Intelligent Identification and Grading Early Warning Method and System for Underground Gas Geological Anomalies in Coal Mines

Abstract

This invention relates to the field of coal mine gas geological disaster prevention and intelligent identification technology, and discloses a method and system for intelligent identification and graded early warning of underground gas geological anomalies in coal mines. The method includes: S1, collecting multi-source heterogeneous monitoring data in the target area of ​​the coal mine and performing standardization processing and spatiotemporal benchmark unification; S2, constructing a multi-dimensional feature fingerprint database based on historical or accumulated standardized datasets; S3, acquiring real-time multi-source monitoring data for the current tunneling or exploration process and extracting its features; based on the multi-dimensional feature fingerprint database, using evidence theory to fuse and calculate multiple extracted real-time feature evidences to obtain a comprehensive confidence assessment result for the existence of a specific type of gas geological anomaly in the undiscovered area ahead; S4, generating graded early warning information based on the comprehensive confidence assessment result and a preset risk level threshold. This invention achieves advanced identification, risk assessment, and targeted early warning of geological disasters such as underground gas outbursts.

Description

Technical Field

[0001] This invention relates to the field of coal mine gas geological disaster prevention and intelligent identification technology, specifically to a method and system for intelligent identification and graded early warning of coal mine gas geological anomalies. Background Technology

[0002] Coal mine gas outbursts are a major geological hazard threatening underground operational safety. Accurately predicting gas-related geological anomalies (such as faults, tectonic coal seams, karst caves, and high-pressure gas pockets) is crucial for preventing and controlling gas outbursts. Current coal mine gas early warning technologies primarily rely on sensor threshold triggering alarms within safety monitoring systems, which presents several technical challenges.

[0003] First, the problem of data silos is prominent. Geological exploration data (static text, drawings), drilling construction data (manual records), and real-time monitoring data (time series) are independent of each other and lack an effective integration mechanism. This makes it impossible to fully explore the correlation information between multiple data sources, resulting in a single basis for judgment.

[0004] Secondly, false alarms occur frequently and lack in-depth analysis. Alarms based solely on single indicator thresholds such as gas concentration are easily affected by interference from power outages, blasting, etc., resulting in false alarms. Furthermore, the lack of in-depth analysis of the disaster-causing mechanism makes it difficult to distinguish between genuine anomalies and interference signals.

[0005] Third, traditional clustering algorithms have poor adaptability. Coal mine monitoring data has multi-source heterogeneous characteristics. Existing clustering algorithms require pre-setting the number of clusters and cannot dynamically balance the contribution of multi-source data, making it difficult for the clustering results to accurately reflect the gas geological characteristics.

[0006] Fourth, there is a lack of spatial early warning capabilities and adaptive identification mechanisms. Existing systems are mostly limited to "current concentration exceeding the limit alarm" and cannot predict geological anomalies within a certain distance ahead. At the same time, there are difficulties in conflict handling during the fusion of multi-source evidence, and there is a lack of adaptive adjustment capabilities based on on-site verification results, making it difficult to continuously optimize the identification accuracy.

[0007] Fifth, new mining areas suffer from cold start-up and insufficient system robustness. New mining areas or mines lacking historical data struggle to quickly build effective identification models; when key sensors fail or communication is interrupted, the system is prone to paralysis and cannot provide basic early warning support.

[0008] These problems result in insufficient reliability, forward-looking nature, and practicality of existing early warning systems, making it difficult to meet the needs of intelligent coal mine construction for precise prevention and control of gas disasters. Summary of the Invention

[0009] The core objective of this invention is to address the numerous pain points of existing coal mine gas geological anomaly identification and early warning technologies, and to provide an intelligent identification and hierarchical early warning scheme that is both scientific, practical, and robust. Specifically, it is an intelligent identification and hierarchical early warning method and system for gas geological anomalies that integrates multi-source heterogeneous data and is based on advanced clustering algorithms and evidence theory. This system enables advanced identification, risk assessment, and targeted early warning of geological disasters such as underground gas outbursts, and can be widely applied to the intelligent upgrading and transformation of safety production in various coal mines.

[0010] To achieve the above objectives, the following technical solution is adopted:

[0011] In a first aspect, embodiments of the present invention provide an intelligent identification and hierarchical early warning method for underground gas geological anomalies in coal mines, applied to advanced geological exploration and disaster early warning in underground mining faces and roadways, comprising the following steps: S1, collecting multi-source heterogeneous monitoring data in the target area of ​​the underground coal mine, including: static geological exploration data, dynamic engineering operation data, and real-time environmental perception data; standardizing the multi-source heterogeneous monitoring data and unifying the spatiotemporal reference to form a spatiotemporally aligned standardized dataset; S2, based on the historical or accumulated standardized dataset, mining feature clusters representing different gas geological anomaly patterns through adaptive clustering analysis, and combining... S3. Based on the knowledge of geological experts, annotate and construct a multi-dimensional feature fingerprint database of associated feature clusters, anomaly types and causes; S4. For the current tunneling or exploration process, acquire real-time multi-source monitoring data and extract its features; based on the multi-dimensional feature fingerprint database, use DS evidence theory to fuse and calculate multiple extracted real-time feature evidences to obtain a comprehensive confidence assessment result of the existence of a specific type of gas geological anomaly in the unrevealed area ahead; S5. Based on the comprehensive confidence assessment result and the preset risk level threshold, generate and release graded early warning information including early warning level, spatial location of the anomaly, anomaly type, comprehensive confidence level and core evidence.

[0012] Furthermore, in step S1, the multi-source heterogeneous monitoring data is standardized and unified in terms of spatiotemporal reference. Specifically, this includes: transforming unstructured static geological maps and text data into structured data to extract spatial coordinates and attribute information; performing noise reduction processing on time series data such as microseismic signals and drilling rig parameters; establishing a unified downhole three-dimensional spatial coordinate system and millisecond-level time reference, and unifying all data from all sources to the spatiotemporal reference through coordinate back-calculation, positioning calibration, and timestamp alignment.

[0013] Furthermore, in step S2, the adaptive clustering analysis employs an adaptive clustering algorithm based on multiple views and entropy regularization. The specific process includes: defining a multi-view dataset, treating different types of monitoring data as independent views, and assigning dynamically adjustable view weights to each view; constructing a multi-objective optimization function, which includes at least a weighted clustering error term, an entropy regularization term based on clustering ratio, and a synergistic term of membership degree and clustering ratio; iteratively optimizing the multi-objective optimization function while updating sample membership, cluster centers, view weights, and clustering ratios, and automatically merging or eliminating invalid clusters during the iteration process until convergence conditions are met, thereby outputting the final feature cluster partitioning, cluster centers of each cluster, and weights of each data view.

[0014] Furthermore, in step S2, during the iterative process of the adaptive clustering analysis, the calculation of the updated clustering ratio integrates the statistical results of the current sample membership relationship and the entropy adjustment of the clustering ratio in the previous round; the update of the view weights is dynamically allocated based on the combined contribution of the weighted clustering error of each view, the clustering ratio entropy term, and the membership degree-ratio synergy term.

[0015] Furthermore, in step S2, the construction of the multidimensional feature fingerprint database also includes a cold start adaptation mechanism for new mining areas: when the target mining area lacks sufficient local historical data, the initial feature fingerprint database template is migrated from mature mining areas with geological condition similarity exceeding a preset threshold based on the similarity of key geological parameters, and localized parameter correction is performed; and / or, a learning period of a preset duration is initiated, during which the system operates in learning mode, accumulates local initial data and combines it with real-time expert annotation to gradually construct a local feature fingerprint database.

[0016] Furthermore, in step S3, the fusion calculation of multiple extracted real-time feature evidences using DS evidence theory specifically includes: setting a basic probability allocation function for different types of real-time feature evidence; the initial value of the basic probability allocation function is set based on expert experience and can be dynamically optimized according to the mapping relationship trained from historical data; calculating the conflict coefficient K between the multiple evidence sources participating in the fusion; and calculating the conflict coefficient K based on a preset threshold. Based on the comparison results, the corresponding evidence combination rule is selected for fusion: when In this case, a weighted average method is first used to correct each piece of evidence before they are merged; when At that time, the Dempster combination rule is directly used for fusion to obtain the comprehensive confidence score identification result.

[0017] Furthermore, the conflict coefficient K is calculated using the following formula: Where R is the number of evidence sources, Let r be the set of propositions about the type of gas geological anomaly supported by the r-th source of evidence. For the r-th source of evidence, the set of propositions The basic probability allocation; r is the index value of the evidence source, 1≤r≤R; the preset threshold It is 0.7.

[0018] Furthermore, step S3 also includes an adaptive feedback adjustment mechanism to optimize the weights of evidence sources: using the actual geological conditions revealed downhole as feedback labels to evaluate historical identification results; and calculating the contribution of each evidence source r within the sliding time window based on the verification data within that window period. The contribution level Defined as the frequency with which the final identification result matches the actual disclosure when the criterion of the evidence source r is triggered; based on the stated contribution. The weights of each evidence source in subsequent fusion calculations are dynamically adjusted using a Bayesian update strategy. The weight update formula is as follows: In the formula, For the first The updated weights of evidence source r in the next iteration; For the first The weight of evidence source r in the next iteration; The contribution of evidence source r; : Weight iteration count identifier, indicating the current iteration round; Indicates the next iteration round after the update; : Evidence source sequence number replacement identifier, used to traverse all evidence sources for weight normalization calculation.

[0019] Furthermore, the adaptive feedback adjustment mechanism also includes the following overfitting prevention and robustness design: setting a confidence interval for the weight adjustment range of each evidence source, limiting the weight change range to between 0.5 times and 2.0 times its initial value; and / or setting manual intervention trigger conditions: when the cumulative weight adjustment range of a certain evidence source exceeds 50% within 3 consecutive time windows, or when the comprehensive identification accuracy for all warning levels is less than 60%, triggering a system alarm and prompting manual review;

[0020] When some real-time monitoring data sources fail or communication is interrupted, the system automatically switches to a downgraded identification mode, continues to perform evidence fusion and identification based on the remaining valid data sources, and raises the warning level or marks the missing data status and corresponding uncertainty in the generated warning information.

[0021] Furthermore, in step S4, the warning level is divided into four levels: red, orange, yellow, and blue, which correspond to the comprehensive confidence interval from high to low and the handling suggestions for different degrees of urgency, respectively. At the same time, the system outputs an interpretability report of the identification results, which includes at least the contribution analysis of each evidence source and a visual display of the changing trends of key features.

[0022] Secondly, embodiments of the present invention also provide an intelligent identification and hierarchical early warning system for coal mine gas geological anomalies, used to implement the method described in the first aspect, comprising: a data sensing module: deployed underground in the coal mine, used to collect multi-source heterogeneous monitoring data of the target area, including: static geological exploration data, dynamic engineering operation data, and real-time environmental sensing data; performing standardization processing and spatiotemporal benchmark unification on the multi-source heterogeneous monitoring data to form a spatiotemporally aligned standardized dataset; and a fingerprint database construction module: used to mine feature clusters representing different gas geological anomaly patterns based on the historical or accumulated standardized dataset through adaptive clustering analysis, and combined with geological experts' data. Knowledge is labeled to construct a multi-dimensional feature fingerprint database of associated feature clusters, anomaly types, and causes; fusion calculation module: used to acquire real-time multi-source monitoring data for the current tunneling or exploration process and extract its features; based on the multi-dimensional feature fingerprint database, evidence theory is used to fuse and calculate multiple extracted real-time feature evidences to obtain a comprehensive confidence assessment result of the existence of a specific type of gas geological anomaly in the unrevealed area ahead; decision warning module: based on the comprehensive confidence assessment result and preset risk level thresholds, generates and publishes graded warning information including warning level, anomaly spatial location, anomaly type, comprehensive confidence level, and core evidence.

[0023] Compared with the prior art, the present invention achieves the following beneficial effects:

[0024] 1. High adaptability and accuracy of clustering algorithm: The adaptive clustering algorithm based on multiple views and entropy regularization proposed in this invention can adapt to the characteristics of multi-source heterogeneous data in coal mines by dynamically adjusting view weights and balancing cluster size. It is more conducive to accurately mining gas geological anomaly feature clusters and helps to eliminate abnormal data points caused by interference signals such as power outages and blasting, thereby reducing the proportion of invalid feature clusters.

[0025] 2. Highly practical feature fingerprint database: The constructed multi-dimensional feature fingerprint database integrates data features, expert experience and geological genesis, supports dynamic updates and cold start adaptation. New mining areas can quickly build initial models by transplanting data from similar mining areas. Effective identification can be achieved after a 3-6 month basic data collection period, solving the problem of historical data dependence and having a wide range of applications.

[0026] 3. Excellent accuracy and adaptability of fusion identification: The fusion identification technology based on DS evidence theory effectively solves the problem of conflict between multiple sources of evidence. Combined with the adaptive feedback adjustment mechanism, the weight of evidence sources can be continuously optimized according to the on-site verification results. Even when positive samples are scarce, overfitting can be prevented by weighted confidence intervals and manual intervention thresholds. The identification accuracy gradually improves with the expected operating time, and the overall confidence accuracy can reach more than 85%.

[0027] 4. Clearly advanced and targeted early warning: It achieves advanced spatial identification and can predict geological anomalies at a distance of 20 meters or more in advance. The early warning information includes level, coordinates, type, confidence level and core basis. The generated "Outburst Prevention Measures Recommendation Form" clearly defines the operable content such as borehole layout and support adjustment, which can directly guide the on-site outburst prevention work and greatly improve the practicality and feasibility of the early warning.

[0028] 5. High system robustness and security: It has a robust design that can automatically degrade when key sensors fail or communication is interrupted, and continue to provide basic early warnings based on existing effective data; at the same time, it strengthens human-machine collaboration, does not have mandatory control authority, reduces decision-making risks through interpretable design, and takes into account both system reliability and safe production management requirements.

[0029] 6. Meets the needs of intelligent coal mine construction: The overall solution realizes a closed-loop process of data fusion, intelligent identification, hierarchical early warning, and adaptive optimization. The software and hardware co-design is compatible with existing coal mine geological survey management systems, safety monitoring systems, etc. The phased implementation strategy reduces the difficulty of implementation, meets the national coal mine intelligent construction acceptance standards, and has significant value for promotion and application.

[0030] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0031] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0032] Figure 1 This is a flowchart illustrating the intelligent identification and hierarchical early warning method for underground gas geological anomalies in coal mines according to an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram illustrating the principle of the adaptive clustering algorithm in an embodiment of the present invention;

[0034] Figure 3This is a schematic diagram of fusion identification and feedback adjustment based on DS evidence theory in an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of a mobile terminal according to an embodiment of the present invention;

[0036] Figure 5 This is a schematic diagram of the modules of the intelligent identification and hierarchical early warning system for underground gas geological anomalies in coal mines, according to an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments 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. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0039] Figure 1 This is a flowchart illustrating a method for intelligent identification and hierarchical early warning of underground gas geological anomalies in coal mines, according to an embodiment of the present invention. Figure 1 As shown, a method for intelligent identification and hierarchical early warning of underground gas geological anomalies in coal mines is applied to advanced geological exploration and disaster early warning in underground mining faces and roadways, including the following steps:

[0040] S1. Collect multi-source heterogeneous monitoring data in the target area of ​​the coal mine, including: static geological exploration data, dynamic engineering operation data and real-time environmental perception data; standardize the multi-source heterogeneous monitoring data and unify the spatiotemporal reference to form a standardized dataset with spatiotemporal alignment.

[0041] Step S1 is used to achieve standardized acquisition and spatiotemporal calibration of multi-source heterogeneous data, including: structuring unstructured static geological maps and text data to extract spatial coordinates and attribute information; denoising time-series data such as microseismic signals and drilling rig parameters; establishing a unified downhole three-dimensional spatial coordinate system and millisecond-level time reference, and unifying all data from all sources to the spatiotemporal reference through coordinate back-calculation, positioning calibration, and timestamp alignment. Specifically, it includes the following steps:

[0042] S11. Data source classification and hardware adaptation requirements:

[0043] Category A: Static geological data: including fault displacement, coal seam thickness, geological sketches, borehole columnar sections, etc., which are core data reflecting the basic geological conditions of coal mines;

[0044] Category B: Engineering Operation Data: This includes drilling depth, drill cuttings volume (S-value), initial gas emission velocity (q-value) in the borehole, borehole azimuth, etc., which are directly related to geological response information during downhole construction.

[0045] Category C: Real-time sensing data: including gas concentration, wind speed, micro-vibration / acoustic emission signals, drilling rig torque / speed / thrust force, downhole environmental temperature and humidity, etc., which can capture dynamic changes downhole in real time;

[0046] Hardware compatibility requirements: The data sensing layer needs to be equipped with high-precision intelligent drilling rig sensors, with the acquisition accuracy of torque, rotational speed, and thrust reaching ±1%; the positioning error of the micro-vibration sensor should not exceed 5m; the spatial positioning accuracy of the UWB high-precision positioning module should be controlled within ±0.3m. The transmission network adopts an "industrial ring network + 5G / Wi-Fi 6" dual-mode architecture to ensure downhole data transmission latency ≤100ms, guaranteeing the continuity and stability of real-time data. Simultaneously, the hardware configuration must support fault self-diagnosis functions, enabling real-time monitoring of sensor operating status and communication link connectivity.

[0047] S12, Data Processing and Spatiotemporal Calibration:

[0048] S121. Unstructured data transformation:

[0049] To achieve the structured conversion of geological text and graphics, preliminary processing can be performed using mature commercial or open-source tools. For example, the optical character recognition (OCR) function of software such as ABBYY Fine Reader and Adobe Acrobat can be used to extract textual information from paper geological reports; the graphics processing and coordinate registration functions of software such as ArcGIS, AutoCAD, or QGIS (open source) can be used to vectorize geological maps and perform preliminary spatial positioning, thereby realizing the conversion of unstructured data into structured data.

[0050] S122. Time Series Data Processing:

[0051] For time-series data such as microseismic signals and drilling rig operating parameters, a wavelet denoising algorithm is used for noise reduction. The principle of this algorithm is to utilize the multi-resolution analysis characteristics of wavelet functions to decompose the signal into wavelet coefficients of different scales. After thresholding the high-frequency coefficients containing noise, the signal is reconstructed through inverse wavelet transform, thereby effectively preserving the effective precursor information in the data.

[0052] S123, Unification of Spatiotemporal Reference:

[0053] A three-dimensional coordinate system (x, y, z, T) is established underground, where x / y / z are spatial coordinates, calibrated based on permanent guide points in the coal mine; T is a timestamp, accurate to the millisecond level. A "laser scanning + UWB positioning calibration" technology is employed to perform spatiotemporal alignment of data collected from different periods and equipment: static geological data is calibrated by back-calculation using borehole coordinates, while dynamic data, such as drilling rig construction data and microseismic signals, are uploaded in real-time via a UWB module to show equipment location. Combined with the timestamp, this achieves a dual-dimensional "space-time" calibration, ensuring that multi-source data can be fused and analyzed under the same spatiotemporal reference.

[0054] S2. Based on the historical or accumulated standardized dataset, feature clusters representing different gas geological anomaly patterns are mined through adaptive clustering analysis, and labeled in combination with geological expert knowledge to construct a multi-dimensional feature fingerprint database of associated feature clusters, anomaly types and causes.

[0055] Step S2 is used to construct a multidimensional feature fingerprint database of gas geological anomalies.

[0056] First, this embodiment of the invention sets up a cold start data adaptation mechanism: for new mining areas or mines lacking historical data, a dual-mode adaptation scheme of data migration and initial accumulation is designed. For data migration, a characteristic fingerprint database of mature mining areas with similar geological conditions is selected as the initial template. The similarity of geological conditions is calculated based on a weighted similarity score of key geological parameters such as coal seam thickness, dip angle, fault density, and roof and floor lithology. The consistency must be ≥70%. Regional adaptation is performed through a geological parameter correction model to ensure the effectiveness of the initial fingerprint database. For initial data accumulation, the system automatically starts the "basic data collection period" (lasting 3-6 months, which can be adjusted by the user according to the tunneling progress). During this period, the system operates in "learning mode," with its main functions being data accumulation and initial fingerprint database construction. Only reference warnings are provided, focusing on collecting Class B engineering operation data and core Class C real-time data. An initial local dataset is constructed by combining expert annotations, providing a foundation for subsequent identification.

[0057] In a preferred embodiment of the present invention, the construction of the multidimensional feature fingerprint database includes a cold start adaptation mechanism for new mining areas: when the target mining area lacks sufficient local historical data, the initial feature fingerprint database template is migrated from mature mining areas with geological condition similarity exceeding a preset threshold based on the similarity of key geological parameters, and localized parameter correction is performed; and / or, a learning period of a preset duration is initiated, during which the system operates in learning mode, accumulates local initial data and combines it with real-time expert annotation to gradually construct a local feature fingerprint database.

[0058] Then, based on the historical or accumulated standardized dataset, this embodiment of the invention uses adaptive clustering analysis to mine feature clusters representing different gas geological anomaly patterns, and combines this with geological expert knowledge for annotation, constructing a multi-dimensional feature fingerprint database of associated feature clusters, anomaly types, and causes. Specifically, it includes the following steps:

[0059] S21. Feature Extraction and Cluster Analysis:

[0060] An adaptive clustering algorithm based on multiple views and entropy regularization is employed, inspired by the K-means clustering framework, by introducing view weights. Clustering proportional entropy term Membership degree - proportional coherence term (See details below) It realizes three major functions: automatically determining the number of clusters, balancing the cluster size, and dynamically weighting multi-source data. It does not require pre-setting the number of clusters, adapts to the clustering needs and characteristics of multi-source heterogeneous data in coal mines, and accurately mines the characteristic patterns of gas geological anomalies.

[0061] This algorithm constructs a multi-objective optimization function to simultaneously optimize cluster partitioning, cluster centers, view weights, and cluster ratios, achieving collaborative clustering of multi-view data. The core idea is to treat different types of monitoring data as different "views," dynamically allocate view weights to highlight the contribution of effective data, use entropy terms to control the rationality of cluster ratios, and ultimately automatically converge to the optimal cluster structure. This solves the problems of traditional clustering algorithms requiring pre-setting the number of clusters and being difficult to adapt to multi-source data.

[0062] In this invention, the set of samples with similar feature patterns partitioned from the data through adaptive clustering analysis is called a feature cluster. For ease of description, the final feature clusters are sequentially labeled as follows: , , ..., Where c is the total number of clusters automatically determined by the algorithm. Each feature cluster (k=1, 2, ..., c) is composed of its constituent sample sets and cluster centers. and the proportion of this cluster in the total. Common definition.

[0063] Furthermore, in step S21, the specific process of the adaptive clustering algorithm based on multiple views and entropy regularization includes: defining a multi-view dataset, treating different types of monitoring data as independent views, and assigning dynamically adjustable view weights to each view; constructing a multi-objective optimization function, including at least a weighted clustering error term, an entropy regularization term based on cluster ratio, and a synergistic term of membership degree and cluster ratio; iteratively optimizing the multi-objective optimization function, simultaneously updating sample membership, cluster centers, view weights, and cluster ratios, and automatically merging or eliminating invalid clusters during the iteration process until the convergence condition is met, thereby outputting the final feature cluster partition, cluster centers of each cluster, and weights of each data view. In the iterative process of adaptive clustering analysis, the update calculation of the cluster ratio integrates the statistical results of the current sample membership and the entropy adjustment of the cluster ratio in the previous round; the update of view weights is dynamically allocated based on the combined contribution of the weighted clustering error, cluster ratio entropy term, and membership degree-ratio synergistic term of each view. Figure 2 The diagram shown is a schematic representation of the adaptive clustering algorithm according to an embodiment of the present invention. The specific implementation process is as follows:

[0064] Set up a multi-view dataset ,in , View weight and Cluster center binary membership matrix Clustering ratio and The specific algorithm flow is as follows:

[0065] : A multi-view dataset of coal mine gas geology, covering all monitoring data such as underground gas concentration, drill cuttings volume, drilling rig torque, and micro-vibration frequency; The first in the dataset Each sample data corresponds to a single set of underground gas geological monitoring data (multi-dimensional monitoring values ​​of a single time period / single point). Gas Geology Multiview Dataset The total number of sample data, i.e., the total number of downhole monitoring data; : No. The sample at the th The feature data under each view corresponds to a certain type of gas geological feature value in a single set of monitoring data; View number identifier: Each type of monitoring data in the coal mine scenario is an independent view (e.g., gas concentration is 1 view, drill cuttings volume is 1 view), used to distinguish different types of gas geological monitoring data views; The total number of views, i.e., the total number of categories of coal mine gas geological monitoring data; : No. Each view has a feature dimension, which corresponds to the number of feature indicators for a single type of monitoring data. : No. The range of feature data under each view is: 3D real space; : No. The weight coefficients of each view are used to dynamically characterize the contribution of this type of monitoring data to the clustering and identification of gas geological anomalies. : No. The set of all cluster centers under each view corresponds to the core value of the gas geological characteristics cluster in the monitoring data of that class. The total number of clusters, i.e. the number of categories of gas geological feature clusters, can be automatically iterated and optimized by the algorithm to determine the optimal value; A binary membership matrix is ​​used to represent the affiliation relationship between sample data and cluster categories. : Elements belonging to the membership matrix, with values ​​of 0 or 1; Indicates the first The sample belongs to the first Gas-like geological feature clusters Conversely; : No. The cluster ratio of each cluster represents the proportion and weight of that type of gas geological feature cluster in the overall data.

[0066] S211, Initialization Settings:

[0067] Let the initial number of clusters be... (One class per point) In actual engineering implementation, sampling or pre-clustering can be used to set smaller initial values ​​to improve efficiency;

[0068] Clustering ratio Cluster center (Each data point serves as the initial center);

[0069] View weight Balance parameters Iteration threshold View weight index ;

[0070] Define a multi-objective optimization function:

[0071] ;

[0072] in, The number of clusters during algorithm initialization is the initial preset value; The first step during algorithm initialization The clustering ratio is the initial preset value; The first step during algorithm initialization The number of cluster centers is the initial preset value; The first step during algorithm initialization Each view weight is set to an initial preset value. The clustering ratio entropy balance parameter during algorithm initialization is the initial preset value; The membership degree-proportional coordination term balancing parameter during algorithm initialization is the initial preset value; The iteration threshold is a very small positive real number. When the change in cluster centers during iteration is less than this value, the algorithm is considered to have converged and the iteration is terminated. The balance parameter of the clustering ratio entropy term is used to regulate the constraint strength of entropy regularization on the balance of cluster size. : The membership degree-proportional synergy term balance parameter, used to regulate the constraint strength of the clustering ratio on the matching degree of sample membership; View weight index: Used to adjust the distribution range and sensitivity of view weights; it is a positive real constant. The weighted view weights strengthen the weight of high-contribution views and weaken the impact of low-contribution views. : No. The first view The first cluster center Each feature dimension value; The multi-objective optimization function of this algorithm serves as the core basis for calculation. The optimal clustering result is obtained when the function value converges to the minimum value. Clustering error term: Characterizes the sum of weighted Euclidean distances from sample data to the corresponding cluster center, reflecting the fit of the cluster; Clustering proportional entropy term, used to balance the sample size of various gas geological feature clusters, to avoid the clustering results being too sparse or too concentrated; : Membership degree - proportional coordination term, used to match the rationality of sample membership relationship and cluster ratio, and improve the accuracy of gas geological feature clustering; ln: natural logarithm function, the core function for calculating entropy term, used to quantify the disorder of cluster ratio.

[0073] S212, Calculate the membership matrix :

[0074] Based on the current cluster centers, view weights, and cluster proportions, the affiliation is determined by minimizing the objective function. If and only if Minimum, otherwise 0.

[0075] : No. The binary membership matrix after the next iteration represents the association between each gas geological monitoring sample and the feature cluster in this round of iteration; : The core element of the membership matrix, with values ​​of only 0 or 1; Representing the The gas geological monitoring sample belongs to the first Gas-like geological feature clusters This means it does not belong to that category; : The number of iterations in the algorithm For the current iteration round, This refers to the previous iteration round; The core calculation formula for sample attribution determination is used to quantify the first... The sample belongs to the first The matching degree of a feature cluster is calculated by setting a smaller value, indicating a higher matching degree.

[0076] S213, Update balance parameters :

[0077] Employing an empirical attenuation strategy With the number of effective clusters Reduce and gradually weaken This factor causes the algorithm to place more emphasis on distance metrics in its later stages.

[0078] in, : No. The membership degree-proportional synergy term balance parameter after the next iteration is a dynamically updated control coefficient; The natural constant, with a value of approximately 2.71828, is the core constant of the exponential decay strategy. : An empirical exponential decay formula for balance parameters, and a dedicated decay strategy adapted for coal mine engineering; : No. The number of effective clusters after each iteration, i.e. the number of gas geological feature clusters that have converged.

[0079] S214, Update clustering ratio :

[0080] The cluster size can be balanced by adjusting the following formula:

[0081] ;

[0082] : No. After the nth iteration The clustering ratio of gas-like geological feature clusters represents the weight of this type of feature cluster in the overall dataset; : No. After the nth iteration The clustering ratio of gas-like geological feature clusters is the historical value of the previous iteration; The basic term for the sample proportion of the clustering ratio is the first term. The ratio of the number of samples contained in a class feature cluster to the total number of samples; The balancing parameter ratio is used to regulate the magnitude and rate of cluster ratio iteration updates; The entropy summation term of the full cluster represents the overall disorder of the proportion of all gas geological feature clusters and is used to balance the cluster size. The total number of views of gas geological monitoring data (e.g., drill cuttings, gas concentration, drilling rig torque, and micro-vibration frequency are each in one independent view).

[0083] S215, Optimize the number of clusters :

[0084] Removal of cluster ratio For invalid clusters, renormalize the proportion and membership of valid clusters, and effectively control the noise threshold to ≤3%. It is the threshold for invalid clustering, a critical value for eliminating noisy clustering adapted to the coal mine scenario, which is determined by the total number of samples and the dimension of view features.

[0085] S216, Update view weights :

[0086] The formula derived using the Lagrange multiplier method dynamically weights multi-source data, highlighting high-contribution data sources:

[0087] ;

[0088] in, : No. After the nth iteration The weight coefficient of each view dynamically represents the contribution of this type of gas geological monitoring data to the clustering and identification of anomalies. The higher the weight value, the greater the reference value of this type of data. View sequence number is an alternative identifier used to distinguish different monitoring data views and avoid confusion with the main view sequence number. Confusion, range of values ​​and Consistent; : No. The feature dimension of each view, that is, the number of feature indicators of this type of gas geological monitoring data; : No. The sample at the th The first view under the first view Each feature value corresponds to a specific indicator value for a single monitoring data point; : No. The first view The first cluster center of the class feature cluster These characteristic values ​​serve as core reference values ​​for this type of gas geological feature; : No. The weighted sum of clustering errors for each view represents the degree of fit between the monitoring data sample and the corresponding cluster center. : No. The clustering ratio entropy correction term under each view is used to balance the sample size of each gas geological feature cluster within that view; : No. Membership degree-proportion co-correction term under each view, matching the rationality of sample affiliation relationship and cluster proportion; : The index correction factor for view weight, based on the view weight index It is deduced that the sensitivity and differentiation magnitude used to regulate weight allocation are determined. The full-view weight normalization summation term ensures that the sum of the weight coefficients of all views is 1, satisfying the mathematical constraints of weight allocation.

[0089] S217, Update Cluster Centers :

[0090] Calculate the mean of each category based on the new membership matrix:

[0091] ;

[0092] : No. Under the first view, the first After the nth iteration The first cluster center of gas-like geological features Each feature dimension value serves as the core clustering benchmark value for that type of feature; : No. Under the view, belonging to the first All samples of the class feature cluster A weighted summation of 1 feature value, with weights representing sample membership; : No. The summation term of the effective sample count of a feature cluster is the total number of gas geological monitoring samples contained in that feature cluster.

[0093] S218. Convergence Judgment:

[0094] like If the clustering result is not found, then stop iterating and outputting the clustering results; otherwise, let... Return to step S212 and continue iterating.

[0095] : The function to find the maximum value, iterating through the 1st... After the next iteration, select the cluster center with the largest change among all valid cluster centers; : No. Class feature cluster, the first Next and first The Euclidean distance between the cluster centers in each iteration represents the magnitude of the iterative change in the cluster centers; The iteration threshold is a very small positive real constant; when the maximum change in cluster centers is less than this value, the clustering result is considered to be stable. Algorithm iteration count identifier For the current iteration round, This refers to the previous iteration round.

[0096] This process can automatically uncover inherent patterns in the data, forming several "feature clusters," such as "sudden increase in drill cuttings + torque fluctuation cluster" and "nonlinear increase in gas emission + increase in micro-vibration frequency cluster," while effectively eliminating abnormal data points caused by interference factors such as power outages and blasting.

[0097] S22, Expert Annotation and Feature Mapping:

[0098] S221. Introducing the geological expert annotation process:

[0099] A mapping relationship between "cluster clusters" and "geological anomaly types" is constructed to fully reflect the core positioning of "enhancing and extending expert experience." At least three senior geological engineers manually annotate the feature clusters obtained from clustering, clearly defining the geological anomaly type corresponding to each cluster, including faults, tectonic coal formations, karst caves, and high-pressure gas pockets. The annotation process employs a "majority voting" rule, meaning that a consensus reached by more than two-thirds of the experts confirms the annotation result, forming a standardized annotation dataset to ensure the accuracy and authority of the annotation results. Simultaneously, experts can manually supplement feature clusters or adjust feature weights based on actual geological conditions on-site, achieving deep collaboration between machine clustering and human experience.

[0100] S222, Construction of Feature Fingerprint Database:

[0101] The "data feature clusters + expert annotation results + geological interpretation" are linked and stored to form a multi-dimensional feature fingerprint database. This fingerprint database contains 4 types of anomalies and 12 sets of core feature combinations. For example, when identified as a concealed fault with a drop of ≥3m, the core feature combination is a sudden increase in drill cuttings (≥10kg / m), fluctuations in drilling torque (coefficient of variation ≥0.2), an increase in the frequency of microseismic events (≥5 times / h), and a nonlinear increase in gas emission. The geological interpretation is that the fault zone rock mass is fractured, and stress is released after the borehole is exposed, leading to rapid gas emission. When identified as a tectonic coal seam, the core feature combination is a persistently high amount of drill cuttings (≥8kg / m), a decrease in drilling propulsion force (≤5kN), and a sudden change in the initial gas emission velocity q value (≥5L / min). The geological interpretation is that the tectonic coal has a loose structure, low strength, and strong gas adsorption capacity, making it easy to desorb and emit gas after borehole disturbance. The fingerprint database supports dynamic updates and can integrate data adapted to new mining areas, new expert experience, and on-site verification results to continuously enrich the feature dimensions.

[0102] The multidimensional feature fingerprint database constructed in step S2 of this embodiment of the invention mines feature clusters in historical data through the adaptive clustering algorithm in step S21, and establishes the association relationship of "feature cluster-geological anomaly type-cause" by combining geological expert annotation and causal interpretation, providing accurate knowledge support for real-time identification. At the same time, a cold start adaptation mechanism is designed to solve the problem of data scarcity in new mining areas.

[0103] S3. For the current tunneling or exploration process, acquire real-time multi-source monitoring data and extract its features; based on the multi-dimensional feature fingerprint database, use DS evidence theory to fuse and calculate the extracted real-time feature evidence to obtain a comprehensive confidence assessment result for the existence of a specific type of gas geological anomaly in the unrevealed area ahead.

[0104] Step S3 is used to achieve adaptive fusion identification based on DS evidence theory.

[0105] This invention employs an adaptive fusion identification technique based on DS evidence theory, introduces a conflict coefficient correction formula to resolve multi-source evidence conflict issues, determines the basic probability allocation (BPA) through a dual-mode approach of "expert experience + data-driven" and combines a sliding time window and Bayesian update weight adjustment strategy to achieve adaptive optimization of the identification model, thereby improving the accuracy and reliability of multi-source evidence fusion.

[0106] In a preferred embodiment, step S3 employs DS evidence theory to fuse multiple extracted real-time feature evidences. Specifically, this includes: setting a basic probability allocation function for different types of real-time feature evidence; the initial value of the basic probability allocation function is set based on expert experience and can be dynamically optimized according to the mapping relationship trained from historical data; calculating the conflict coefficient K among the multiple evidence sources participating in the fusion; and calculating the conflict coefficient K based on a preset threshold. (For example Based on the comparison results (=0.7), the corresponding evidence combination rule is selected for fusion: when In this case, a weighted average method is first used to correct each piece of evidence before they are merged; when In this case, the Dempster combination rule is directly used for fusion to obtain the comprehensive confidence score. For example... Figure 3 The diagram shown is a schematic representation of the fusion identification and feedback adjustment based on DS evidence theory in an embodiment of the present invention. Specifically, it includes the following steps:

[0107] S31. Determination of the Basic Probability Assignment (BPA) function:

[0108] A dual-mode approach combining expert experience and data-driven analysis was employed to determine the Basic Probability Assignment (BPA) function. In the initial stage, based on geological expert knowledge, an initial BPA was set for each evidence source (such as drill cuttings volume S-value, microseismic signals, and torque fluctuations). For example, the initial confidence level for fault identification when drill cuttings volume S exceeded the threshold was set to 0.35. Subsequently, a mapping model between BPA and characteristic parameter thresholds was established through training with historical data. For instance, when the sudden increase in drill cuttings volume ≥ 30%, the BPA automatically increased to 0.45, achieving dynamic optimization of the BPA.

[0109] S32, Evidence Fusion Calculation:

[0110] S321, Primary Identification:

[0111] Threshold judgment for a single indicator, such as the amount of drill cuttings. At that time, preliminary evidence of "suspected anomaly" is generated. (Drill cuttings volume index) It is one of the core engineering indicators for predicting coal mine gas outbursts, with units of .

[0112] S322, Fusion Identification:

[0113] This paper adopts the Dempster-Shafer Evidence Theory (DS Evidence Theory) and introduces a conflict coefficient correction formula to resolve the problem of conflict among multiple sources of evidence. The principle of this theory is to represent the uncertainty of evidence through a trust function and a likelihood function, fusing information from multiple evidence sources to obtain a comprehensive confidence level. Specifically, DS Evidence Theory is a classic theory for fusing uncertain information from multiple sources, suitable for the identification of gas geological anomalies in coal mines with multiple evidence sources; the trust function in DS Evidence Theory represents the degree of trust in the proposition, quantifying the support of evidence for the identification of gas geological anomalies; and the likelihood function in DS Evidence Theory represents the maximum possible degree of trust in the proposition, quantifying the upper limit of the uncertainty of the evidence.

[0114] Its core formula is as follows: Let the source of evidence be... , ... Their basic probability distributions are as follows: , ... Conflict coefficient The calculation formula is:

[0115] ;

[0116] in, , ... : Multiple evidence sources involved in the fusion, corresponding to different types of gas geological monitoring evidence in the coal mine scenario (such as evidence of abnormal drill cuttings volume, evidence of abnormal microseismic signals, etc.); R: Total number of evidence sources involved in the fusion; , ... The BPA corresponding to each source of evidence represents the confidence assignment value of each piece of evidence for different gas geological anomaly propositions. Conflict coefficient: quantifies the degree of conflict between multiple sources of evidence; its value ranges from [value range missing]. ; Source of evidence The key element is the set of propositions related to the gas geological anomaly supported by this evidence (such as "there is a hidden fault" or "there is a tectonic coal belt"). The intersection of all sources of evidence is an empty set, indicating that the propositions supported by each piece of evidence contradict each other. The product of the basic probability assignments of each source of evidence to the focal element represents the joint confidence level of multiple pieces of evidence simultaneously supporting the corresponding contradictory proposition. The sum of the joint confidence levels of all combinations of contradictory propositions, i.e., the conflict coefficient. The core of its computation. Let r be the set of propositions supported by the r-th source of evidence. For the r-th source of evidence, the set of propositions The basic probability assignment; r is the index value of the evidence source, 1≤r≤R.

[0117] when (in When evidence conflicts arise, a weighted average method is used to correct the evidence before fusion. The weighted average method is a correction method for conflicting evidence, reducing the impact of conflicts on the fusion result by assigning reasonable weights to each source of evidence. In this case, the Dempster combination rule is directly used to calculate the overall confidence level. The Dempster combination rule formula is as follows:

[0118] ;

[0119] in, After fusion, the proposition The basic probability allocation, i.e., the comprehensive confidence level, characterizes the degree of trust in the proposition of gas geological anomalies after the fusion of multiple evidences; G: the target proposition of interest after fusion, such as "there is a hidden fault 20 meters ahead"; An empty set represents a proposition that holds true when there are no gas-related geological anomalies. Normalization factor: Used to eliminate the interference of evidence conflicts on the fusion result and ensure that the sum of the basic probability distributions after fusion is 1; The intersection of all evidence sources is the sum of the joint confidence levels of the target proposition G, representing that multiple pieces of evidence jointly support the proposition. Overall trust level.

[0120] S33, Adaptive feedback adjustment mechanism (optimizing imbalanced sample fitting).

[0121] The adaptive feedback adjustment mechanism in step S33, used to optimize the weights of evidence sources, includes: using the actual geological conditions revealed downhole as feedback labels to evaluate historical identification results; and calculating the contribution of each evidence source r within the sliding time window based on the verification data within that window period. The contribution level Defined as the frequency with which the final identification result matches the actual disclosure when the criterion of the evidence source r is triggered; based on the stated contribution. The weights of each evidence source in subsequent fusion calculations are dynamically adjusted using a Bayesian update strategy. Specifically, the steps include the following:

[0122] S331, Feedback Data Definition:

[0123] The results of the working face excavation, such as the actual fault drop and the type of anomaly, are used as feedback labels and divided into three categories: "correct identification", "deviation in identification", and "incorrect identification".

[0124] S332, Weight Adjustment Strategy:

[0125] A "sliding time window + Bayesian update" algorithm is adopted, with a time window length of 30 days (or 10 tunneling cycles). Evidence source weights are adjusted only based on validation data within the window. The sliding time window is a time range constraint mechanism for weight adjustment, ensuring that weights are updated only based on recently valid validation data; 30 days / 10 tunneling cycles is the specific length of the sliding time window, adapting to the engineering settings of coal mine tunneling operation cycles; Bayesian update is the core logic of weight adjustment, correcting the prior allocation of evidence source weights based on the posterior information of the validation results. The weight update formula is as follows: Let the... In the next iteration, the weight of evidence source r is: The contribution of this evidence source in the verification results is Then the first The weights for the next iteration are:

[0126] ;

[0127] in, Defined as in the first Within a time window, when the source of evidence When the criterion is triggered, the frequency with which the final judgment result matches the actual disclosed situation (i.e., the source of evidence) (accuracy). Weight iteration count identifier For the current iteration round, This is the next iteration round after the update. : No. Source of evidence at the next iteration The weighting coefficient represents the importance of the evidence source in the current round. : No. Source of evidence at the next iteration The updated weight coefficients are based on the optimized importance of the validation results. Source of evidence The contribution, i.e. the first Within a given time window, when the evidence source criterion is triggered, the frequency (accuracy) of the matching result with the actual disclosed situation, with a value range of [value missing]. . The range of values ​​for contribution: 0 represents no contribution at all, and 1 represents complete accuracy. : Evidence source sequence number replacement identifier, used to traverse all evidence sources for weight normalization calculation. : The sum of the weights and contributions of all evidence sources, used for weight normalization to ensure that the sum of the weights of all evidence sources is 1 after the update.

[0128] S333. Prevention and control of sample imbalance and overfitting:

[0129] To address the extreme scarcity of positive samples from coal mine disasters (occurring only once every few years), in addition to using the Synthetic Minority Oversampling Technique (SMOTE) to expand the effective sample size, a dual safeguard mechanism is implemented: First, a confidence interval for weight adjustment is set, limiting the weight variation of each evidence source to the range of [0.5, 2.0] of the initial value. The initial weights are set based on expert experience, and the confidence interval aims to prevent over-adjustment due to a single sample or short-term fluctuations. Second, a manual intervention threshold is set. When the cumulative weight adjustment of an evidence source exceeds 50% within three consecutive time windows, or when the overall accuracy of the judgment for all warning levels is below 60%, the system automatically triggers an expert review reminder, where geological experts assess whether manual weight calibration is necessary to prevent model overfitting.

[0130] As a preferred embodiment of the present invention, the adaptive feedback adjustment mechanism includes a sample imbalance and overfitting prevention design: setting a confidence interval for the weight adjustment range of each evidence source, limiting the weight change range to between 0.5 times and 2.0 times its initial value; and / or setting a manual intervention trigger condition: when the cumulative weight adjustment range of a certain evidence source exceeds 50% within three consecutive time windows, or when the comprehensive identification accuracy for all warning levels is less than 60%, a system alarm is triggered and a manual review is prompted.

[0131] S334, Robust Design:

[0132] When critical sensors (such as microseismic sensors and gas concentration sensors) fail or communication is interrupted, the system automatically activates a degraded operation mode: If some of the Class C real-time sensing data fails, and identification is based solely on Class A static data and Class B engineering operation data, it is recommended to raise the warning level by one level (e.g., upgrade the original yellow warning to an orange warning) and mark it "Some data is missing, and it is recommended to strengthen manual inspection"; if communication is completely interrupted, the local edge computing module saves the critical data and simultaneously issues an equipment fault alarm, pushing it to the maintenance personnel's terminal to ensure that the system can still provide basic early warning support under fault conditions and ensure the safety of downhole operations.

[0133] As a preferred embodiment of the present invention, the adaptive feedback adjustment mechanism includes a robust design: when some real-time monitoring data sources fail or communication is interrupted, the system automatically switches to a degraded identification mode, continues to perform evidence fusion and identification based on the remaining valid data sources, and raises the warning level or marks the data missing status and corresponding uncertainty in the generated warning information.

[0134] S4. Based on the comprehensive confidence level assessment result and the preset risk level threshold, generate and publish graded early warning information including early warning level, spatial location of the anomaly, anomaly type, comprehensive confidence level and core evidence.

[0135] Step S4 is used to implement risk classification and targeted early warning issuance (strengthening human-machine collaborative decision-making).

[0136] In a specific embodiment of the present invention, in step S4, the warning level is divided into at least four levels: red, orange, yellow, and blue, which correspond to the comprehensive confidence interval from high to low and the handling suggestions for different degrees of urgency, respectively; at the same time, the system outputs an interpretability report of the identification results, which includes at least the contribution analysis of each evidence source and a visualization of the changing trends of key features. Figure 4 This is a schematic diagram of a mobile terminal according to an embodiment of the present invention. Step S4 specifically includes the following steps:

[0137] S41, Level 4 Early Warning Model:

[0138] A four-level early warning model is established based on the overall confidence score and the degree of potential hazard. A red warning is issued when the overall confidence score (Score) is ≥ 80%, indicating a high probability of a gas outburst, threatening personnel safety, requiring mandatory production shutdown and activation of the emergency rescue plan. An orange warning is issued when 60% ≤ C < 80%, meaning there is a high probability of geological anomalies and a risk of outburst, requiring a halt to tunneling and the implementation of advanced exploration. A yellow warning is issued when 40% ≤ C < 60%, indicating a suspected anomaly, requiring enhanced monitoring, increasing the monitoring frequency to once every 15 minutes. A blue warning is issued when C < 40%, indicating no significant anomaly risk, requiring continued routine monitoring. All warning levels are for decision-making reference only; the final decision-making authority regarding production shutdowns and tunneling rests with the mine manager, chief engineer, and other on-site management personnel. The system does not have the authority to forcibly control equipment operation, fully reflecting the design principle of "human-machine collaboration and human-led operation."

[0139] S42, Targeted Early Warning and Interpretable Design:

[0140] The early warning information includes five elements: warning level, coordinates of the anomaly (x, y, z), anomaly type, overall confidence level, and core evidence. For example: "Red Warning - A hidden fault with a drop of ≥3m is suspected to exist 20 meters ahead (coordinates X=1250.3m, Y=820.5m, Z=-450.2m), overall confidence level 85%; core evidence: a sudden increase of 35% in drill cuttings (evidence weight 0.32), an increase in the frequency of microseismic events to 8 times / h (evidence weight 0.28), and a drilling rig torque fluctuation coefficient of 0.25 (evidence weight 0.25)".

[0141] The system automatically generates a "Gout Reduction Measures Recommendation Form," specifying actionable content such as "exploratory borehole layout parameters (diameter, depth, and spacing)" and "support scheme adjustment suggestions," and pushes it to the mobile terminals of the mine manager, chief engineer, and on-site team leaders, providing precise guidance for on-site outburst prevention work. Simultaneously, the integrated computing layer outputs an "Evidence Contribution Report," clarifying the influence weight of each evidence source on the identification results and establishing a "feature-anomaly" correlation visualization interface. Engineers can trace the change curves of key features during the identification process, clearly understand the model's decision-making logic, and facilitate making final decisions based on the actual on-site conditions.

[0142] This invention adopts a three-tier architecture: "Data Perception Layer - Fusion Computing Layer - Decision Early Warning Layer." The core logical chain is: "Standardized Collection of Multi-Source Data → Feature Extraction and Expert Annotation Fusion → Adaptive Evidence Fusion Identification → Targeted Hierarchical Early Warning." Step S1 is fundamental, completing the collection, processing, and spatiotemporal unification of multi-source data, providing high-quality data support for subsequent analysis. Step S2, based on the data processed in S1, constructs a fingerprint database containing geological anomaly features and type mappings, forming the core knowledge foundation for intelligent identification. Step S3 utilizes the fingerprint database constructed in S2 to identify anomalies in real-time data through an improved fusion algorithm, a crucial link connecting data and decision-making. Step S4, based on the identification results of S3, performs risk classification and issues targeted early warnings, representing the final implementation of the technical solution. These four steps are progressive and interconnected, incorporating core concepts such as human-machine collaboration, phased implementation, and robust design, forming a complete technical closed loop that is both practical and secure. This invention enables spatial advanced identification and hierarchical targeted early warning of gas geological anomalies, clearly defining the specific coordinates, type, and overall confidence level of the anomalies, generating actionable suggestions for gas outburst mitigation measures, and strengthening the system's robustness design to ensure degraded operation even in the event of critical equipment failure or communication interruption, thus guaranteeing the continuity of early warning. It also strengthens the human-machine collaborative decision-making mechanism, positioning the system as an "enhancement and extension of expert experience," retaining the final decision-making authority of personnel, and enhancing model trust through interpretable design (evidence contribution reports, feature-anomaly correlation visualization), thus adapting to the requirements of coal mine safety production management.

[0143] like Figure 5 The diagram shown is a schematic representation of a module of an intelligent identification and hierarchical early warning system for coal mine gas geological anomalies according to an embodiment of the present invention. This system is used to implement the aforementioned intelligent identification and hierarchical early warning method for coal mine gas geological anomalies, including:

[0144] Data sensing module 210: Deployed underground in coal mines, it is used to collect multi-source heterogeneous monitoring data of the target area, including: static geological exploration data, dynamic engineering operation data and real-time environmental sensing data; and to standardize and unify the spatiotemporal reference of the multi-source heterogeneous monitoring data to form a standardized dataset with spatiotemporal alignment.

[0145] Fingerprint database construction module 220: Based on the historical or accumulated standardized dataset, it mines feature clusters that characterize different gas geological anomaly patterns through adaptive clustering analysis, and combines geological expert knowledge for annotation to construct a multi-dimensional feature fingerprint database with associated feature clusters, anomaly types and causes.

[0146] Fusion calculation module 230: used to acquire real-time multi-source monitoring data for the current tunneling or exploration process and extract its features; based on the multi-dimensional feature fingerprint database, use evidence theory to perform fusion calculation on the extracted real-time feature evidence to obtain a comprehensive confidence judgment result on the existence of a specific type of gas geological anomaly in the unrevealed area ahead;

[0147] Decision-making and early warning module 240: Based on the comprehensive confidence level identification result and the preset risk level threshold, generate and release graded early warning information including early warning level, spatial location of the anomaly, anomaly type, comprehensive confidence level and core evidence.

[0148] The fusion computing module 230 is further configured to execute the adaptive feedback adjustment mechanism, including weighted confidence interval limitation, manual intervention trigger judgment, and downgrade identification mode switching. It should be understood that the specific functions of the adaptive feedback adjustment performed by the fusion computing module 230 correspond to the steps defined in the aforementioned method.

[0149] This system adopts an integrated solution with co-designed hardware and software, requiring seamless compatibility between hardware configuration and software algorithms. The intelligent drilling rig must support real-time sensor data upload interfaces, utilize the Modbus TCP protocol, and possess parameter calibration capabilities. The downhole microseismic monitoring system must synchronize its clock with the fusion computing layer to ensure consistent event timestamps. The software system must be compatible with mainstream coal mine geological survey management systems and safety monitoring system databases, such as SQL Server and MySQL, supporting standardized data interface integration to achieve data interoperability and collaborative operation between systems. Simultaneously, the hardware must be resistant to downhole humidity, dust, and electromagnetic interference, while the software system must support local edge computing and cloud-based collaborative analysis to improve data processing efficiency and system reliability.

[0150] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0151] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A method for intelligent identification and hierarchical early warning of underground gas geological anomalies in coal mines, applied to advanced geological exploration and disaster early warning in underground mining faces and roadways, characterized in that... Includes the following steps: S1. Collect multi-source heterogeneous monitoring data in the target area of ​​the coal mine, including: static geological exploration data, dynamic engineering operation data and real-time environmental perception data; standardize the multi-source heterogeneous monitoring data and unify the spatiotemporal reference to form a standardized dataset with spatiotemporal alignment. S2. Based on the historical or accumulated standardized dataset, feature clusters representing different gas geological anomaly patterns are mined through adaptive clustering analysis, and labeled in combination with geological expert knowledge to construct a multi-dimensional feature fingerprint database of associated feature clusters, anomaly types and causes. S3. For the current tunneling or exploration process, acquire real-time multi-source monitoring data and extract its features; based on the multi-dimensional feature fingerprint database, use DS evidence theory to fuse and calculate the extracted real-time feature evidence to obtain a comprehensive confidence assessment result for the existence of a specific type of gas geological anomaly in the unrevealed area ahead. S4. Based on the comprehensive confidence level identification result and the preset risk level threshold, generate and publish graded early warning information including early warning level, spatial location of the anomaly, type of anomaly, comprehensive confidence level and core evidence. In step S1, the multi-source heterogeneous monitoring data undergoes standardization processing and spatiotemporal benchmark unification, specifically including: Unstructured static geological maps and text data are transformed into structured data to extract spatial coordinates and attribute information; Denoising processing is performed on time-series data such as microseismic signals and drilling rig parameters; Establish a unified downhole three-dimensional spatial coordinate system and a millisecond-level time reference, and unify all data from all sources to the aforementioned spatiotemporal reference through coordinate back-calculation, positioning calibration, and timestamp alignment; In step S2, the adaptive clustering analysis employs an adaptive clustering algorithm based on multiple views and entropy regularization, the specific process of which includes: Define a multi-view dataset, treat different types of monitoring data as independent views, and assign dynamically adjustable view weights to each view; Construct a multi-objective optimization function, which includes at least a weighted clustering error term, an entropy regularization term based on clustering ratio, and a synergistic term of membership degree and clustering ratio; The multi-objective optimization function is iteratively optimized, and the sample membership, cluster center, weight of each view and cluster ratio are updated simultaneously. Invalid clusters are automatically merged or removed during the iteration process until the convergence condition is met, thereby outputting the final feature cluster division, cluster center of each cluster and weight of each data view.

2. The method according to claim 1, characterized in that, In step S2, during the iterative process of the adaptive clustering analysis, the update calculation of the clustering ratio integrates the statistical results of the current sample membership relationship and the entropy adjustment of the clustering ratio in the previous round; the update of the view weight is dynamically allocated based on the combined contribution of the weighted clustering error of each view, the clustering ratio entropy term, and the membership degree-ratio synergy term.

3. The method according to claim 2, characterized in that, In step S2, the construction of the multidimensional feature fingerprint database also includes a cold start adaptation mechanism for new mining areas: When the target mining area lacks sufficient local historical data, the initial feature fingerprint template is migrated from mature mining areas with geological condition similarity exceeding a preset threshold based on the similarity of key geological parameters, and localized parameter correction is performed. And / or, initiate a learning period of a preset duration. During this learning period, the system runs in learning mode, accumulating local initial data and combining it with real-time expert annotations to gradually build a local feature fingerprint database.

4. The method according to claim 1, characterized in that, In step S3, the fusion calculation of multiple extracted real-time feature evidences using DS evidence theory specifically includes: A basic probability allocation function is set for different types of real-time feature evidence; the initial value of the basic probability allocation function is set based on expert experience and can be dynamically optimized according to the mapping relationship trained from historical data. Calculate the conflict coefficient K among the multiple evidence sources involved in the fusion; Based on the conflict coefficient K and the preset threshold Based on the comparison results, the corresponding evidence combination rule is selected for fusion: when In this case, a weighted average method is first used to correct each piece of evidence before they are merged; when At that time, the Dempster combination rule is directly used for fusion to obtain the comprehensive confidence score identification result.

5. The method according to claim 4, characterized in that, The conflict coefficient K is calculated using the following formula: Where R represents the number of evidence sources. Let r be the set of propositions about the type of gas geological anomaly supported by the r-th source of evidence. For the r-th source of evidence, the set of propositions The basic probability allocation; r is the index value of the evidence source, 1≤r≤R; the preset threshold It is 0.

7.

6. The method according to claim 5, characterized in that, Step S3 also includes an adaptive feedback adjustment mechanism to optimize the weights of evidence sources: The actual geological conditions revealed underground are used as feedback labels to evaluate the historical identification results; Based on the verification data within the sliding time window, calculate the contribution of each evidence source r within that sliding time window period. The contribution level Defined as the frequency at which the final identification result matches the actual disclosure when the criterion of the evidence source r is triggered; Based on the contribution The weights of each evidence source in subsequent fusion calculations are dynamically adjusted using a Bayesian update strategy. The weight update formula is as follows: In the formula, For the first The updated weights of evidence source r in the next iteration; For the first The weight of evidence source r in the next iteration; The contribution of evidence source r; : Weight iteration count identifier, indicating the current iteration round; Indicates the next iteration round after the update; : Evidence source sequence number replacement identifier, used to traverse all evidence sources for weight normalization calculation.

7. The method according to claim 6, characterized in that, The adaptive feedback adjustment mechanism also includes the following overfitting prevention and robust design: Set a confidence interval for the weight adjustment range of each evidence source, limiting the weight change range to between 0.5 times and 2.0 times its initial value; and / or, set manual intervention trigger conditions: when the cumulative weight adjustment range of a certain evidence source exceeds 50% within 3 consecutive time windows, or when the comprehensive identification accuracy for all warning levels is less than 60%, trigger a system alarm and prompt manual review. When some real-time monitoring data sources fail or communication is interrupted, the system automatically switches to a downgraded identification mode, continues to perform evidence fusion and identification based on the remaining valid data sources, and raises the warning level or marks the missing data status and corresponding uncertainty in the generated warning information.

8. A coal mine gas geological anomaly intelligent identification and hierarchical early warning system, used to implement the method described in any one of claims 1-7, characterized in that, include: Data sensing module: Deployed underground in coal mines, it is used to collect multi-source heterogeneous monitoring data of the target area, including: static geological exploration data, dynamic engineering operation data and real-time environmental sensing data; it performs standardization processing and spatiotemporal benchmark unification on the multi-source heterogeneous monitoring data to form a spatiotemporally aligned standardized dataset; Fingerprint database construction module: Based on the historical or accumulated standardized dataset, it uses adaptive clustering analysis to mine feature clusters that characterize different gas geological anomaly patterns, and combines geological expert knowledge for annotation to construct a multi-dimensional feature fingerprint database with associated feature clusters, anomaly types and causes. Fusion calculation module: used to acquire real-time multi-source monitoring data for the current tunneling or exploration process and extract its features; based on the multi-dimensional feature fingerprint database, the extraction of multiple real-time feature evidences is fused and calculated using evidence theory to obtain a comprehensive confidence judgment result on the existence of a specific type of gas geological anomaly in the unrevealed area ahead; Decision-making and early warning module: Based on the comprehensive confidence level assessment results and preset risk level thresholds, it generates and publishes graded early warning information that includes early warning level, spatial location of the anomaly, anomaly type, comprehensive confidence level, and core evidence.

Images