coal and rock heterogeneity and operational procedures, resulting in delayed early warning results and a high risk of missed or false alarms.

[0003] With the development of intelligent and information technologies in coal mines, dynamic monitoring technologies such as acoustic emission, electromagnetic radiation, fiber optic strain, and online gas concentration monitoring have begun to be applied in coal and gas outburst early warning research. Among these, acoustic emission signals can reflect the elastic wave release characteristics during the loading and fracturing process of coal and rock mass; electromagnetic radiation signals can reflect charge separation and radiation behavior during the fracturing process; fiber optic strain signals can reflect the stress-strain evolution law of coal and surrounding rock; and gas concentration signals can reflect the desorption, migration, and abnormal outburst state of gas within the coal mass. These multiple signals characterize the outburst incubation and evolution process from different perspectives, possessing real-time and continuous characteristics, thus providing the possibility to improve the dynamics and predictability of outburst prediction.

[0004] However, existing dynamic monitoring technologies still have significant shortcomings in application. The complex underground environment, with factors such as mechanical vibration, ventilation disturbances, electromagnetic interference, and human-caused noise, can significantly impact monitoring signals like AE and EMR. This results in raw monitoring data containing a high degree of noise, exhibiting strong non-stationarity and low signal-to-noise ratio characteristics, making it difficult to extract effective precursor features. Most studies still focus on single-signal analysis or rely solely on a single monitoring indicator for early warning. Thresholds are often determined empirically or based on local statistical patterns. There is a lack of effective mechanisms for unified modeling and fusion decision-making of heterogeneous information such as sound, light, electricity, and gas. The system's intelligent identification and uncertainty handling capabilities for multi-source data are insufficient, leading to inadequate adaptability and robustness under different geological and mining conditions, and limited reliability of early warnings.

[0005] From the perspective of disaster mechanism, the formation of coal and gas outburst disasters is accompanied by the synchronous evolution of multiple physical processes, such as coal and rock fracturing, dynamic stress adjustment, and gas desorption and migration. A single monitoring method cannot fully reflect this comprehensive process. Therefore, integrating multi-source monitoring information, including acoustic, optical, electrical, and gas sources, for comprehensive early warning has become an important research direction. However, existing methods still face shortcomings in practical applications: First, multi-source monitoring data differ in sampling frequency, dimensions, and time series, constituting heterogeneous information that is difficult to process collaboratively; second, downhole data is prone to missing and anomalies, and the reliability of each information source changes dynamically with operating conditions, making it difficult for traditional fixed-weight fusion methods to achieve adaptive evaluation; third, in strong noise environments, precursor signals are weak and exhibit significant nonlinear characteristics. Relying on manual experience to extract features or set thresholds makes it difficult to fully uncover effective early warning patterns, thus limiting the accuracy and predictability of early warnings.

[0006] Therefore, there is an urgent need for a coal and gas outburst early warning method and system based on multi-source information fusion that can solve the above problems. Summary of the Invention

[0007] This solution addresses the problems and needs raised above by proposing a coal and gas outburst early warning method and system based on multi-source information fusion. The above technical objectives are achieved by adopting the following technical features, and other technical effects are also brought about.

[0008] One objective of this invention is to propose a coal and gas outburst early warning method based on multi-source information fusion, comprising the following steps: S10: Synchronous acquisition of multi-source monitoring data: Multi-source monitoring sensor units are deployed at the coal mine tunneling face to synchronously acquire multi-source monitoring data during the outburst incubation process; among which, the multi-source monitoring data includes: AE signal, EMR signal, FBG strain signal and CH4 concentration signal; S20: Joint noise filtering and reconstruction of AE and EMR signals: The AE and EMR signals are processed by a joint noise filtering and reconstruction method combining singular value decomposition (SVD) and ensemble empirical mode decomposition (EEMD) to obtain the reconstructed signal. S30: Construction of time-frequency features of multi-source signals: Combining the FBG strain signal and CH4 concentration signal obtained by synchronous acquisition, time-domain, frequency-domain and time-frequency-domain feature analysis are carried out on the multi-source monitoring data respectively to extract the precursor information that can characterize the damage evolution, stress adjustment and gas anomaly changes of coal and rock mass, and construct a multi-dimensional feature expression. S40: Intelligent recognition of single-source precursor features: The obtained AE two-dimensional time-frequency image, EMR two-dimensional time-frequency image, FBG two-dimensional time-frequency image and CH4 two-dimensional time-frequency image are input together with the corresponding time domain feature parameters and frequency domain feature parameters into the intelligent recognition model of single-source precursor features to identify the prominent precursor status corresponding to each single monitoring source. S50: Multi-source information fusion early warning: After obtaining the single-source precursor identification results and corresponding probabilities of the four types of monitoring sources, namely AE signal, EMR signal, FBG strain signal and CH4 concentration signal, a multi-source information fusion method based on dynamic weighted Bayesian inference is further adopted as the core model for multi-source information fusion decision-making. The identification results of different monitoring sources are uniformly expressed, dynamically weighted, fused, inferred and output as decisions to achieve comprehensive identification of coal and gas outburst danger states.

[0009] Furthermore, the coal and gas outburst early warning method and system based on multi-source information fusion according to the present invention may also have the following technical features: In one example of the present invention, step S20 specifically includes the following steps: S21: For any original signal sequence Let n be the number of sampling points in the current sliding analysis window. Construct its corresponding Hankel matrix according to the phase space reconstruction idea:

[0010] Where m + c - 1 = n; then, singular value decomposition is performed on the Hankel matrix to obtain:

[0011] Where U is the left singular vector matrix, V is the right singular vector matrix, and Σ is the singular value diagonal matrix. And satisfy ; S22: Set a threshold based on the singular value mutation characteristics or cumulative energy ratio, determine the effective singular value retention order based on the threshold, and truncate the singular value matrix to obtain the preliminary noise filtering matrix. Then, the initial noise filtering matrix The initial noise-filtered signal is obtained by performing anti-diagonal averaging reconstruction. ; S23: Initial noise filtering signal Performing ensemble empirical mode decomposition yields p intrinsic mode function (IMF) components and residual terms, expressed as follows:

[0012] in, The i-th intrinsic mode component Residual terms; S24: Quantitative screening of each IMF component is performed using the variance contribution rate. For the i-th intrinsic mode component, its variance contribution rate... Defined as:

[0013] in, Let represent the variance of the i-th IMF component; based on this, calculate the cumulative variance contribution rate of the q IMF components. :

[0014] S25: Select the one that satisfies The first q IMF components are used to reconstruct the signal, thereby obtaining the acoustic emission reconstructed signal. and electromagnetic radiation reconstructed signal Where θ is a preset cumulative contribution rate threshold, and the final reconstructed signal is... Represented as: .

[0015] In one example of the present invention, step S30 specifically includes the following steps: S31: Construct time-domain features for multi-source signals, including energy value, peak amplitude, root mean square value, strain change, and gas concentration change rate; where, for discrete signal sequence x i Its energy value E is expressed as:

[0016] Where n is the number of sampling points in the current analysis window; For the gas concentration sequence C(t), the rate of change of gas concentration is further calculated, and its expression is:

[0017] Where Δt is the time interval between adjacent sampling times; By extracting time-domain parameters, statistical characteristics of amplitude changes, energy accumulation, and rate of change of various signals within the current time window are obtained. S32: Construct frequency domain features for multi-source signals, perform spectral analysis on the reconstructed AE signal, EMR signal, as well as the FBG strain signal and CH4 concentration signal after unified time window processing, and extract frequency domain feature parameters including peak frequency, center frequency, and centroid frequency to characterize the distribution characteristics of signal frequency components and the variation law of dominant frequency band. If the discrete spectrum is P(f) i If f is the centroid frequency, then f g Represented as:

[0018] Among them, f i For the i-th discrete frequency point, P(f i () represents the spectral value at the corresponding frequency point; S33: Based on this, further construct time-frequency domain features; use wavelet decomposition and Hilbert-Huang transform methods to perform joint time-frequency analysis on various signals, constructing a two-dimensional time-frequency feature image; during wavelet decomposition, decompose the signal into multiple frequency bands, and calculate the energy proportion of each frequency band as time-frequency feature parameters; by calculating the energy proportion of different frequency bands, the distribution characteristics of signal energy at various scales can be effectively characterized, and a two-dimensional time-frequency feature representation with discriminative significance can be constructed accordingly; where, if the energy of the j-th frequency band is E j If the total number of frequency bands is L, then the energy percentage ρ of the j-th frequency band is... j Represented as:

[0019] In the formula, E l Energy per frequency band; S34: Finally, the two-dimensional time-frequency images I of AE are obtained respectively. AE EMR two-dimensional time-frequency image I EMR FBG two-dimensional time-frequency image I FBG and CH4 two-dimensional time-frequency image I CH4 Simultaneously, by combining the time-domain and frequency-domain characteristic parameters corresponding to various signals, a structured feature set of multi-source signals is formed.

[0020] In one example of the present invention, step S40 specifically includes the following steps: S41: Preprocess the two-dimensional time-frequency images of each single source to ensure that the two-dimensional time-frequency images constructed from different monitoring sources have a unified data expression form before being input into the recognition model; S42: Extract directional gradient histogram features and local binary pattern texture features from each single-source two-dimensional time-frequency image to characterize energy distribution edges, local texture changes, frequency band clustering features, and abnormal evolution patterns in the image; at the same time, incorporate the extracted time-domain feature parameters and frequency-domain feature parameters into the recognition input to form a single-source comprehensive feature that combines image representation capabilities and statistical representation capabilities. S43: For any single-source monitoring source k, where, Its corresponding image descriptor feature vector and statistical eigenvectors They are respectively recorded as:

[0021]

[0022] The image descriptor feature vector With statistical eigenvectors The features are concatenated to form a single-source comprehensive feature vector. Among them, Z k This represents the fusion feature vector of single-source monitoring information; S44: Perform principal component analysis to reduce the dimensionality of the comprehensive feature vector. Based on this, input the dimensionality-reduced feature vector into a support vector machine classifier to identify single-source precursor states. Let the dimensionality reduction transformation matrix be P, then the dimensionality-reduced feature vector is represented as:

[0023] in, The input feature vector is used for the dimensionality-reduced single-source recognition.

[0024] In one example of the present invention, in step S44, the dimensionality-reduced feature vector is input into a support vector machine classifier to identify single-source precursor states, specifically including the following steps: S441: The support vector machine uses a radial basis kernel function and employs a one-to-many strategy to construct a three-class classification recognition model; to achieve hierarchical recognition of prominent hazardous states, the single-source recognition state set is... Set as:

[0025] Where w1 represents a safe state, w2 represents a threat state, and w3 represents a dangerous state; For any state category Its support vector machine discriminant function is expressed as:

[0026] in, Indicates single-source monitoring information regarding status The output value of the discriminant function; N is the number of support vectors; For Lagrange multipliers; Labels for training samples; For kernel functions; For bias terms; Let be the feature vector of the i-th training sample obtained during the training process; S442: Convert the discriminant values ​​output by the support vector machine into probabilistic form; for single-source monitoring information belonging to the state... The recognition probability is expressed as:

[0027] in, This indicates that the single-source monitoring information belongs to the status category. The recognition probability; l is the summation index; S443: Based on the probability values ​​corresponding to each state category, determine the optimal identification result of the current single-source monitoring information, and obtain the single-source precursor identification results and corresponding probability outputs for the four monitoring sources: AE signal, EMR signal, FBG strain signal, and CH4 concentration signal. This realizes the transformation from multi-source time-frequency feature construction to intelligent identification of single-source states. The optimal identification result... The expression is: .

[0028] In one example of the present invention, step S50 specifically includes the following steps: S51: Construct a multi-source fusion identification framework; the early warning status identification framework is as follows: Where w1 represents a safe state, w2 represents a threat state, and w3 represents a dangerous state; S52: Construct dynamic quality evaluation indicators for any monitoring source k The dynamic quality evaluation index comprehensively considers signal quality, identification confidence, and data integrity. S53: Normalize the dynamic quality evaluation indicators of each monitoring source to obtain the dynamic fusion weight of each single-source monitoring information at the current moment. The expression is as follows:

[0029] in, q represents the fusion weight of a single monitoring source at time t; r (t) represents the dynamic quality evaluation index of the r-th monitoring source at time t; S54: The recognition probability output from each single-source monitoring information With dynamic weights By combining these factors, a weighted likelihood term is constructed for each monitoring source with respect to different early warning states. Its expression is:

[0030] Where ε is a positive constant, and its value ranges from 1 × 10⁻⁶. -4 ≤ ε ≤ 1×10 -3 ; Based on this, the prior probability π of each warning state is introduced. j By combining the weighted likelihood terms of the four monitoring sources with Bayesian fusion, the posterior probabilities of the system being in each warning state at the current moment are obtained:

[0031] in, This represents the set of multi-source monitoring observations corresponding to time t; This indicates that the system is in state w after fusion. j The posterior probability; S55: Introduces a time-recursive smoothing mechanism to smoothly update the fusion results of adjacent time steps. Its expression is:

[0032] in, This represents the fusion probability after smoothing at time t; It is a smoothing coefficient, and satisfies 0 < γ ≤ 1; Based on the smoothed fusion probability, the state corresponding to the largest probability is taken as the comprehensive warning state at the current moment. This enables the transformation from multi-source single-source identification results to comprehensive state identification results, including comprehensive early warning state. The expression is:

[0033] S56: Constructing a comprehensive early warning index R t Through the comprehensive early warning index R t The discrete classification results are further transformed into continuous risk characterization quantities to reflect the overall changing trend and phased evolution characteristics of the current prominent hazards at the working face. Among them, the comprehensive early warning index R t The expression is:

[0034] Where j is the warning status category index, j=1,2,3, corresponding to safe status w1, threat status w2 and dangerous status w3 respectively; λ1, λ2, λ3 are the risk mapping coefficients corresponding to safe, threat and dangerous statuses respectively, and satisfy 1≥λ3>λ2>λ1≥0.

[0035] In one example of the present invention, the method further includes: Step S60: Warning output and result display: After completing the multi-source information fusion warning, based on the current comprehensive warning status and the comprehensive warning index R... t The judgment result triggers the corresponding level of early warning output; at the same time, the software platform of the ground monitoring center displays and records the early warning result in real time.

[0036] In one example of the present invention, in step S60, based on the current comprehensive early warning status and the comprehensive early warning index R... t The judgment result triggers the corresponding level of early warning output, specifically including the following: When the comprehensive early warning index R tWhen the judgment range corresponding to the corresponding warning level is reached, and the fusion decision result indicates that the current working face is in a threatened or dangerous state, the corresponding alarm signal and interface prompt information are generated according to the pre-set warning strategy. Among them, the low-level warning is used to indicate that there is a prominent evolution risk in the working face and to prompt the strengthening of monitoring and analysis, while the high-level warning is used to indicate that there is a significant prominent danger in the working face and to trigger response measures such as evacuation, shutdown and anti-outburst disposal.

[0037] Another objective of this invention is to propose a coal and gas outburst early warning system based on multi-source information fusion, comprising: The multi-source monitoring data synchronous acquisition module is configured to deploy multi-source monitoring sensor units at the coal mine tunneling face to synchronously acquire multi-source monitoring data during the outburst incubation process; the multi-source monitoring data includes: AE signal, EMR signal, FBG strain signal and CH4 concentration signal; The joint noise filtering and reconstruction module is configured to process the AE signal and EMR signal using a joint noise filtering and reconstruction method that combines singular value decomposition (SVD) and ensemble empirical mode decomposition (EEMD) to obtain the reconstructed signal. The time-frequency feature construction module is configured to combine the synchronously acquired FBG strain signal and CH4 concentration signal to perform time-domain, frequency-domain and time-frequency-domain feature analysis on multi-source monitoring data, extract precursor information that can characterize the evolution of coal and rock mass damage, stress adjustment and gas anomaly changes, and construct multi-dimensional feature expression. The single-source precursor feature recognition module is configured to input the obtained AE two-dimensional time-frequency image, EMR two-dimensional time-frequency image, FBG two-dimensional time-frequency image and CH4 two-dimensional time-frequency image, together with the corresponding time domain feature parameters and frequency domain feature parameters, into the single-source precursor feature intelligent recognition model to identify the prominent precursor state corresponding to each single monitoring source. The multi-source information fusion early warning module is configured to, after obtaining the single-source precursor identification results and corresponding probabilities of the four types of monitoring sources (AE signal, EMR signal, FBG strain signal, and CH4 concentration signal), further employ a multi-source information fusion method based on dynamic weighted Bayesian inference as the core model for multi-source information fusion decision-making. This model unifies, dynamically assigns weights, fuses inferences, and outputs decisions on the identification results of different monitoring sources, thereby achieving comprehensive identification of coal and gas outburst hazard states.

[0038] In one example of the present invention, the single-source precursor feature recognition module includes: The image preprocessing unit is configured to preprocess each single-source two-dimensional time-frequency image to ensure that the two-dimensional time-frequency images constructed from different monitoring sources have a unified data expression form before being input into the recognition model; The feature extraction module is configured to extract directional gradient histogram features and local binary pattern texture features from each single-source two-dimensional time-frequency image. These features are used to characterize the energy distribution edges, local texture changes, frequency band clustering features, and abnormal evolution patterns in the image. At the same time, the extracted time-domain feature parameters and frequency-domain feature parameters are incorporated into the recognition input, thereby forming a single-source comprehensive feature that combines image representation capabilities and statistical representation capabilities. The single-source integrated feature module is configured to be used for any single-source monitoring source k, wherein, Its corresponding image descriptor feature vector and statistical eigenvectors They are respectively recorded as:

[0039]

[0040] The image descriptor feature vector With statistical eigenvectors The features are concatenated to form a single-source comprehensive feature vector. Among them, Z k This represents the fusion feature representation of single-source monitoring information; A single-source precursor identification unit is configured to perform principal component analysis (PCA) dimensionality reduction on the comprehensive feature vector. Based on this, the dimensionality-reduced feature vector is input into a support vector machine (SVM) classifier to identify the state of the single-source precursor. Let the dimensionality reduction transformation matrix be P, then the dimensionality-reduced feature vector is represented as:

[0041] in, The input feature vector is used for the dimensionality-reduced single-source recognition.

[0042] Compared with the prior art, the present invention has the following beneficial effects: (1) It integrates parameters from four different physical mechanisms: acoustic (AE), optical (fiber strain), electrical (EMR), and gas (CH4), overcoming the one-sidedness and limitations of single-indicator early warning and significantly improving the completeness of early warning information. The SVD-EEMD joint algorithm can effectively filter out strong downhole environmental noise and reconstruct effective signals containing prominent precursor features, providing a high-quality data foundation for subsequent feature extraction, status recognition, and fusion early warning.

[0043] (2) The method of combining image feature descriptors with support vector machines is used to automatically identify two-dimensional time-frequency feature maps, which reduces the subjectivity caused by traditional manual threshold setting and experience-based interpretation. At the same time, the method of using dynamic weighted Bayesian inference to fuse multi-source identification results can comprehensively consider the data quality, identification confidence level and data integrity of each monitoring source. The method has better interpretability and robustness.

[0044] (3) The present invention is particularly suitable for coal mine safety monitoring scenarios. The proposed intelligent early warning system for coal and gas outbursts can achieve stable and reliable early warning under the conditions of complex monitoring environment and dynamic changes in the quality of multi-source information, which is convenient for on-site deployment and engineering implementation.

[0045] The preferred embodiments of the invention will be described in more detail below with reference to the accompanying drawings, so as to facilitate an understanding of the features and advantages of the invention. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0047] Figure 1 A flowchart of a coal and gas outburst early warning method based on multi-source information fusion according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a single-source precursor feature recognition model based on image feature descriptors and support vector machines according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a multi-source information fusion early warning model based on dynamic weighted Bayesian inference according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the overall architecture of the intelligent early warning system for coal and gas outbursts according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the downhole layout of an intelligent monitoring system according to an embodiment of the present invention; Figure 6 This is a graph of the original electromagnetic radiation signal according to an embodiment of the present invention; Figure 7 This is a graph of the original electromagnetic radiation signal after SVD noise filtering according to an embodiment of the present invention. Figure 8 This is a graph of the original acoustic emission signal according to an embodiment of the present invention; Figure 9 This is a graph of the original acoustic emission signal after SVD noise filtering according to an embodiment of the present invention; Figure 10This is a diagram showing the EEMD signal decomposition results according to an embodiment of the present invention; Figure 11 This is a comparison chart of changes in early warning indicators according to an embodiment of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The same reference numerals in the drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0049] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “comprising” or “including” mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as “connected” or “linked” are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as “upper,” “lower,” “left,” and “right” are used only to indicate relative positional relationships; these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0050] According to a first aspect of the present invention, a coal and gas outburst early warning method based on multi-source information fusion, such as... Figure 1 As shown, it includes the following steps: S10: Synchronous Acquisition of Multi-Source Monitoring Data: Multi-source monitoring sensor units are deployed at the coal mine tunneling face to synchronously acquire multi-source monitoring data during the outburst incubation process. The multi-source monitoring data includes acoustic emission (AE) signals, electromagnetic radiation (EMR) signals, FBG strain signals, and CH4 concentration signals. The multi-source monitoring sensor units are installed at key locations corresponding to the monitoring targets, such as the tunneling face, coal wall perimeter, and roadway sidewalls. The installation positions, spacing, and quantities of various sensors are configured according to the site conditions to ensure effective coverage of the tunneling face. Each sensor is connected to the underground monitoring substation according to unified acquisition requirements. The underground monitoring substation collects, numbers, and initially organizes the data from different channels, and transmits it to the ground monitoring center through the underground industrial ring network and ground exchange device, achieving centralized reception and unified management of the monitoring data. Simultaneously, to ensure the temporal consistency of the multi-source heterogeneous monitoring data in subsequent fusion analysis, a unified clock reference is implemented for different monitoring channels during the acquisition process, enabling various monitoring data to form corresponding data sequences at the same time scale. The four types of raw monitoring signals acquired at different times t are summarized as follows:

[0051] Where n is the number of sampling points within the current time window. The above four types of raw monitoring signals, after being collected synchronously, serve as the basic input for subsequent data preprocessing, feature extraction, and fusion-based early warning analysis.

[0052] S20: Joint denoising and reconstruction of acoustic emission (AE) and electromagnetic radiation (EMR) signals: Due to various interference factors such as mechanical vibration, electromagnetic disturbance, and operational noise in the underground tunneling environment, acoustic emission (AE) and electromagnetic radiation (EMR) signals are usually non-stationary and noisy time-series signals. If directly used for subsequent feature extraction and early warning identification, they are prone to causing the overwhelming of effective precursor information, severe mode mixing, and a decrease in recognition accuracy. To improve the extraction capability of effective information from monitoring data, a joint denoising and reconstruction method combining Singular Value Decomposition (SVD) and Ensemble Empirical Mode Decomposition (EEMD) is used to process the AE and EMR signals to obtain a reconstructed signal with both a high signal-to-noise ratio and strong precursor characterization capability. S30: Construction of time-frequency features of multi-source signals: Further combining the FBG strain signal and CH4 concentration signal obtained by synchronous acquisition, time-domain, frequency-domain and time-frequency-domain feature analysis are carried out on the multi-source monitoring data respectively to extract the precursor information that can characterize the damage evolution, stress adjustment and gas anomaly changes of coal and rock mass, and construct a multi-dimensional feature expression for subsequent intelligent identification and fusion early warning. S40: Intelligent recognition of single-source precursor features: The obtained AE two-dimensional time-frequency image, EMR two-dimensional time-frequency image, FBG two-dimensional time-frequency image and CH4 two-dimensional time-frequency image are input together with the corresponding time domain feature parameters and frequency domain feature parameters into the intelligent recognition model of single-source precursor features to identify the prominent precursor status corresponding to each single monitoring source. S50: Multi-source Information Fusion Early Warning: After obtaining the single-source precursor identification results and corresponding probabilities of the four types of monitoring sources—AE signal, EMR signal, FBG strain signal, and CH4 concentration signal—a multi-source information fusion method based on dynamic weighted Bayesian inference is further adopted as the core model for multi-source information fusion decision-making. This method unifies, dynamically weights, fuses, infers, and outputs the identification results of different monitoring sources to achieve comprehensive identification of coal and gas outburst hazard states. The dynamic weighted Bayesian inference method can achieve the fusion of multi-source identification results within a unified probability framework while preserving the differences among multiple monitoring sources. Furthermore, it can adaptively adjust the degree of influence based on the current data quality and identification stability of each monitoring source, thereby improving the accuracy, robustness, and feasibility of the comprehensive early warning results.

[0053] In short, this early warning method can simultaneously acquire, effectively filter, extract features, intelligently identify, and fuse multiple monitoring signals such as acoustic emission, electromagnetic radiation, fiber optic strain, and gas concentration, thereby achieving real-time, continuous, and accurate early warning of coal and gas outburst hazards.

[0054] This early warning method integrates parameters from four different physical mechanisms: acoustic (AE), optical (fiber optic strain), electrical (EMR), and gas (CH4). This overcomes the limitations and biases of single-indicator early warning systems, significantly improving the completeness of early warning information. The SVD-EEMD joint algorithm effectively filters out strong downhole environmental noise and reconstructs effective signals containing prominent precursory features, providing a high-quality data foundation for subsequent feature extraction, status identification, and fusion-based early warning.

[0055] This early warning method automatically identifies two-dimensional time-frequency feature maps by combining image feature descriptors with support vector machines, reducing the subjectivity of traditional manual threshold setting and experience-based interpretation. At the same time, by fusing multi-source identification results through dynamic weighted Bayesian inference, it can comprehensively consider the data quality, identification confidence level, and data integrity of each monitoring source. The method has better interpretability and robustness.

[0056] This early warning method is particularly suitable for coal mine safety monitoring scenarios. The proposed intelligent early warning system for coal and gas outbursts can achieve stable and reliable early warning under complex monitoring environments and dynamic changes in the quality of multi-source information, which is convenient for field deployment and engineering implementation.

[0057] In one example of the present invention, step S20 specifically includes the following steps: S21: For any original signal sequence Let n be the number of sampling points in the current sliding analysis window. Construct its corresponding Hankel matrix according to the phase space reconstruction idea:

[0058] Where m + c - 1 = n; then, singular value decomposition is performed on the Hankel matrix to obtain:

[0059] Where U is the left singular vector matrix, V is the right singular vector matrix, and Σ is the singular value diagonal matrix. And satisfy ; S22: Set a threshold based on the singular value mutation characteristics or cumulative energy ratio, preferably 0.9. Determine the effective singular value retention order based on the threshold, and truncate the singular value matrix, retaining the larger singular values ​​that mainly represent effective information and setting the smaller singular values ​​corresponding to noise components to zero, thereby obtaining the preliminary noise filtering matrix. Then, the initial noise filtering matrix The initial noise-filtered signal is obtained by performing anti-diagonal averaging reconstruction. This step utilizes the energy distribution difference between the effective signal and noise in the singular spectrum of the matrix to suppress random noise and local glitches, while preserving the overall trend of the original signal and key anomalous fluctuation information. After SVD processing, local glitches in electromagnetic radiation and acoustic emission signals are improved, and the overall trend is well preserved.

[0060] S23: Initial noise filtering signal Performing ensemble empirical mode decomposition yields p intrinsic mode function (IMF) components and residual terms, expressed as follows:

[0061] in, The i-th intrinsic mode component The residual term; the purpose of using EEMD is to reduce the mode aliasing phenomenon that easily occurs in traditional methods by introducing ensemble white noise to assist decomposition, so that the multi-scale features of non-stationary signals can be separated more stably. EEMD is particularly suitable for non-stationary acoustic emission and electromagnetic radiation intensity signals acquired downhole, and can enhance the ability to extract frequency variation features and effective precursor information. To avoid the problem of information omission or redundancy accumulation caused by fixed selection of the first few IMF components, S24: The variance contribution rate is used to quantitatively screen each IMF component. For the i-th intrinsic mode component, its variance contribution rate is defined as:

[0062] in, Let represent the variance of the i-th IMF component; based on this, calculate the cumulative variance contribution rate of the q IMF components. :

[0063] S25: Select the one that satisfies The first q IMF components are used to reconstruct the signal, thereby obtaining the acoustic emission reconstructed signal. and electromagnetic radiation reconstructed signal Where θ is a preset cumulative contribution rate threshold, preferably 90%, i.e., θ = 0.9. The final reconstructed signal is expressed as:

[0064] Through the above processing, while suppressing complex downhole noise interference, the effective temporal features related to the instability and evolution of coal and rock mass in the AE and EMR signals can be better preserved, providing high-quality input data for subsequent two-dimensional time-frequency feature construction, precursor feature identification, and multi-source fusion early warning. The signal reconstructed using SVD-EEMD outperforms the simple SVD-filtered signal and the fixed IMF combined signal in precursor feature identification. The joint denoising and reconstruction method has good effectiveness and applicability, and can improve the accuracy of subsequent identification models.

[0065] In one example of the present invention, step S30 specifically includes the following steps: S31: Construct time-domain features for multi-source signals, including energy value, peak amplitude, root mean square value, strain change, and gas concentration change rate; where, for discrete signal sequence x i Its energy value E is expressed as:

[0066] Where n is the number of sampling points in the current analysis window.

[0067] For the gas concentration sequence C(t), to reflect the rate of change of gas concentration between adjacent time points, the rate of change of gas concentration is further calculated, and its expression is as follows:

[0068] Where Δt is the time interval between adjacent sampling times; By extracting the time-domain parameters mentioned above, we can obtain statistical characteristics such as amplitude changes, energy accumulation, and rate of change of various signals within the current time window, thereby providing a basic quantitative basis for abnormal state identification. S32: Construct frequency domain features for multi-source signals, perform spectral analysis on the reconstructed AE signal, EMR signal, as well as the FBG strain signal and CH4 concentration signal after unified time window processing, and extract frequency domain feature parameters including peak frequency, center frequency, and centroid frequency to characterize the distribution characteristics of signal frequency components and the variation law of dominant frequency band. If the discrete spectrum is P(f) i If f is the centroid frequency, then f g It can be represented as:

[0069] Among them, f i For the i-th discrete frequency point, P(f i () represents the spectral value at the corresponding frequency point; through frequency domain feature extraction, the differences in energy distribution and spectral centroid shift characteristics of the signal in different frequency ranges can be further characterized, thereby enhancing the ability to identify abnormal responses during the gestation process; S33: Based on this, further time-frequency domain features are constructed; since multi-source monitoring signals in the coal and gas outburst incubation process generally have nonlinear, non-stationary, and transient change characteristics, it is difficult to fully reflect their local time-varying laws by simply relying on time-domain or frequency-domain analysis. Therefore, wavelet decomposition and Hilbert-Huang transform methods are used to perform joint time-frequency analysis on various signals to construct a two-dimensional time-frequency feature image; the two-dimensional time-frequency feature image is used to synchronously characterize the energy distribution evolution characteristics of the signal on the time axis and frequency axis, so as to improve the ability to express and distinguish the precursor information of outburst evolution. When performing wavelet decomposition, the signal is decomposed into multiple frequency bands, and the energy proportion of each frequency band is calculated as the time-frequency feature parameter; by calculating the energy proportion of different frequency bands, the distribution characteristics of signal energy at various scales can be effectively characterized, and a two-dimensional time-frequency feature representation with discriminative significance can be constructed accordingly; where, if the energy of the j-th frequency band is E j If the total number of frequency bands is L, then the energy percentage ρ of the j-th frequency band is... j It can be represented as:

[0070] In the formula, E l Energy per frequency band; Compared to a single statistical measure, the time-frequency image can more fully preserve local anomaly information, energy transfer information, and frequency band evolution information in multi-source signals, making it more suitable as input features for subsequent recognition models. This step constitutes a key link between the preceding and following technical stages.

[0071] S34: Through the above processing, the final two-dimensional time-frequency image I of AE is obtained respectively. AE EMR two-dimensional time-frequency image I EMR FBG two-dimensional time-frequency image I FBGand CH4 two-dimensional time-frequency image I CH4 Simultaneously, by combining the time-domain and frequency-domain feature parameters corresponding to various signals, a structured feature set of multi-source signals can be formed. The structured feature set and the two-dimensional time-frequency image together serve as the input basis for the subsequent construction of a prominent precursor feature recognition model and the multi-source information fusion early warning analysis, thereby realizing the transformation from raw monitoring data to identifiable feature expression.

[0072] In one example of the present invention, such as Figure 2 As shown, step S40 specifically includes the following steps: S41: Preprocess the two-dimensional time-frequency images of each single source to ensure that the two-dimensional time-frequency images constructed from different monitoring sources have a unified data expression form before being input into the recognition model; S42: Extract directional gradient histogram features and local binary pattern texture features from each single-source two-dimensional time-frequency image to characterize energy distribution edges, local texture changes, frequency band clustering features, and abnormal evolution patterns in the image; at the same time, incorporate the extracted time-domain feature parameters and frequency-domain feature parameters into the recognition input to form a single-source comprehensive feature that combines image representation capabilities and statistical representation capabilities. S43: For any single-source monitoring source k, where, Its corresponding image descriptor feature vector and statistical eigenvectors They are respectively recorded as:

[0073]

[0074] The image descriptor feature vector is concatenated with the statistical feature vector to form a single-source comprehensive feature vector. Among them, Z k This represents the fusion feature vector of single-source monitoring information; k can take four monitoring sources: AE, EMR, FBG, and CH4. This method combines morphological features from two-dimensional time-frequency images with the original time-domain and frequency-domain statistical information. Furthermore, considering the high dimensionality of the combined features after stitching, to reduce redundant information, lower computational complexity, and improve classification stability, S44: Perform principal component analysis to reduce the dimensionality of the comprehensive feature vector. Based on this, input the dimensionality-reduced feature vector into a support vector machine classifier to identify single-source precursor states. Let the dimensionality reduction transformation matrix be P, then the dimensionality-reduced feature vector is represented as:

[0075] in, The input feature vector is used for the dimensionality-reduced single-source recognition. Principal component analysis is used to retain the main components in the comprehensive features that contribute significantly to state identification, thereby improving the efficiency of subsequent classification and recognition.

[0076] In one example of the present invention, in step S44, the dimensionality-reduced feature vector is input into a support vector machine classifier to identify single-source precursor states, specifically including the following steps: S441: The support vector machine uses a radial basis kernel function and employs a one-to-many strategy to construct a three-class classification recognition model; to achieve hierarchical recognition of prominent dangerous states, the single-source recognition state set is set as:

[0077] Where w1 represents a safe state, w2 represents a threat state, and w3 represents a dangerous state; For any state category Its support vector machine discriminant function can be expressed as:

[0078] in, Indicates single-source monitoring information regarding status The output value of the discriminant function; N is the number of support vectors; For Lagrange multipliers; Labels for training samples; For kernel functions; For bias term; y i This represents the feature vector of the i-th training sample obtained during the training process; S442: Simultaneously, to facilitate subsequent multi-source fusion processing, the discriminant values ​​output by the support vector machine are converted into probabilistic form; for single-source monitoring information belonging to the state... The recognition probability is expressed as:

[0079] in, This indicates that the single-source monitoring information belongs to the status category. The recognition probability; l is the summation index, l=1, 2, 3; S443: Based on the probability values ​​corresponding to each state category, the optimal identification result of the current single-source monitoring information can be determined. This yields the single-source precursor identification results and corresponding probability outputs for the four monitoring sources: AE signal, EMR signal, FBG strain signal, and CH4 concentration signal. This realizes the transformation from multi-source time-frequency feature construction to intelligent single-source state identification. Simultaneously, the probability values ​​of each state can also reflect the degree of support the monitoring source provides for different risk states, offering a foundational input for subsequent multi-source information fusion analysis. The optimal identification result... The expression is: .

[0080] In one example of the present invention, such as Figure 3 As shown, step S50 specifically includes the following steps: S51: Construct a multi-source fusion identification framework; the early warning status identification framework is as follows: Where w1 represents a safe state, w2 represents a threat state, and w3 represents a dangerous state; S52: To reflect the effectiveness and reliability of different monitoring sources at the current moment, a dynamic quality evaluation index is constructed for any monitoring source k. The dynamic quality evaluation index comprehensively considers three aspects: signal quality, identification confidence, and data integrity. Signal quality characterizes the current signal-to-noise level of the monitoring source; identification confidence characterizes the reliability of the single-source classification result; and data integrity characterizes the proportion of missing, abnormal, and invalid data within the current time window. The dynamic quality evaluation index is defined as follows:

[0081] in, This represents the normalized signal quality index corresponding to the k-th monitoring source at time t; This represents the confidence index for normalized identification; This represents the proportion of normalized outliers or missing values; η1, η2, and η3 are weighting coefficients, and satisfy the following conditions: .

[0082] S53: Normalize the dynamic quality evaluation indicators of each monitoring source to obtain the dynamic fusion weight of each single-source monitoring information at the current moment. The expression is as follows:

[0083] in, q represents the fusion weight of the k-th monitoring source at time t; r (t) represents the dynamic quality evaluation index of the r-th monitoring source at time t. By constructing weights, monitoring sources with higher monitoring quality, stronger identification stability, and better data integrity can make a greater contribution in the fusion process, while the influence of monitoring sources that are significantly affected by noise, have unstable identification results, or have missing or abnormal data is reduced accordingly.

[0084] S54: The recognition probability output from each single-source monitoring information With dynamic weights By combining these factors, a weighted likelihood term for each monitoring source with respect to different early warning states is constructed, and its expression is as follows:

[0085] Where ε is a very small positive constant introduced to prevent the probability value from being zero, and its value ranges from 1 × 10⁻⁶. -4 ≤ ε ≤ 1×10 -3 When the quality of the monitoring source is high, A larger output state probability has a stronger impact on the final fusion result; when the monitoring source is affected by noise interference or the recognition confidence is low. The smaller the monitoring source, the weaker its influence on the fusion results.

[0086] Based on this, the prior probability π of each warning state is introduced. j By combining the weighted likelihood terms of the four monitoring sources with Bayesian fusion, the posterior probabilities of the system being in each warning state at the current moment are obtained:

[0087] in, This represents the set of multi-source monitoring observations corresponding to time t; This indicates that the system is in state w after fusion. j The posterior probability; S55: Considering the complex downhole conditions and the short-term fluctuations and transient disturbances in the monitoring signals, a time-recursive smoothing mechanism is introduced to improve the temporal continuity and stability of the fusion results. This mechanism smoothly updates the fusion results at adjacent time points, and its expression is as follows:

[0088] in, This represents the fusion probability after smoothing at time t; It is a smoothing coefficient, and satisfies 0 < γ ≤ 1; Based on the smoothed fusion probability, the state corresponding to the largest probability is taken as the comprehensive warning state at the current moment. This enables the transformation from multi-source single-source identification results to comprehensive state identification results, including comprehensive early warning state. The expression is:

[0089] S56: Furthermore, to improve the interpretability and continuous characterization capability of the fusion results in engineering applications, a comprehensive early warning index R is constructed. t Through the comprehensive early warning index R t The discrete classification results are further transformed into continuous risk characterization quantities to reflect the overall changing trend and phased evolution characteristics of the current prominent hazards at the working face. Among them, the comprehensive early warning index R t The expression is:

[0090] Where j is the warning status category index, j=1,2,3, corresponding to safe status w1, threat status w2 and dangerous status w3 respectively; λ1, λ2, λ3 are the risk mapping coefficients corresponding to safe, threat and dangerous statuses respectively, and satisfy 1≥λ3>λ2>λ1≥0; Comprehensive Early Warning Index R t The corresponding warning level classification thresholds are not preset fixed constants, but are determined through joint calibration based on laboratory test results and field monitoring data. Specifically, by combining coal and rock loading failure tests, outburst simulation tests, historical monitoring samples, and field operating condition data, the distribution range of the comprehensive warning index under different dangerous stages is statistically analyzed and calibrated to determine the threshold ranges corresponding to different warning levels such as safety, threat, and danger. Joint calibration through a combination of laboratory and field methods ensures that the determined warning level classification standards are more consistent with specific mine geological conditions, mining technology characteristics, and monitoring system response patterns, thereby improving the engineering applicability of the warning model and the accuracy of on-site identification.

[0091] Through the above steps, this invention achieves the transformation from single-source identification results to multi-source fusion decision results, forming a comprehensive early warning method that takes into account dynamic changes in data quality, differences in single-source identification, and temporal continuity. This step, together with the aforementioned steps S1 to S4, constitutes a chain of intelligent early warning technology for coal and gas outbursts based on multi-source information fusion. This enables the early warning results to not only reflect the local abnormal characteristics of a single monitoring source but also to reflect the overall dangerous state under the conditions of multi-physics field information coupling, providing a basis for dynamic assessment and advanced early warning of coal and gas outburst hazards.

[0092] In one example of the present invention, the method further includes: Step S60: Warning output and result display: After completing the multi-source information fusion warning, based on the current comprehensive warning status and the comprehensive warning index R... t The judgment result automatically triggers the corresponding level of early warning output; at the same time, the software platform of the ground monitoring center displays and records the early warning result in real time.

[0093] In one example of the present invention, in step S60, based on the current comprehensive early warning status and the comprehensive early warning index R... t The judgment result will automatically trigger the corresponding level of early warning output, specifically including the following: When the comprehensive early warning index R t When the judgment range corresponding to the corresponding warning level is reached, and the fusion decision result indicates that the current working face is in a threatened or dangerous state, the system generates corresponding alarm signals and interface prompts according to the pre-set warning strategy. Among them, the lower-level warning is used to indicate that there is a prominent evolution risk at the working face and to prompt the strengthening of monitoring and analysis, while the higher-level warning is used to indicate that there is a significant prominent danger at the working face and to trigger response measures such as evacuation, shutdown and anti-outburst disposal.

[0094] When the comprehensive early warning index R t When the judgment range corresponding to the corresponding warning level is reached, and the fusion decision result indicates that the current working face is in a threatened or dangerous state, the system generates corresponding alarm signals and interface prompts according to the pre-set warning strategy. Among them, the lower-level warning is used to indicate that there is a risk of outburst evolution at the working face and to prompt the strengthening of monitoring and analysis, while the higher-level warning is used to indicate that there is a significant outburst hazard at the working face and to trigger response measures such as evacuation, shutdown, and outburst prevention. The correspondence between the warning levels and the triggering conditions can be jointly calibrated and dynamically corrected by combining laboratory test results, field monitoring data, and actual mine outburst prevention requirements to improve the pertinence and engineering applicability of the warning output.

[0095] The displayed content includes the change curves of four types of monitoring signals, the identification results of individual precursors, the fusion posterior probability results, the change curve of the comprehensive early warning index, and the current early warning status and historical early warning records. By synchronously displaying multi-source monitoring data and fusion decision results, monitoring personnel can intuitively grasp the dynamic evolution of coal and gas outburst hazard states, and provide a basis for subsequent early warning assessment, emergency response, and historical data tracing analysis, thereby achieving a complete closed loop from multi-source monitoring, intelligent identification, fusion early warning to result output and visualization.

[0096] According to a second aspect of the present invention, a coal and gas outburst early warning system based on multi-source information fusion, such as... Figure 4 and Figure 5 As shown, it includes: A multi-source monitoring data synchronous acquisition module is configured to deploy multi-source monitoring sensor units at the coal mine tunneling face to synchronously acquire multi-source monitoring data during the outburst incubation process. The multi-source monitoring data includes: AE signal, EMR signal, FBG strain signal, and CH4 concentration signal. The multi-source monitoring sensor units are installed at key locations corresponding to the monitoring targets, such as the tunneling face, coal wall perimeter, and roadway sidewalls. The installation positions, spacing, and quantities of various sensors are configured according to the site conditions to ensure effective coverage of the tunneling face. Each sensor is connected to an underground monitoring substation according to unified acquisition requirements. The underground monitoring substation collects, numbers, and initially organizes the data from different channels, and transmits it to the ground monitoring center through the underground industrial ring network and ground exchange device, achieving centralized reception and unified management of the monitoring data. Simultaneously, to ensure the temporal consistency of the multi-source heterogeneous monitoring data in subsequent fusion analysis, a unified clock reference is implemented for different monitoring channels during the acquisition process, enabling various monitoring data to form corresponding data sequences at the same time scale. The joint noise filtering and reconstruction module is configured to process acoustic emission (AE) signals and electromagnetic radiation (EMR) signals using a joint noise filtering and reconstruction method combining Singular Value Decomposition (SVD) and Ensemble Empirical Mode Decomposition (EEMD) to obtain reconstructed signals with both high signal-to-noise ratio and strong precursor characterization capabilities. Due to various interference factors in the underground tunneling environment, such as mechanical vibration, electromagnetic disturbances, and operational noise, acoustic emission (AE) signals and electromagnetic radiation (EMR) signals are typically non-stationary and highly noisy time-series signals. Directly using them for subsequent feature extraction and early warning identification can easily lead to the overwhelming of effective precursor information, severe modal mixing, and decreased recognition accuracy. This module aims to improve the extraction capability of effective information from monitoring data.

[0097] The time-frequency feature construction module is configured to combine the synchronously acquired FBG strain signal and CH4 concentration signal to conduct time-domain, frequency-domain and time-frequency-domain feature analysis on multi-source monitoring data, extract precursor information that can characterize the evolution of coal and rock mass damage, stress adjustment and gas anomaly changes, and construct a multi-dimensional feature expression for subsequent intelligent identification and fusion early warning. The single-source precursor feature recognition module is configured to input the obtained AE two-dimensional time-frequency image, EMR two-dimensional time-frequency image, FBG two-dimensional time-frequency image and CH4 two-dimensional time-frequency image, together with the corresponding time domain feature parameters and frequency domain feature parameters, into the single-source precursor feature intelligent recognition model to identify the prominent precursor state corresponding to each single monitoring source. The multi-source information fusion early warning module is configured to, after obtaining the single-source precursor identification results and corresponding probabilities of the four monitoring sources (AE signal, EMR signal, FBG strain signal, and CH4 concentration signal), further employ a multi-source information fusion method based on dynamic weighted Bayesian inference as the core model for multi-source information fusion decision-making. This model uniformly expresses, dynamically weights, fuses, infers, and outputs decisions based on the identification results of different monitoring sources, thereby achieving a comprehensive identification of coal and gas outburst hazard states. The dynamic weighted Bayesian inference method can achieve the fusion of multi-source identification results within a unified probability framework while preserving the differences among multiple monitoring sources. Furthermore, it can adaptively adjust the degree of influence based on the current data quality and identification stability of each monitoring source, thereby improving the accuracy, robustness, and feasibility of the comprehensive early warning results.

[0098] Specifically, the early warning system includes: an underground monitoring and sensing module and a data transmission module, and a surface data analysis and early warning module. The multi-source monitoring data synchronous acquisition module includes the underground monitoring and sensing module and the data transmission module; the surface data analysis and early warning module includes: a joint noise filtering and reconstruction module, a time-frequency feature construction module, a single-source precursor feature identification module, and a multi-source information fusion early warning module. The modules work collaboratively according to the technical path of "underground acquisition - transmission and aggregation - surface processing - early warning output" to achieve continuous perception, dynamic analysis, and graded early warning of coal and gas outburst hazards during tunneling.

[0099] The underground monitoring sensor module is installed at the coal mine tunneling face and its corresponding key monitoring areas to synchronously collect multi-physics response information from the working face. The underground monitoring sensor module includes acoustic emission sensors, electromagnetic radiation sensors, fiber optic strain sensors, and gas concentration sensors to acquire acoustic emission monitoring data, electromagnetic radiation monitoring data, fiber optic strain monitoring data, and gas concentration monitoring data, respectively. The various sensors are deployed according to the spatial location of the tunneling face, the monitoring targets, and the on-site working conditions, and are connected to the underground monitoring substation, thereby achieving unified access and synchronous acquisition of multi-source monitoring information.

[0100] The data transmission module is used to transmit multi-source monitoring data collected underground to the ground monitoring center in real time. The data transmission module includes underground monitoring substations, an industrial ring network switch, and a ground communication interface. The underground monitoring substations receive, aggregate, and initially process data from each monitoring channel; the industrial ring network switch ensures stable transmission from underground to the surface; and the ground communication interface enables data interaction with the ground monitoring center. Through this data transmission module, unified uploading, unified identification, and unified management of monitoring data from different sources and of different types can be achieved, providing a reliable data link for subsequent ground-based fusion analysis.

[0101] The ground data analysis and early warning module is used to process, analyze, identify, and fuse early warning data transmitted to the ground from multiple sources. Preferably, the ground data analysis and early warning module sequentially performs signal preprocessing, feature extraction and identification, multi-source fusion early warning, and human-computer interaction. In the signal preprocessing stage, namely the joint noise filtering and reconstruction module and the time-frequency feature construction module, joint noise filtering and reconstruction based on SVD-EEMD is performed on the AE and EMR signals. In the feature extraction and identification stage, namely the single-source precursor feature identification module, two-dimensional time-frequency feature images of each single-source monitoring signal are constructed, image descriptor features and time-domain and frequency-domain statistical features are extracted, and the single-source precursor status is identified based on an identification model combining image feature descriptors and support vector machines. In the multi-source fusion early warning stage, namely the multi-source information fusion early warning module, dynamic weighting, posterior probability fusion, time smoothing update, and decision output are performed on the identification results of each single source based on a dynamically weighted Bayesian inference model to form a comprehensive early warning result and a comprehensive early warning index. The human-computer interaction interface is used to visualize the monitoring curves, fusion early warning results, early warning levels, and historical data.

[0102] Through the aforementioned system configuration, this invention achieves integrated functions of multi-source synchronous monitoring, intelligent identification, and fusion early warning for coal and gas outburst hazards. Compared with traditional single-parameter monitoring systems, this system can simultaneously utilize multiple monitoring information such as sound, light, electricity, and gas to provide a more comprehensive characterization of the outburst hazard state at the working face. By combining single-source intelligent identification with multi-source dynamic fusion, it improves the accuracy, real-time performance, and reliability of early warning results under complex working conditions, thereby providing technical support for safe production and outburst prevention decisions in coal mine tunneling faces.

[0103] This early warning system integrates parameters from four different physical mechanisms: acoustic (AE), optical (fiber optic strain), electrical (EMR), and gas (CH4). This overcomes the limitations and biases of single-indicator early warning systems, significantly improving the completeness of early warning information. Employing the SVD-EEMD joint algorithm, it effectively filters out strong downhole environmental noise and reconstructs effective signals containing prominent precursory features, providing a high-quality data foundation for subsequent feature extraction, status identification, and fusion-based early warning.

[0104] This early warning system automatically identifies two-dimensional time-frequency feature maps by combining image feature descriptors with support vector machines, reducing the subjectivity of traditional manual threshold setting and experience-based interpretation. At the same time, it fuses multi-source identification results through a dynamic weighted Bayesian inference method, which can comprehensively consider the data quality, identification confidence level and data integrity of each monitoring source. The method has better interpretability and robustness.

[0105] This early warning system is particularly suitable for coal mine safety monitoring scenarios. The proposed intelligent early warning system for coal and gas outbursts can achieve stable and reliable early warning under complex monitoring environments and dynamic changes in the quality of multi-source information, making it easy to deploy on-site and implement in engineering.

[0106] In one example of the present invention, the single-source precursor feature recognition module includes: The image preprocessing unit is configured to preprocess each single-source two-dimensional time-frequency image to ensure that the two-dimensional time-frequency images constructed from different monitoring sources have a unified data expression form before being input into the recognition model; The feature extraction module is configured to extract directional gradient histogram features and local binary pattern texture features from each single-source two-dimensional time-frequency image. These features are used to characterize the energy distribution edges, local texture changes, frequency band clustering features, and abnormal evolution patterns in the image. At the same time, the extracted time-domain feature parameters and frequency-domain feature parameters are incorporated into the recognition input, thereby forming a single-source comprehensive feature that combines image representation capabilities and statistical representation capabilities. The single-source integrated feature module is configured to be used for any single-source monitoring source k, wherein, Its corresponding image descriptor feature vector and statistical eigenvectors They are respectively recorded as:

[0107]

[0108] The image descriptor feature vector With statistical eigenvectors The features are concatenated to form a single-source comprehensive feature vector. Among them, Z k This represents the fusion feature representation of single-source monitoring information; through this method, morphological features in two-dimensional time-frequency images can be combined with the original time-domain and frequency-domain statistical information. Furthermore, considering the high dimensionality of the composite features after stitching, measures are taken to reduce redundant information, lower computational complexity, and improve classification stability. A single-source precursor identification unit is configured to perform principal component analysis (PCA) dimensionality reduction on the comprehensive feature vector. Based on this, the dimensionality-reduced feature vector is input into a support vector machine (SVM) classifier to identify the state of the single-source precursor. Let the dimensionality reduction transformation matrix be P, then the dimensionality-reduced feature vector is represented as:

[0109] in, The input feature vector is used for the dimensionality-reduced single-source recognition.

[0110] Principal component analysis is used to retain the main components in the comprehensive features that contribute significantly to state identification, thereby improving the efficiency of subsequent classification and recognition.

[0111] It should be noted that the coal and gas outburst early warning system based on multi-source information fusion of the present invention can also perform any of the processing described in the coal and gas outburst early warning method based on multi-source information fusion previously, and the specific details are not repeated here.

[0112] Specific examples: (I) Background Environment This embodiment describes the excavation of a coal roadway in a coal mine. The face is located in the 30 mining area at level 3, with an underground elevation ranging from -400.0m to -479.0m, a coal seam thickness ranging from 6.41m to 13.10m, and an average thickness of 8.78m; the coal seam dip angle ranges from 20°42′ to 39°31′, with an average dip angle of 28°30′; the initial gas content of the 3002 face is 14.3m. 3 / t to 17.8m 3 / t, with an initial gas pressure of 1.4MPa to 2.6MPa, belongs to the outburst hazard zone.

[0113] A chamber was excavated and installed on the side of the working face near the tunneling face. The chamber dimensions were 5m wide, 5m deep, and 4m high. The monitoring range for this section was 60m ahead in the tunneling direction, and the actual length of the fiber optic cable inserted was 58m. Two fiber optic strain monitoring boreholes were drilled, with a diameter of Φ=95mm and a height of 1.5m from the bottom plate. The length of the casing protection section inside the borehole was 10m. Grouting was stopped after the grouting pressure reached 3.0MPa to ensure coupling between the fiber optic strain sensor and the coal and rock mass. The electromagnetic radiation sensor is a KBD7 type, with a monitoring frequency range of 30Hz to 1kHz; the acoustic emission sensor is a GDD12 type, with an acoustic emission signal bandwidth of 11Hz to 5.5kHz; the fiber optic strain demodulator is a 2-channel device with a wavelength range of 1527nm to 1568nm, a wavelength resolution of 1pm, repeatability of ±2pm, demodulation rate of 1Hz, and dynamic range of 35dB; the gas sensor is deployed on the anchor network of the lower side of the 3002 lower roadway, specifically located in the return airflow 20m from the tunnel face, for continuously acquiring CH4 volume fraction signals. The acoustic-electric-gas monitoring signals are converted into RS485 signals by the KJT96-F monitoring substation and the KJ2000N substation, and then transmitted to the ground monitoring center through the underground industrial ring network, with a signal transmission cable length of approximately 500m.

[0114] This embodiment selects continuous monitoring data from 00:00 on November 15, 2025 to 00:00 on November 19, 2025 as the application sample. All four types of monitoring sources were uniformly converted to 1-second timestamp data, and statistical analysis was performed in 10-minute sliding analysis windows, with each window containing 600 sampling points, forming 576 sets of synchronous analysis windows; of these, 420 sets were used as model training samples, 78 sets as validation samples, and 78 sets as test samples. Electromagnetic radiation and acoustic emission intensity are expressed in mV, FBG strain in με, and CH4 concentration as volume fraction percentage. The basic thresholds for on-site early warning are set as follows: electromagnetic radiation intensity 90mV, acoustic emission intensity 40mV, and CH4 concentration 0.5%; the early warning coefficient for electromagnetic radiation intensity is 1.6, and the early warning coefficient for acoustic emission intensity is 1.8.

[0115] (II) Step-by-step implementation S10: Synchronous acquisition of multi-source monitoring data. In the 3002 lower roadway monitoring section, KBD7 electromagnetic radiation sensors and GDD12 acoustic emission sensors were suspended near the tunneling face, CH4 sensors were placed on the return air side of the working face, and fiber optic strain cables were laid inside the coal and rock mass along two Φ=95mm boreholes. During the acquisition process, the intensity signals obtained from the electromagnetic radiation channel were mainly distributed between 18mV and 72mV, with a local anomaly window where the original peak value reached 168.2mV; the intensity signals obtained from the acoustic emission channel were mainly distributed between 6mV and 28mV, with a local anomaly window where the original peak value reached 82.6mV; the fiber optic strain signal exhibited a non-stationary curve that gradually accumulated with tunneling disturbances, with a maximum strain increment of 78με in a single window and a maximum cumulative strain of 185με; the CH4 concentration was 0.22% to 0.35% in most windows, reaching a maximum of 0.48% in the anomaly window. The original missing rate of the four types of data during the acquisition period was 0.7%. After timestamp alignment and linear completion, 576 sets of AE-EMR-FBG-CH4 synchronous window data were formed. The signal morphology was manifested as enhanced acoustic emission pulse, sudden increase in electromagnetic radiation, step accumulation of strain, and slow increase in gas concentration.

[0116] S20: Joint noise filtering and reconstruction of AE and EMR signals. For example... Figures 6-9 As shown, a 300×301 Hankel matrix was constructed for 600 AE signals and 600 EMR signals in each 10-minute analysis window, with the singular value truncation threshold set to the cumulative energy percentage η = 0.90. When processing representative windows before and after the high-energy event on November 17, 2025, the EMR signal retained 22 valid singular values, and the AE signal retained 24 valid singular values. When processing all 576 window sets, the EMR signal retained an average of 17 valid singular values, and the AE signal retained an average of 19 valid singular values.

[0117] After initial noise filtering using SVD, such as Figure 10 As shown, EEMD decomposition was performed on the AE and EMR signals respectively, with 100 iterations. White noise amplitude was added at 0.2 times the standard deviation of the corresponding signal. In AE reconstruction, 10 IMF components and 1 residual term were obtained. The top 5 IMF components with a cumulative variance contribution rate of 92.7% were selected for reconstruction. In EMR reconstruction, 9 IMF components and 1 residual term were obtained. The top 4 IMF components with a cumulative variance contribution rate of 91.5% were selected for reconstruction. After processing a total of 1152 sets of acoustic-electric signals (576 AE windows and 576 EMR windows), the average signal-to-noise ratio (SNR) of EMR increased from 8.9 dB to 19.4 dB, and the average SNR of AE increased from 9.6 dB to 20.8 dB. The peak value of EMR reconstruction in the abnormal window was 126.4 mV, and the peak value of AE reconstruction was 51.3 mV. Random glitches were reduced by an average of 73.2%, while preserving the trend of continuous signal rise before high-energy events.

[0118] S30: Construction of Time-Frequency Features for Multi-Source Signals. Time-domain, frequency-domain, and time-frequency-domain features were calculated for the reconstructed AE and EMR signals, as well as the synchronously obtained fiber optic strain and CH4 concentration signals. Eight structured features were extracted from each monitoring source within a window, including peak value, root mean square value, energy, rate of change, peak frequency, center frequency, centroid frequency, and the energy proportion of the main frequency band. Each window was then converted into a 224×224 pixel two-dimensional time-frequency image. After processing 576 groups of windows, a total of 2304 two-dimensional time-frequency images (AE, EMR, FBG, and CH4) were obtained, forming a 576×32-dimensional structured feature matrix.

[0119] Taking the abnormal window from 20:20 to 20:30 on November 17, 2025 as an example, the peak value of the EMR reconstructed signal was 126.4mV, the root mean square value was 55.8mV, and the calculated energy was 1.87×10⁻⁶. 6 mV 2 The centroid frequency is 312Hz, and the high-frequency band energy accounts for 46.8%. The peak value of the AE reconstructed signal is 51.3mV, the root mean square value is 20.9mV, and the calculated energy is 2.62×10⁻⁶. 5 mV 2 The centroid frequency was 1.14 kHz, with high-frequency band energy accounting for 52.1%; the FBG strain increment was 78 με, with low-frequency strain energy accounting for 68.4%; the CH4 concentration increased from 0.31% to 0.48%, with a concentration change rate of 0.017% / min, and low-frequency change energy accounting for 74.2%. These values ​​indicate that the acoustic and electrical signals exceeded the on-site critical values ​​within the same window, and the fiber optic strain and gas concentration also showed a synchronous abnormal trend.

[0120] S40: Intelligent Recognition of Single-Source Precursor Features. For each 224×224 pixel two-dimensional time-frequency image, grayscale normalization and size unification are performed. 1764-dimensional directional gradient histogram features and 256-dimensional local binary pattern texture features are extracted and concatenated with 8-dimensional structured statistical features to obtain a 2028-dimensional single-source comprehensive feature vector. This vector is then reduced to 32 dimensions through principal component analysis and input into a radial basis function (RBF) kernel support vector machine (SVM) classifier. The SVM parameters are set to C=8 and γ=0.03125, and the recognition states are set as safe state w1, threat state w2, and dangerous state w3.

[0121] In 78 test samples, the single-source identification accuracy rates were 91.0% for AE, 89.7% for EMR, 84.6% for FBG, and 82.1% for CH4. For the abnormal window from 20:20 to 20:30 on November 17, 2025, the probabilities of safety, threat, and danger output by AE were 0.04, 0.18, and 0.78, respectively; those by EMR were 0.07, 0.23, and 0.70; those by FBG were 0.12, 0.34, and 0.54; and those by CH4 were 0.19, 0.42, and 0.39. Therefore, AE, EMR, and FBG all identified this window as a dangerous state, while CH4 was identified as a threat state because its highest concentration did not reach the 0.5% critical value.

[0122] S50: Multi-source information fusion early warning. Dynamic quality evaluation indicators are calculated for each monitoring source. Signal quality is normalized based on the reconstructed signal-to-noise ratio, identification confidence is normalized based on the single-source maximum category probability, and data integrity is normalized based on the proportion of valid sampling points in the current window. In the above-mentioned abnormal window, the dynamic quality evaluation values ​​for AE, EMR, FBG, and CH4 are 0.91, 0.89, 0.73, and 0.58, respectively, with corresponding normalized dynamic fusion weights of 0.293, 0.286, 0.235, and 0.186. In Bayesian fusion, the state prior probabilities are set to π1=1 / 3, π2=1 / 3, and π3=1 / 3, with a zero-prevention constant ε=1×10⁻⁶. -6 The time-recursive smoothing coefficient γ = 0.65.

[0123] After fusing the probabilities of each individual source with dynamic weights, the posterior probabilities of safety, threat, and danger for this anomaly window before smoothing are 0.06, 0.42, and 0.52, respectively; after time-progressive smoothing, the posterior probabilities of safety, threat, and danger are 0.054, 0.270, and 0.676, respectively. Risk mapping coefficients are set to λ1=0.10, λ2=0.55, and λ3=0.90, and the comprehensive early warning index R is calculated. t =0.762. The warning level threshold is set to R. tA value < 0.45 indicates a safe state, and 0.45 ≤ R t A value <0.70 indicates a threat state. t A value ≥0.70 indicates a dangerous state, therefore the system outputs a "dangerous" comprehensive warning result.

[0124] (III) Conclusion Analysis The final result of this embodiment is as follows: Figure 11 As shown. Figure 11 This paper illustrates the correspondence between the comprehensive early warning index, the optimal prediction index for a single indicator, and the occurrence time of high-energy events during the monitoring period from November 15th to November 19th, 2025. Specifically, the red dashed line represents the time of high-energy events around November 17th, 2025; the blue curve represents the trend of the comprehensive early warning index obtained using the multi-source information fusion method of this invention; and the orange curve represents the trend of the optimal early warning index obtained using the single indicator early warning method.

[0125] Depend on Figure 11 It can be seen that before the high-energy event, the blue comprehensive warning index had already begun to rise gradually from the low range, and then rose rapidly near the high-energy event, showing a clear characteristic of increased risk. Although the orange single-indicator warning index also showed a certain increase near the event, its overall amplitude was lower than that of the blue comprehensive warning index, indicating a relatively insufficient risk response intensity. In particular, near the anomaly window, the blue comprehensive warning index was significantly higher than the orange single-indicator warning index, indicating that this invention, by integrating multi-source information such as acoustic emission, electromagnetic radiation, fiber optic strain, and gas concentration, can form a more comprehensive and sensitive characterization of coal and rock mass fracture, stress adjustment, and abnormal gas changes.

[0126] From the perspective of early warning mechanisms, single-indicator early warning methods mainly rely on changes in a certain type of monitoring parameter. When this indicator does not reach a set threshold or is affected by on-site noise or local operating conditions, insufficient early warning response is likely to occur. For example, in this embodiment, the peak value of the CH4 concentration anomaly window is 0.48%, which has not exceeded the on-site concentration threshold of 0.5%. Using a single gas indicator can only provide a weak risk warning. However, this invention jointly analyzes AE, EMR, FBG strain, and CH4 concentration change rate, and further uses a dynamic weighted Bayesian inference method to fuse the identification results of each single source. This allows monitoring sources with higher signal quality and stronger identification confidence to obtain more reasonable weights in the fusion decision, thereby improving the response amplitude and identification reliability of the comprehensive early warning index.

[0127] This invention not only effectively captures multi-source anomaly precursors during the coal and gas outburst incubation process, but also transforms abnormal responses under different physical mechanisms into a continuous, intuitive, and interpretable comprehensive risk characterization quantity. Compared with single-indicator early warning, the comprehensive early warning index of this invention has a more significant amplification effect on dangerous conditions and can more accurately reflect the stage-by-stage evolution trend of the outburst risk level at the working face.

[0128] Furthermore, this invention can unify and convert four heterogeneous monitoring information types—sound, light, electricity, and gas—into a continuous comprehensive early warning index, enabling more obvious and reliable risk response outputs before and after high-energy events. Compared with traditional single-indicator early warning methods, it has a better early warning effect, proving that the method of this invention is effective and superior in dynamic early warning scenarios for coal and gas outbursts.

[0129] The foregoing description, with reference to preferred embodiments, details an exemplary implementation of the coal and gas outburst early warning method and system based on multi-source information fusion proposed in this invention. However, those skilled in the art will understand that various modifications and alterations can be made to the above specific embodiments without departing from the concept of this invention, and various combinations can be made to the various technical features and structures proposed in this invention without exceeding the protection scope of this invention, which is determined by the appended claims.

Claims

1. A method for early warning of coal and gas outbursts based on multi-source information fusion, characterized in that, Includes the following steps: S10: Synchronous acquisition of multi-source monitoring data: Multi-source monitoring sensor units are deployed at the coal mine tunneling face to synchronously acquire multi-source monitoring data during the outburst incubation process; among which, the multi-source monitoring data includes: AE signal, EMR signal, FBG strain signal and CH4 concentration signal; S20: Joint noise filtering and reconstruction of AE and EMR signals: The AE and EMR signals are processed by a joint noise filtering and reconstruction method combining singular value decomposition (SVD) and ensemble empirical mode decomposition (EEMD) to obtain the reconstructed signal. S30: Construction of time-frequency features of multi-source signals: Combining the FBG strain signal and CH4 concentration signal obtained by synchronous acquisition, time-domain, frequency-domain and time-frequency-domain feature analysis are carried out on the multi-source monitoring data respectively to extract the precursor information that can characterize the damage evolution, stress adjustment and gas anomaly changes of coal and rock mass, and construct a multi-dimensional feature expression. S40: Intelligent recognition of single-source precursor features: The obtained AE two-dimensional time-frequency image, EMR two-dimensional time-frequency image, FBG two-dimensional time-frequency image and CH4 two-dimensional time-frequency image are input together with the corresponding time domain feature parameters and frequency domain feature parameters into the intelligent recognition model of single-source precursor features to identify the prominent precursor status corresponding to each single monitoring source. S50: Multi-source information fusion early warning: After obtaining the single-source precursor identification results and corresponding probabilities of the four types of monitoring sources, namely AE signal, EMR signal, FBG strain signal and CH4 concentration signal, a multi-source information fusion method based on dynamic weighted Bayesian inference is further adopted as the core model for multi-source information fusion decision-making. The identification results of different monitoring sources are uniformly expressed, dynamically weighted, fused, inferred and output as decisions to achieve comprehensive identification of coal and gas outburst danger states.

2. The coal and gas outburst early warning method based on multi-source information fusion according to claim 1, characterized in that, Step S20 specifically includes the following steps: S21: For any original signal sequence Let n be the number of sampling points in the current sliding analysis window. Construct its corresponding Hankel matrix according to the phase space reconstruction idea: Where m + c - 1 = n; then, singular value decomposition is performed on the Hankel matrix to obtain: Where U is the left singular vector matrix, V is the right singular vector matrix, and Σ is the singular value diagonal matrix. And satisfy ; S22: Set a threshold based on the singular value mutation characteristics or cumulative energy ratio, determine the effective singular value retention order based on the threshold, and truncate the singular value matrix to obtain the preliminary noise filtering matrix. Then, the initial noise filtering matrix The initial noise-filtered signal is obtained by performing anti-diagonal averaging reconstruction. ; S23: Initial noise filtering signal Performing ensemble empirical mode decomposition yields p intrinsic mode function (IMF) components and residual terms, expressed as follows: in, The i-th intrinsic mode component Residual terms; S24: Quantitative screening of each IMF component is performed using the variance contribution rate. For the i-th intrinsic mode component, its variance contribution rate... Defined as: in, Let represent the variance of the i-th IMF component; based on this, calculate the cumulative variance contribution rate of the q IMF components. : S25: Select the one that satisfies The first q IMF components are used to reconstruct the signal, thereby obtaining the acoustic emission reconstructed signal. and electromagnetic radiation reconstructed signal Where θ is a preset cumulative contribution rate threshold, and the final reconstructed signal is... Represented as: 。 3. The coal and gas outburst early warning method based on multi-source information fusion according to claim 1, characterized in that, Step S30 specifically includes the following steps: S31: Construct time-domain features for multi-source signals, including energy value, peak amplitude, root mean square value, strain change, and gas concentration change rate; where, for discrete signal sequence x i Its energy value E is expressed as: Where n is the number of sampling points in the current analysis window; For the gas concentration sequence C(t), the rate of change of gas concentration is further calculated, and its expression is: Where Δt is the time interval between adjacent sampling times; By extracting time-domain parameters, statistical characteristics of amplitude changes, energy accumulation, and rate of change of various signals within the current time window are obtained. S32: Construct frequency domain features for multi-source signals, perform spectral analysis on the reconstructed AE signal, EMR signal, as well as the FBG strain signal and CH4 concentration signal after unified time window processing, and extract frequency domain feature parameters including peak frequency, center frequency, and centroid frequency to characterize the distribution characteristics of signal frequency components and the variation law of dominant frequency band. If the discrete spectrum is P(f) i If f is the centroid frequency, then f g Represented as: Among them, f i For the i-th discrete frequency point, P(f i () represents the spectral value at the corresponding frequency point; S33: Based on this, further construct time-frequency domain features; use wavelet decomposition and Hilbert-Huang transform methods to perform joint time-frequency analysis on various signals, constructing a two-dimensional time-frequency feature image; during wavelet decomposition, decompose the signal into multiple frequency bands, and calculate the energy proportion of each frequency band as time-frequency feature parameters; by calculating the energy proportion of different frequency bands, the distribution characteristics of signal energy at various scales can be effectively characterized, and a two-dimensional time-frequency feature representation with discriminative significance can be constructed accordingly; where, if the energy of the j-th frequency band is E j If the total number of frequency bands is L, then the energy percentage ρ of the j-th frequency band is... j Represented as: In the formula, E l Energy per frequency band; S34: Finally, the two-dimensional time-frequency images I of AE are obtained respectively. AE EMR two-dimensional time-frequency image I EMR FBG two-dimensional time-frequency image I FBG and CH4 two-dimensional time-frequency image I CH4 Simultaneously, by combining the time-domain and frequency-domain characteristic parameters corresponding to various signals, a structured feature set of multi-source signals is formed.

4. The coal and gas outburst early warning method based on multi-source information fusion according to claim 1, characterized in that, Step S40 specifically includes the following steps: S41: Preprocess the two-dimensional time-frequency images of each single source to ensure that the two-dimensional time-frequency images constructed from different monitoring sources have a unified data expression form before being input into the recognition model; S42: Extract directional gradient histogram features and local binary pattern texture features from each single-source two-dimensional time-frequency image to characterize energy distribution edges, local texture changes, frequency band clustering features, and abnormal evolution patterns in the image; at the same time, incorporate the extracted time-domain feature parameters and frequency-domain feature parameters into the recognition input to form a single-source comprehensive feature that combines image representation capabilities and statistical representation capabilities. S43: For any single-source monitoring source k, where, Its corresponding image descriptor feature vector and statistical eigenvectors They are respectively recorded as: The image descriptor feature vector With statistical eigenvectors The features are concatenated to form a single-source comprehensive feature vector. Among them, Z k This represents the fusion feature vector of single-source monitoring information; S44: Perform principal component analysis to reduce the dimensionality of the comprehensive feature vector. Based on this, input the dimensionality-reduced feature vector into a support vector machine classifier to identify single-source precursor states. Let the dimensionality reduction transformation matrix be P, then the dimensionality-reduced feature vector is represented as: in, The input feature vector is used for the dimensionality-reduced single-source recognition.

5. The coal and gas outburst early warning method based on multi-source information fusion according to claim 4, characterized in that, In step S44, the dimensionality-reduced feature vector is input into the support vector machine classifier to identify single-source precursor states, specifically including the following steps: S441: The support vector machine uses a radial basis kernel function and employs a one-to-many strategy to construct a three-class classification recognition model; to achieve hierarchical recognition of prominent hazardous states, the single-source recognition state set is... Set as: Where w1 represents a safe state, w2 represents a threat state, and w3 represents a dangerous state; For any state category Its support vector machine discriminant function is expressed as: in, Indicates single-source monitoring information regarding status The output value of the discriminant function; N is the number of support vectors; For Lagrange multipliers; Labels for training samples; For kernel functions; For bias terms; Let be the feature vector of the i-th training sample obtained during the training process; S442: Convert the discriminant values ​​output by the support vector machine into probabilistic form; for single-source monitoring information belonging to the state... The recognition probability is expressed as: in, This indicates that the single-source monitoring information belongs to the status category. The recognition probability; l is the summation index; S443: Based on the probability values ​​corresponding to each state category, determine the optimal identification result of the current single-source monitoring information, and obtain the single-source precursor identification results and corresponding probability outputs for the four monitoring sources: AE signal, EMR signal, FBG strain signal, and CH4 concentration signal. This realizes the transformation from multi-source time-frequency feature construction to intelligent identification of single-source states. The optimal identification result... The expression is: 。 6. The coal and gas outburst early warning method based on multi-source information fusion according to claim 1, characterized in that, Step S50 specifically includes the following steps: S51: Construct a multi-source fusion identification framework; the early warning status identification framework is as follows: Where w1 represents a safe state, w2 represents a threat state, and w3 represents a dangerous state; S52: Construct dynamic quality evaluation indicators for any monitoring source k The dynamic quality evaluation index comprehensively considers signal quality, identification confidence, and data integrity. S53: Normalize the dynamic quality evaluation indicators of each monitoring source to obtain the dynamic fusion weight of each single-source monitoring information at the current moment. The expression is as follows: in, q represents the fusion weight of a single monitoring source at time t; r (t) represents the dynamic quality evaluation index of the r-th monitoring source at time t; S54: The recognition probability output from each single-source monitoring information With dynamic weights By combining these factors, a weighted likelihood term is constructed for each monitoring source with respect to different early warning states. Its expression is: Where ε is a positive constant, and its value ranges from 1 × 10⁻⁶. -4 ≤ ε ≤ 1×10 -3 ; Based on this, the prior probability π of each warning state is introduced. j By combining the weighted likelihood terms of the four monitoring sources with Bayesian fusion, the posterior probabilities of the system being in each warning state at the current moment are obtained: in, This represents the set of multi-source monitoring observations corresponding to time t; This indicates that the system is in state w after fusion. j The posterior probability; S55: Introduces a time-recursive smoothing mechanism to smoothly update the fusion results of adjacent time steps. Its expression is: in, This represents the fusion probability after smoothing at time t; It is a smoothing coefficient, and satisfies 0 < γ ≤ 1; Based on the smoothed fusion probability, the state corresponding to the largest probability is taken as the comprehensive warning state at the current moment. This enables the transformation from multi-source single-source identification results to comprehensive state identification results, including comprehensive early warning state. The expression is: S56: Constructing a comprehensive early warning index R t Through the comprehensive early warning index R t The discrete classification results are further transformed into continuous risk characterization quantities to reflect the overall changing trend and phased evolution characteristics of the current prominent hazards at the working face. Among them, the comprehensive early warning index R t The expression is: Where j is the warning status category index, j=1,2,3, corresponding to safe status w1, threat status w2 and dangerous status w3 respectively; λ1, λ2, λ3 are the risk mapping coefficients corresponding to safe, threat and dangerous statuses respectively, and satisfy 1≥λ3>λ2>λ1≥0.

7. The coal and gas outburst early warning method based on multi-source information fusion according to claim 1, characterized in that, It also includes: Step S60: Early warning output and result display: After completing the multi-source information fusion early warning, based on the current comprehensive early warning status and the comprehensive early warning index R t The judgment result triggers the corresponding level of early warning output; at the same time, the software platform of the ground monitoring center displays and records the early warning result in real time.

8. The coal and gas outburst early warning method based on multi-source information fusion according to claim 7, characterized in that, In step S60, based on the current comprehensive early warning status and the comprehensive early warning index R... t The judgment result triggers the corresponding level of early warning output, specifically including the following: When the comprehensive early warning index R t When the judgment range corresponding to the corresponding warning level is reached, and the fusion decision result indicates that the current working face is in a threatened or dangerous state, the corresponding alarm signal and interface prompt information are generated according to the pre-set warning strategy. Among them, the low-level warning is used to indicate that there is a prominent evolution risk in the working face and to prompt the strengthening of monitoring and analysis, while the high-level warning is used to indicate that there is a significant prominent danger in the working face and to trigger response measures such as evacuation, shutdown and anti-outburst disposal.

9. A coal and gas outburst early warning system based on multi-source information fusion, characterized in that, include: The multi-source monitoring data synchronous acquisition module is configured to deploy multi-source monitoring sensor units at the coal mine tunneling face to synchronously acquire multi-source monitoring data during the outburst incubation process; the multi-source monitoring data includes: AE signal, EMR signal, FBG strain signal and CH4 concentration signal; The joint noise filtering and reconstruction module is configured to process the AE signal and EMR signal using a joint noise filtering and reconstruction method that combines singular value decomposition (SVD) and ensemble empirical mode decomposition (EEMD) to obtain the reconstructed signal. The time-frequency feature construction module is configured to combine the synchronously acquired FBG strain signal and CH4 concentration signal to perform time-domain, frequency-domain and time-frequency-domain feature analysis on multi-source monitoring data, extract precursor information that can characterize the evolution of coal and rock mass damage, stress adjustment and gas anomaly changes, and construct multi-dimensional feature expression. The single-source precursor feature recognition module is configured to input the obtained AE two-dimensional time-frequency image, EMR two-dimensional time-frequency image, FBG two-dimensional time-frequency image and CH4 two-dimensional time-frequency image, together with the corresponding time domain feature parameters and frequency domain feature parameters, into the single-source precursor feature intelligent recognition model to identify the prominent precursor state corresponding to each single monitoring source. The multi-source information fusion early warning module is configured to, after obtaining the single-source precursor identification results and corresponding probabilities of the four types of monitoring sources (AE signal, EMR signal, FBG strain signal, and CH4 concentration signal), further employ a multi-source information fusion method based on dynamic weighted Bayesian inference as the core model for multi-source information fusion decision-making. This model unifies, dynamically assigns weights, fuses inferences, and outputs decisions on the identification results of different monitoring sources, thereby achieving comprehensive identification of coal and gas outburst hazard states.

10. The coal and gas outburst early warning system based on multi-source information fusion according to claim 9, characterized in that, The single-source precursor feature recognition module includes: The image preprocessing unit is configured to preprocess each single-source two-dimensional time-frequency image to ensure that the two-dimensional time-frequency images constructed from different monitoring sources have a unified data expression form before being input into the recognition model; The feature extraction module is configured to extract directional gradient histogram features and local binary pattern texture features from each single-source two-dimensional time-frequency image. These features are used to characterize the energy distribution edges, local texture changes, frequency band clustering features, and abnormal evolution patterns in the image. At the same time, the extracted time-domain feature parameters and frequency-domain feature parameters are incorporated into the recognition input, thereby forming a single-source comprehensive feature that combines image representation capabilities and statistical representation capabilities. The single-source integrated feature module is configured to be used for any single-source monitoring source k, wherein, Its corresponding image descriptor feature vector and statistical eigenvectors They are respectively recorded as: The image descriptor feature vector With statistical eigenvectors The features are concatenated to form a single-source comprehensive feature vector. Among them, Z k This represents the fusion feature representation of single-source monitoring information; A single-source precursor identification unit is configured to perform principal component analysis (PCA) dimensionality reduction on the comprehensive feature vector. Based on this, the dimensionality-reduced feature vector is input into a support vector machine (SVM) classifier to identify the state of the single-source precursor. Let the dimensionality reduction transformation matrix be P, then the dimensionality-reduced feature vector is represented as: in, The input feature vector is used for the dimensionality-reduced single-source recognition.

Images