A Method for Monitoring Fault Activation Characteristics Based on Multimodal Fusion of Catastrophic Precursor Information

Abstract

This invention relates to the field of mine safety monitoring technology, specifically to a fault activation characteristic monitoring method based on multimodal fusion of disaster precursor information, comprising: sample preparation; system construction; synchronous data acquisition; based on the multi-source data acquired in S3, feature vectors are extracted, and a data fusion algorithm is used to identify the initiation, expansion, and penetration stages of fault activation; based on the feature vectors extracted in S4, normalization processing is performed to construct a fault activation index calculation model, and multi-level early warning thresholds are set according to the FAI value range and combined with the characteristic trend slope; a multimodal fault activation early warning mechanism is established, and the reliability of the early warning results in S5 is verified and presented in multiple dimensions; this application constructs a three-in-one multidimensional monitoring system of acoustic emission, digital image, and parallel electrical resistivity, which synchronously acquires multi-source data such as fracture spatial location, surface strain field, resistivity change, and fracture number, making up for the shortcomings of traditional single monitoring methods that are difficult to comprehensively characterize the complex process of fault activation.

Description

Technical Field

[0001] This invention relates to the field of mine safety monitoring technology, specifically to a fault activation feature monitoring method based on multimodal fusion of disaster precursor information. Background Technology

[0002] Faults, as common adverse geological structures in coal mining and tunnel excavation, are the core cause of dynamic disasters such as water inrush and rock bursts due to their activation process under the influence of mining effects. Most mine water inrush accidents originate from faults that were originally non-conductive. Under the disturbance of mining activities, their internal structure changes and eventually forms a dominant water inrush channel. Fault activation is a complex dynamic process, which is affected by the mining effects and the fault's own characteristics, such as spatial location, occurrence, and the properties of the filling material. Therefore, accurate monitoring of the fault activation process is crucial for ensuring safe production in mines.

[0003] Currently, scholars both domestically and internationally have conducted extensive theoretical analyses, numerical simulations, indoor experiments, and field measurements. However, existing fault activation monitoring technologies still have significant limitations: First, there is insufficient research on the impact of different lithological components in fault fracture zones on the degree of activation, making it difficult to accurately simulate fault response characteristics under actual geological conditions. Second, monitoring methods are limited, relying heavily on single physical parameter monitoring, failing to fully leverage the advantages of electrical resistivity tomography in fracture propagation identification, and lacking precise identification technology for rock fracture images, resulting in a lack of refined description of fault activation characteristics. Third, the hydrogeological conditions in deep mines are complex, and the reliability of single monitoring data is insufficient, making them susceptible to environmental interference and prone to false alarms and missed alarms, failing to meet the engineering requirements for accurate early warning.

[0004] Therefore, constructing a multi-dimensional, multi-parameter, and full-process fault activation monitoring system, and achieving accurate early warning through multi-source information fusion, has become an urgent need for mine water control. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a fault activation feature monitoring method based on multimodal fusion of disaster precursor information, which solves the problems of low precision, single method and insufficient reliability of fault activation monitoring in the prior art.

[0006] To address the above problems, the technical solution of this invention is as follows: a fault activation feature monitoring method based on multimodal fusion of disaster precursor information, comprising the following steps: S1. Sample preparation: Based on the physical and mechanical properties of the target geological fault, rock samples of the fault fracture zone containing preset dip angles and lithological components are prepared using similar materials. S2. System Construction: Establish a fault activation characteristic monitoring and identification system that integrates acoustic emission monitoring module, digital image monitoring module, parallel electrical resistivity monitoring module and data fusion and analysis module. The system uses a rock stress-seepage coupled true triaxial test system as the loading platform. S3. Data Synchronous Acquisition: The sample prepared in S1 is subjected to biaxial loading, and the acoustic emission signal, digital image sequence of the sample surface, resistivity data, and number of cracks are synchronously triggered and acquired, and preprocessed. S4. Based on the multi-source data collected in S3, feature vectors are extracted. The feature vectors include acoustic emission event density, strain concentration, resistivity change rate, and number of fractures per unit area. A data fusion algorithm is used to identify the initiation, expansion, and connection stages of fault activation. S5. Based on the feature vector extracted in S4, normalize it and construct a fault activation index calculation model. Set multi-level early warning thresholds according to the FAI value range and the feature trend slope. S6. For the feature representation images corresponding to the multi-source monitoring data collected in S3, perform spatial fusion and overlay processing, calculate the overlap degree of abnormal areas, establish a multi-modal tomographic activation early warning mechanism, and verify the reliability of the early warning results of S5 and present them in multiple dimensions.

[0007] Furthermore, in S1, the material of the fault fracture zone includes aggregate and cementing agent. The aggregate includes clay and breccia with different particle size distributions. The cementing agent includes cement and gypsum. The particle size of the breccia is graded according to Talbot theory.

[0008] Furthermore, in S3, preprocessing includes standardization, filtering and noise reduction, and trend analysis, among which: The expression for standardization is: ,in, The standardized feature matrix, The original feature matrix, The vector of mean values ​​for each feature. The standard deviation vector of each feature; The filtering and noise reduction uses the moving average method, and the expression is: ,in, Let N be the moving average at time t, and N be the size of the moving window. Let be the eigenvalue at time ti; The trend analysis expression is: slope ,intercept Trend strength: ,in, For the sum of squared residuals, Let t be the total sum of squares, and t be the time variable. Let be the eigenvalue at time t.

[0009] Furthermore, in S4, the feature vector is calculated as follows: Acoustic emission event density ,in, S3 represents the cumulative energy of acoustic emission events within the time window collected, where V is the volume of the monitoring area. The length of the time window; Strain Concentration ,in, The maximum principal strain at (x,y) coordinates at time t is calculated based on the digital image sequence acquired by S3. resistivity change rate ,in, The current resistivity value collected by S3. This is the initial resistivity value; Number of cracks per unit area ,in, The effective number of fractures at time t is the statistical result for S3. The area to be monitored.

[0010] Furthermore, in S4, the data fusion algorithm is a combination of principal component analysis and support vector machine algorithms, including: S41. Principal component analysis is used to reduce the dimensionality of the extracted feature vectors. First, the feature vector matrix is ​​standardized, then the covariance matrix is ​​calculated and eigenvalue decomposition is performed. The top k principal components with larger eigenvalues ​​are selected to calculate the principal component scores. S42. Classify the dimensionality-reduced feature data using a support vector machine, employing a radial basis function kernel: Decision function: ,in, , The feature vector of the sample after dimensionality reduction. For kernel function parameters, For Lagrange multipliers, 'b' represents the sample label, 'b' represents the bias term, and the output shows the identification results of the initiation, expansion, and penetration stages of tomographic activation.

[0011] Furthermore, in S5, the normalization expression is: , in, Let be the normalized value of the feature at time t, and its value ranges from [0,1]. These are the original eigenvalues ​​at time t. This is the historical minimum value of this feature. This is the historical maximum value of this feature.

[0012] Furthermore, in S5, the expression for the fault activation index calculation model is as follows: , in, , , , α, β, γ, and ω are the normalized values ​​of the features corresponding to time t, respectively, and α, β, γ, and ω are weight coefficients that satisfy α+β+γ+ω=1. The weight coefficients are determined by multiple linear regression.

[0013] Furthermore, in S5, the weighting coefficients are determined through multiple linear regression, including: Construct a regression model, the expression of which is: ,in, This is an evaluation value for the measured degree of fault activation. , , , This is the normalized value for the corresponding feature; The weighting coefficients are obtained by solving the normal equation using the least squares method. Regression quality assessment indicators: The model quality was verified using the coefficient of determination (R²), root mean square error (RMSE), and coefficient significance test, where R² ≥ 0.8, RMSE ≤ 0.05, and so on. , , for Statistic, For the standard error of the coefficients, For the first Each weight coefficient.

[0014] Furthermore, in S6, the feature characterization image includes a spatially located thermogram of the acoustic emission event, a strain cloud map calculated based on digital images, a resistivity map based on parallel electrical resistivity inversion, and a crack image of the sample surface. The expression for the overlap degree of the anomalous region is: Where C represents the overlap of abnormal regions. The intersection area of ​​the abnormal regions in the four images. The area of ​​the union of the abnormal regions in the four images is given.

[0015] Furthermore, in S6, the multimodal tomographic activation early warning mechanism is a multimodal collaborative early warning mechanism encompassing reasoning mode, auditory mode, and visual mode, wherein... The inference modality determines the warning level based on the FAI value, the slope of the characteristic trend, and the degree of overlap of abnormal regions. The sound modality triggers different frequencies and volumes of alert sounds based on the warning level; The system synchronously displays the warning level, real-time FAI value, overlap value, multimodal image fusion window, and trend curves of key indicators in the visual modality display. The reasoning, audio, and visual modalities respond synchronously in real time, and achieve deep collaboration through positive verification and reverse error correction mechanisms. Compared with existing technologies, this invention has the following advantages: It constructs a three-in-one multi-dimensional monitoring system integrating acoustic emission, digital imaging, and parallel electrical resistivity tomography, and simultaneously collects multi-source data such as fracture spatial location, surface strain field, resistivity change, and fracture number. This makes up for the shortcomings of traditional single monitoring methods, which are difficult to fully characterize the complex process of fault activation. By reflecting energy release through acoustic emission signals, capturing surface deformation through digital images, and detecting internal fracture development through parallel electrical resistivity tomography, the multi-dimensional data cross-validation achieves comprehensive coverage of the physical and mechanical response of fault activation, and greatly improves the integrity and comprehensiveness of feature capture.

[0016] By extracting four core features—acoustic emission event density (AED), strain concentration (SC), resistivity change rate (RCR), and number of fractures per unit area (FS)—and combining principal component analysis (PCA) dimensionality reduction and support vector machine (SVM) classification algorithms, the nonlinear correlation of features is effectively preserved, and the three key stages of fault activation—initiation, expansion, and connection—are accurately identified, thus achieving a refined and quantitative division of the activation process.

[0017] Based on four core features, a fault activation index (FAI) quantitative model is constructed. Reasonable weight coefficients are determined through multiple linear regression, and the warning threshold is dynamically adjusted by combining the feature trend slope. The model divides the warning levels into three levels: safe, attention, and warning. This model transforms multi-source features into intuitive quantitative indicators, avoiding the subjectivity of traditional warnings that rely on experience-based judgment. At the same time, it predicts the risk escalation trend by using the trend slope, effectively solving the problem of warning lag, and achieving accurate quantitative assessment and efficient warning of fault activation risk. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a tomographic sample diagram of the present invention; Figure 3 This is a flowchart of the feature extraction and fusion process of the present invention; Figure 4 This is a diagram illustrating the multi-source image fusion method of the present invention. Figure 5 This diagram illustrates the multimodal early warning mechanism of the present invention. Detailed Implementation

[0019] This application provides a method for monitoring fault activation characteristics based on multimodal fusion of disaster precursor information. The overall technical approach is as follows: Figure 1As shown, firstly, rock samples containing fault fracture zones with highly simulated target geological conditions are prepared. A multi-dimensional monitoring system integrating acoustic emission, digital imaging, and parallel electrical resistivity tomography is constructed. Multi-source data are collected synchronously during the biaxial loading process of the sample, and four core feature vectors are extracted. The PCA-SVM intelligent algorithm is used to identify the activation stage, and the FAI quantitative index is constructed to achieve risk classification. Finally, through the spatial superposition of multi-source images and the collaborative mechanism of inference-sound-visual three modalities, accurate identification and reliable early warning of the entire fault activation process are achieved.

[0020] A fault activation feature monitoring method based on multimodal fusion of disaster precursor information includes the following steps: S1. Sample preparation: Based on the physical and mechanical properties of the target geological fault, rock samples of the fault fracture zone containing preset dip angles and lithological components are prepared using similar materials. S2. System Construction: Establish a fault activation characteristic monitoring and identification system that integrates acoustic emission monitoring module, digital image monitoring module, parallel electrical resistivity monitoring module and data fusion and analysis module. The system uses a rock stress-seepage coupled true triaxial test system as the loading platform. S3. Data Synchronous Acquisition: The sample prepared in S1 is subjected to biaxial loading, and the acoustic emission signal, digital image sequence of the sample surface, resistivity data, and number of cracks are synchronously triggered and acquired, and preprocessed. S4. Based on the multi-source data collected in S3, feature vectors are extracted. The feature vectors include acoustic emission event density, strain concentration, resistivity change rate, and number of fractures per unit area. A data fusion algorithm is used to identify the initiation, expansion, and connection stages of fault activation. S5. Based on the feature vector extracted in S4, normalize it and construct a fault activation index calculation model. Set multi-level early warning thresholds according to the FAI value range and the feature trend slope. S6. For the feature representation images corresponding to the multi-source monitoring data collected in S3, perform spatial fusion and overlay processing, calculate the overlap degree of abnormal areas, establish a multi-modal tomographic activation early warning mechanism, and verify the reliability of the early warning results of S5 and present them in multiple dimensions.

[0021] The specific steps are as follows: S1. Sample Preparation First, based on the physical and mechanical properties of a typical fault in the target mine, such as a fault dip angle of approximately 70° and a fracture zone infill mainly composed of breccia and clay, a rock-like sample containing a pre-defined fault fracture zone was prepared using the principle of similar material simulation. The sample is as follows: Figure 2 As shown.

[0022] Sample design: The final sample size is 150mm×150mm×300mm. To highlight the key observation points, the internal fault fracture zone is locally enlarged with an inclination angle of 70° and dimensions of 20mm×100mm (width and length).

[0023] Material selection and proportioning: The surrounding rock material uses quartz sand and calcium carbonate as aggregates, and ordinary silicate cement and gypsum as binders. The specific mass ratio is: quartz sand: calcium carbonate: cement: gypsum = 1.87:0.47:0.70:0.30. The amount of water added is 20% of the total mass of aggregates and binders. This ratio makes the mechanical properties of the material close to those of typical sandstone.

[0024] Fault fracture zone material: clay and breccia with different particle size distributions are used as aggregates, and cement and gypsum are used as binders. The specific mass ratio is: breccia: cement: clay: gypsum = 5.5: 1.5: 1.2: 1.8.

[0025] Aggregate gradation: The gradation design of breccia aggregate is carried out based on Talbot's continuous gradation theory. The calculation formula is as follows: ,in, For diameter less than mass fraction of particles, For the first The particle diameter of the group Using a particle size of 6mm and a Talbot index n of 0.4, the calculated particle mass fractions for 0-2mm, 2-4mm, and 4-6mm are 64.4%, 20.6%, and 15%, respectively. Aggregates are mixed in this proportion to ensure that the material has the density and mechanical properties that conform to the heterogeneous characteristics of natural crushing zones.

[0026] Preparation process: The specimens were prepared using a layered compaction method. First, a portion of the surrounding rock was poured. Then, the fault fracture zone material, mixed according to the specified ratio, was precisely placed in the mixture. Finally, the remaining surrounding rock was poured. The specimens were cured under standard curing conditions for at least 28 days to achieve the predetermined strength. The completed fault-containing specimens are shown below. Figure 2 As shown.

[0027] S2, System Construction Loading platform: The system uses a rock stress-seepage coupled true triaxial test system as the core loading platform to apply controllable biaxial loads to the specimen.

[0028] Acoustic emission monitoring module: It adopts the PCI-2 acoustic emission system from Physical Acoustics, Inc., equipped with three acoustic emission sensors with a resonant frequency of 150kHz. The sensors are fixed to the non-observation area of ​​the sample surface with Vaseline coupling agent and are used to collect elastic wave signals generated by micro-fractures inside the sample.

[0029] Digital image monitoring module: High-speed camera, such as PhantomVEO series, with a resolution of not less than 1280×800 pixels and a frame rate of not less than 100fps. Before the test, black and white matte paint is sprayed on the surface of the sample to form a random speckle pattern. The camera is calibrated by calibration plate to establish the correspondence between pixels and physical size.

[0030] Parallel electrical resistivity monitoring module: adopts a Wenner three-stage device, with a power supply voltage of 72V, a minimum effective current of 0.5mA, and a depth coefficient of 0.4. Electrode arrays are pre-embedded on both sides of the sample to collect resistivity data.

[0031] Data fusion and analysis module: It adopts a high-performance computer with built-in custom-developed data acquisition and processing software. It is responsible for receiving and synchronously storing data from the above three sub-modules, and executing subsequent data fusion, feature extraction, model calculation and early warning logic.

[0032] S3, Data Synchronous Acquisition and Preprocessing Synchronous acquisition: The acoustic emission monitoring module, digital image monitoring module, parallel electrical resistivity monitoring module, and data fusion and analysis module are started synchronously. The data acquisition interval is uniformly set to 100ms to achieve time synchronization of multi-source data.

[0033] Acoustic emission data: Collect parameters such as the occurrence time, energy, amplitude, and rise time of acoustic emission events, and record the spatial location coordinates of each event; Digital image data: A high-speed camera continuously acquires image sequences of the sample surface, saving one frame every 100ms to ensure a complete record of the sample surface deformation and crack development process; Resistivity data: Collect resistivity data for zero field, primary field, and secondary field, and calculate resistivity values ​​at different times; Crack count: The number of effective cracks with a length ≥1mm on the sample surface was recorded at each sampling time by combining manual observation with image analysis.

[0034] Data preprocessing includes standardization, filtering and noise reduction, and trend analysis, as detailed below: Standardization processing involves standardizing the raw data, such as acoustic emission event density, strain values, resistivity change rate, and number of fractures, to eliminate dimensional differences. The expression is as follows: ,in, The standardized feature matrix, The original feature matrix, The vector of mean values ​​for each feature. The standard deviation vector of each feature; The filtering and noise reduction uses the moving average method to filter the standardized data, with the moving window size adjusted accordingly. Set to 8 sampling intervals, the expression is: ,in, Let N be the moving average at time t, and N be the size of the moving window. Let be the eigenvalue at time ti; The trend analysis expression is: calculate the linear trend of each characteristic parameter, including: slope. ,intercept Trend strength ,in, For the sum of squared residuals, Let t be the total sum of squares, and t be the time variable. Let be the eigenvalue at time t.

[0035] S4. Feature Vector Extraction and Activation Stage Identification Feature vector extraction, based on preprocessed multi-source data, calculates four core feature vectors time-by-time, such as... Figure 3 As shown: Acoustic emission event density ,in, The cumulative energy of acoustic emission events within the time window collected in S3 is given by V, where V is the volume of the monitoring area, i.e., 150mm × 150mm × 300mm = 6.75 × 10⁻⁶. 6 mm³, The time window length can be set to 1 second, for example, at a certain moment. If the cumulative energy of the acoustic emission event in the first 1 second is 5000 μJ, then AED(t) = 5000 / (6.75 × 10⁶ × 1) ≈ 7.41 × 10⁶ μJ. 4μJ / (mm 3 s); Strain concentration: Reflects the degree of strain localization at the fault tip and between the two disks, and is a sensitive indicator of activation entering the propagation stage. It is calculated by processing the acquired image sequence using a digital image correlation algorithm, determining the principal strain values ​​at each point on the sample surface, and taking the maximum principal strain across the entire field as the strain concentration. The formula is as follows: ,in, The maximum principal strain at (x,y) coordinates at time t is calculated based on the digital image sequence acquired by S3. resistivity change rate ,in, The current resistivity value collected by S3. The initial resistivity value is used as follows: crack propagation - increased porosity - broken conductive path - increased resistivity. This parameter is particularly sensitive to through cracks and can predict macroscopic failure 10-15 minutes in advance. The number of fractures per unit area directly reflects the degree of damage to the fault zone surface and is key evidence to verify whether internal microfractures extend to the surface. The formula is as follows: ,in, The effective number of fractures at time t is the statistical result for S3. The area to be monitored is 150mm × 300mm = 4.5 × 10. 4 mm².

[0036] The activation phase identification employs a data fusion algorithm, specifically a combination of principal component analysis and support vector machine (PCA-SVM) algorithm. S41. Principal Component Analysis (PCA) dimensionality reduction: PCA is used to eliminate multicollinearity among the four features, simplifying computational complexity, as detailed below: Standardization process: ,in, The vector of mean values ​​for each feature. The standard deviation vector is calculated using a sliding window that is dynamically updated, and the most recent 100 samples are used. Covariance matrix calculation: where n is the number of samples; Eigenvalue decomposition: Solving , to obtain eigenvalues and eigenvectors ; Principal component selection: Select the top two principal components with a cumulative contribution rate of ≥85% for each eigenvalue, and calculate the principal component scores. This achieves feature dimensionality reduction.

[0037] Support Vector Machine (SVM) classification: An SVM classification model is constructed using the dimensionality-reduced feature data as input, employing a radial basis function kernel. ,in, The cross-validation method was used to determine the value as 0.8. The tomographic activation stage was divided into three categories: initiation, extension, and penetration, with corresponding sample labels of 0, 1, and 2. 200 sets of sample data from known stages were selected to train the model. After training, real-time feature data was input, and the decision function was used: Output the current activation stage, where, For Lagrange multipliers, 'b' represents the sample label, and 'b' represents the bias term.

[0038] S5. Calculation of Fault Activation Index and Determination of Early Warning Threshold Feature normalization: The extracted four feature vectors (AED, SC, RCR, and FN) are normalized, mapping the feature values ​​to the [0,1] interval. The formula is as follows: ,in, Let be the normalized value of the feature at time t, and its value ranges from [0,1]. These are the original eigenvalues ​​at time t. This is the historical minimum value of this feature. This represents the historical maximum value of this feature. For example, the historical minimum value of an AED is 0, and the historical maximum value is 2 × 10. - ³μJ / (mm³ s), at a certain moment, the original value of the AED is 1×10 - ³μJ / (mm³ s), then after normalization =(1×10 3 0) / (2×10 3 0) = 0.5.

[0039] The weighting coefficients are determined using a multiple linear regression method for the FAI model, as follows: Constructing a regression model: ,in, The measured fault activation level evaluation value is determined based on factors such as fracture development and stress changes, and the value range is [0,1]. , , , This is the normalized value for the corresponding feature; 300 sets of sample data under different activation states were selected, including 50 sets of germination stage, 150 sets of expansion stage, and 100 sets of penetration stage. The normal equation was solved by the least squares method to obtain the weight coefficients. Validate model quality: Calculate the coefficient of determination R² ≥ 0.8, root mean square error RMSE ≤ 0.05, and the values ​​of each weight coefficient. t The absolute values ​​of all statistics are ≥2.3, which meets the model requirements. The final weight coefficients are determined to be α=0.35, β=0.25, γ=0.28, and ω=0.12, which satisfies α+β+γ+ω=1.

[0040] FAI calculation and early warning threshold determination: According to the formula Calculate the real-time FAI value and determine the warning level based on the characteristic trend slope: Safety status: FAI < 0.3, at which point the fault is in a stable state with no risk of activation, and the system does not trigger an early warning; Attention status: 0.3≤FAI<0.6, the fault begins to develop fractures. If the characteristic trend slope a≥0.01 / h, it will be upgraded to an early warning status in advance. Warning status: FAI≥0.6, the fault fracture is rapidly expanding and tending to be connected, the risk is extremely high, and the system triggers a high-level warning.

[0041] S6. Establish a multimodal tomographic activation early warning mechanism. Multi-source image fusion and overlap calculation, such as Figure 4As shown, the spatial localization thermogram of the acoustic emission event, the digital image strain cloud map, the parallel electrical resistivity inversion map, and the sample surface crack image are imported into the data fusion and analysis module for spatial overlay processing to ensure that the coordinate system of the four images is consistent. The overlap degree of the abnormal area is calculated using the following formula: Where C represents the overlap of abnormal regions. The intersection area of ​​the abnormal regions in the four images. The area of ​​the union of the abnormal regions in the four images is considered reliable when C ≥ 70% and FAI is at or above the level of interest.

[0042] Multimodal collaborative early warning, such as Figure 5 As shown: Inference modality: Updated every 100ms, based primarily on FAI value, trend slope a, and overlap C; Sound modality: Different sounds are triggered in real time based on the conclusion of the reasoning modality. There is no prompt sound in the safe state; a low-frequency prompt sound of 500Hz (60-70dB) is played in the attention state; and a high-frequency rapid sound of 1000Hz (80-90dB) is played in the warning state.

[0043] Visual modalities: Simultaneously displayed on the monitoring screen: ① Overview of warning levels, current level, real-time FAI value, and C value; ② Multimodal image fusion window; ③ Key indicators, AED, SC, RCR, and FAI trend curves; ④ Current SVM classification activation stage.

[0044] Collaboration Mechanism: The three modes achieve deep collaboration through trigger synchronization and result mutual verification. For example, when inference triggers a red alert, and the visual display shows a C value as high as 75% and the crack is expanding rapidly, a high-frequency alarm sounds simultaneously, then it is confirmed as a valid alert. If the inference alert is triggered but the visual display shows a C value below 50%, the system will automatically downgrade the alert and prompt manual review. The review results can be used to reverse-calibrate the model parameters.

[0045] A complete technical chain has been constructed, encompassing sample preparation, system integration, synchronous acquisition, intelligent analysis, quantitative early warning, and multimodal presentation. Each step features clearly defined parameters and strong operability. The multi-dimensional monitoring system comprehensively captures the physical and mechanical responses of tomographic activation, overcoming the shortcomings of single-method information loss. The PCA-SVM algorithm effectively processes high-dimensional nonlinear data, achieving recognition accuracy that meets engineering requirements. The FAI quantitative model objectively evaluates the degree of activation. The early warning threshold is dynamically adjustable and highly adaptable. The inference-audio-visual three-modal collaborative mechanism ensures reliable verification and efficient transmission of early warning information, significantly reducing the risk of false alarms and missed alarms. The entire method has undergone thorough experimental verification, with clearly defined performance indicators, and can directly guide field applications and equipment development.

[0046] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for monitoring fault activation characteristics based on multimodal fusion of disaster precursor information, characterized in that, Includes the following steps: S1. Sample preparation: Based on the physical and mechanical properties of the target geological fault, rock samples of the fault fracture zone containing preset dip angles and lithological components are prepared using similar materials. S2. System Construction: Establish a fault activation characteristic monitoring and identification system that integrates acoustic emission monitoring module, digital image monitoring module, parallel electrical resistivity monitoring module and data fusion and analysis module. The system uses a rock stress-seepage coupled true triaxial test system as the loading platform. S3. Data Synchronous Acquisition: The sample prepared in S1 is subjected to biaxial loading, and the acoustic emission signal, digital image sequence of the sample surface, resistivity data, and number of cracks are synchronously triggered and acquired, and preprocessed. S4. Based on the multi-source data collected in S3, feature vectors are extracted. The feature vectors include acoustic emission event density, strain concentration, resistivity change rate, and number of fractures per unit area. A data fusion algorithm is used to identify the initiation, expansion, and connection stages of fault activation. S5. Based on the feature vector extracted in S4, normalize it and construct a fault activation index calculation model. Set multi-level early warning thresholds according to the FAI value range and the feature trend slope. S6. For the feature representation images corresponding to the multi-source monitoring data collected in S3, perform spatial fusion and overlay processing, calculate the overlap degree of abnormal areas, establish a multi-modal tomographic activation early warning mechanism, and verify the reliability of the early warning results of S5 and present them in multiple dimensions.

2. The method for monitoring fault activation characteristics according to claim 1, characterized in that: In S1, the materials of the fault fracture zone include aggregates and cementing agents. The aggregates include clay and breccia with different particle size gradations. The cementing agents include cement and gypsum. The particle size of the breccia is graded according to Talbot theory.

3. The method for monitoring fault activation characteristics according to claim 2, characterized in that: In S3, preprocessing includes standardization, filtering and noise reduction, and trend analysis, among which: The expression for standardization is: ,in, The standardized feature matrix, The original feature matrix, The vector of mean values ​​for each feature. The standard deviation vector of each feature; The filtering and noise reduction uses the moving average method, and the expression is: ,in, Let N be the moving average at time t, and N be the size of the moving window. Let be the eigenvalue at time ti; The trend analysis expression is: slope ,intercept Trend strength: ,in, For the sum of squared residuals, Let t be the total sum of squares, and t be the time variable. Let be the eigenvalue at time t.

4. The method for monitoring fault activation characteristics according to claim 1, characterized in that: In S4, the feature vector is calculated as follows: Acoustic emission event density ,in, S3 represents the cumulative energy of acoustic emission events within the time window collected, where V is the volume of the monitoring area. The length of the time window; Strain Concentration ,in, The maximum principal strain at (x,y) coordinates at time t is calculated based on the digital image sequence acquired by S3. resistivity change rate ,in, The current resistivity value collected by S3. This is the initial resistivity value; Number of cracks per unit area ,in, The effective number of fractures at time t is the statistical result for S3. The area to be monitored.

5. The method for monitoring fault activation characteristics according to claim 4, characterized in that: In S4, the data fusion algorithm includes principal component analysis and support vector machine combination algorithm, comprising: S41. Principal component analysis is used to reduce the dimensionality of the extracted feature vectors. First, the feature vector matrix is ​​standardized, then the covariance matrix is ​​calculated and eigenvalue decomposition is performed. The top k principal components with larger eigenvalues ​​are selected to calculate the principal component scores. S42. Classify the dimensionality-reduced feature data using a support vector machine, employing a radial basis function kernel: Decision function: ,in, , The feature vector of the sample after dimensionality reduction. For kernel function parameters, For Lagrange multipliers, 'b' represents the sample label, 'b' represents the bias term, and the output shows the identification results of the initiation, expansion, and penetration stages of tomographic activation.

6. The method for monitoring fault activation characteristics according to claim 5, characterized in that: In S5, the normalization expression is: , in, Let be the normalized value of the feature at time t, and its value ranges from [0,1]. These are the original eigenvalues ​​at time t. This is the historical minimum value of this feature. This is the historical maximum value of this feature.

7. The method for monitoring fault activation characteristics according to claim 6, characterized in that: In S5, the expression for the fault activation index calculation model is: ， in, , , , α, β, γ, and ω are the normalized values ​​of the features corresponding to time t, respectively, and α, β, γ, and ω are weight coefficients that satisfy α+β+γ+ω=1. The weight coefficients are determined by multiple linear regression.

8. The method for monitoring fault activation characteristics according to claim 7, characterized in that: In S5, the weighting coefficients are determined through multiple linear regression, including: Construct a regression model with the following expression: ,in, This is an evaluation value for the measured degree of fault activation. , , , This is the normalized value for the corresponding feature; The weighting coefficients are obtained by solving the normal equation using the least squares method. Regression quality assessment indicators: The model quality was verified using the coefficient of determination (R²), root mean square error (RMSE), and coefficient significance test, where R² ≥ 0.8, RMSE ≤ 0.05, and so on. , , for Statistic, For the standard error of the coefficients, For the first Each weight coefficient.

9. The method for monitoring fault activation characteristics according to claim 8, characterized in that: In S6, the feature characterization image includes a spatially located thermogram of the acoustic emission event, a strain cloud map calculated based on digital images, a resistivity map based on parallel electrical inversion, and a sample surface crack image. The expression for the overlap degree of the anomalous region is: Where C represents the overlap of abnormal regions. The intersection area of ​​the abnormal regions in the four images. The area of ​​the union of the abnormal regions in the four images is given.

10. The method for monitoring fault activation characteristics according to claim 9, characterized in that: In S6, the multimodal tomographic activation early warning mechanism is a multimodal collaborative early warning mechanism involving reasoning mode, sound mode, and visual mode, wherein... The inference modality determines the warning level based on the FAI value, the slope of the characteristic trend, and the degree of overlap of abnormal regions. The sound modality triggers different frequencies and volumes of alert sounds based on the warning level; The system synchronously displays the warning level, real-time FAI value, overlap value, multimodal image fusion window, and trend curves of key indicators in the visual modality display. The reasoning, sound, and vision modalities respond synchronously in real time and achieve deep collaboration through positive verification and reverse error correction mechanisms.

Images