Multi-parameter collaborative sensing and intelligent alarm system for prominent main geological anomalies

Abstract

This invention relates to the field of risk management in underground engineering, and discloses a multi-parameter collaborative perception and intelligent alarm system for prominent key geological anomalies. The system includes: a data acquisition and preprocessing module, which acquires and preprocesses multi-dimensional parameter data, outputting a multi-parameter data matrix; a feature extraction module, which extracts features from the time-series data using a temporal feature extraction network based on the multi-parameter data matrix, obtaining a temporal feature vector; and simultaneously extracts features from image data using a multi-branch hybrid pooling convolutional neural network, obtaining an image feature vector; a feature fusion module, which constructs a cross-parameter channel correlation matrix through a cross-modal feature channel attention mechanism, outputting a multi-parameter collaborative representation vector; and an alarm generation module, which inputs the multi-parameter collaborative representation vector into an integrated learning anomaly discrimination model, outputting anomaly probability values ​​and generating multi-level alarm signals. This achieves accurate perception and early warning of key geological anomalies, significantly improving alarm accuracy and timeliness.

Description

Technical Field

[0001] This invention relates to the field of risk management technology for underground engineering, and in particular to a multi-parameter collaborative sensing and intelligent alarm system for prominent main geological anomalies. Background Technology

[0002] Controlling geological anomalies (such as fault activity and fold deformation) are the core triggers for underground engineering geological disasters, and their monitoring and early warning are crucial for ensuring the safety of mining, tunnel construction, and other operations. With the development of sensing technology, existing monitoring systems can now collect multi-dimensional parameter data such as microseismic activity, gas levels, stress, and acoustic emission; however, they still face many technical bottlenecks in practical applications.

[0003] 1. Challenges in quality control of multi-parameter data: The downhole environment is complex, and the acquired data is easily affected by equipment noise, electromagnetic interference, and environmental disturbances. Traditional denoising methods (such as filtering and thresholding) are difficult to retain weak anomaly precursors while suppressing noise, resulting in low signal-to-noise ratio and affecting the accuracy of subsequent analysis.

[0004] 2. Lack of Structured Representation of Multi-Parameter Information: The morphology and dimensions of different parameters (time-series data, image data) vary significantly, lacking a unified layout definition and embedding method. This makes it impossible to establish spatial-temporal correlations between parameters, resulting in data exhibiting "isolated" characteristics and making it difficult to uncover collaborative anomaly patterns. For example, geological microseismic signals are temporal, while fault distribution is spatial. Existing heterogeneous data fusion often relies on simple vector concatenation, ignoring the physical coupling of signals at specific geological spatial locations.

[0005] 3. Insufficient precise control of noise regions: The existing attention mechanism is not designed for the layout characteristics of multi-parameter data, and cannot accurately distinguish between effective data regions and noise regions, resulting in the easy loss of effective features and incomplete noise suppression during the denoising process;

[0006] 4. Limitations of Image Feature Extraction: In image data such as geological structure vector maps and time-frequency transformation maps, strong response features (such as obvious fault boundaries) and weak response features (such as subtle texture changes) coexist. The single pooling strategy of traditional convolutional neural networks (CNNs) is difficult to fully capture both types of features, especially for unlabeled downhole images. Downhole geological cloud maps often contain 'weak texture' features (such as tiny fractures), and the max pooling of traditional CNNs tends to retain the strongest features, resulting in the discarding of crucial weak precursor signals.

[0007] 5. Poor synergy of multimodal fusion: Existing fusion methods are mostly simple feature splicing or weighted summation, which fail to learn the coupling mode between parameters in depth (such as the time delay correlation between microseismic frequency and gas concentration, and the intensity synergy between stress and acoustic emission), resulting in limited characterization ability of fused features and difficulty in supporting high-precision anomaly detection.

[0008] These problems prevent existing systems from achieving early and reliable detection of precursors to geological anomalies, resulting in insufficient accuracy and timeliness of alarms, which fails to meet the actual needs of underground engineering safety protection. Therefore, there is an urgent need to build a multi-parameter collaborative sensing and alarm system that integrates advanced feature extraction, intelligent noise reduction, and deep fusion technologies to overcome the aforementioned technical bottlenecks. Summary of the Invention

[0009] In view of the above-mentioned problems in the existing technology, the core objective of this invention is to provide a high-precision and high-reliability multi-parameter collaborative perception and intelligent alarm system to address the problems of poor multi-parameter data denoising effect, inconsistent information representation, incomplete feature extraction, and insufficient fusion and collaboration in existing master-controlled geological anomaly monitoring systems.

[0010] This invention provides a multi-parameter collaborative sensing and intelligent alarm system for prominent controlling geological anomalies, comprising a data acquisition and preprocessing module, a feature extraction module, a feature fusion module, and an alarm generation module connected in sequence, wherein:

[0011] The data acquisition and preprocessing module is used to synchronously acquire multi-dimensional parametric data, including microseismic waveforms, gas concentration, stress / strain data, acoustic emission signals, temperature and pressure data, and geological structure vector images, through a downhole sensor array, and perform noise reduction based on a layout condition diffusion model to output a preprocessed multi-parameter data matrix.

[0012] The feature extraction module is used to extract features from the time-series data using a temporal feature extraction network based on the preprocessed multi-parameter data matrix, to obtain a temporal feature vector; and simultaneously to extract features from the image data using a multi-branch hybrid pooling convolutional neural network (PP-CNN), to obtain an image feature vector.

[0013] The feature fusion module is used to receive temporal feature vectors and image feature vectors from the feature extraction module, construct a cross-parameter channel correlation matrix through a cross-modal feature channel attention mechanism (CFCA), align and fuse heterogeneous features, and output a multi-parameter collaborative representation vector.

[0014] The alarm generation module is used to input the multi-parameter collaborative representation vector into the integrated learning anomaly discrimination model, output anomaly probability values, and generate multi-level alarm signals by comparing the anomaly probability values ​​with preset thresholds.

[0015] Furthermore, the denoising process employs a layout-conditional diffusion model, specifically including:

[0016] The multi-parameter layout embedding unit is used to define each parameter data as a multimodal layout element. It uniformly represents the layout position of time series data and image data through normalized bounding box coordinate vectors, and generates layout control tokens by combining coordinate encoding and category encoding.

[0017] The layout mask attention unit is used to construct a 0 / 1 layout mask matrix, where 1 represents the valid data region and 0 represents the noise / invalid region. In the reverse denoising process of the diffusion model, the attention weight is forced to focus on the valid data region through the mask mechanism, thereby suppressing the feature update of the noise region.

[0018] The dual cross-attention unit is used to fuse global textual semantic conditions that represent prior knowledge of geological anomalies with local layout conditions that represent local layout relationships of parameters. The layout control token is used as input to the local layout conditions to construct a local layout condition vector, and the denoising features are output through weighted cross-attention.

[0019] Furthermore, the operations performed by the multi-parameter layout embedding unit include:

[0020] Each parameter data is defined as a multimodal layout element, where time-series data is represented by a time-amplitude two-dimensional matrix, and image data is represented by a space-pixel value two-dimensional matrix.

[0021] For the effective data region of each type of parameter, an 8-axis bounding box is used. Define its layout position to ensure that the layout information of different parameters is represented in a consistent manner;

[0022] The bounding box coordinates are converted into frequency domain vectors using Fourier coding. Meanwhile, the parameter category code c is obtained through the frozen CLIP text encoder;

[0023] Layout control tokens are generated by fusing coordinate encoding and category encoding through a linear layer. Where Linear(·) is a linear transformation layer, and the layout control token Used to implement structured embedding of multi-parameter layout information.

[0024] Furthermore, the operations performed by the layout mask attention unit include:

[0025] In each denoising step of the diffusion model, the layout mask matrix M is introduced into the attention calculation: attention weights in noisy regions are suppressed through mask multiplication, and noise suppression is further enhanced through a learnable mask bias term matrix, so that the attention weights are concentrated on the effective data regions identified by the mask. The specific calculation formula is as follows:

[0026]

[0027] in, : Layout mask output features of the attention layer; : Feature query vector for the current denoising step; : The layout mask matrix for the i-th type of parameter, whose elements are 0 or 1; Activation function; : No. Key vectors of class parameters; : No. Class parameter key vector The transpose of the matrix; Key vector The dimension; : No. Learnable mask bias term matrix of class parameters; : No. The value vector of the class parameter contains the feature information to be extracted; This indicates element-wise multiplication.

[0028] Furthermore, the dual-cross attention unit performs operations including:

[0029] Text representing geological anomalies is input into a pre-trained language encoder to generate global textual conditional features, which in turn yields global key vectors. and global value vector ;

[0030] The layout control token is processed through a linear projection layer to generate local layout condition features, thereby obtaining local key vectors. and local value vector ;

[0031] The query vector Q and the global key-value pair are calculated using the cross-attention function Ψ. , The first attention result, and Q with local key-value pairs , The second attention result is then used to weight and sum the first and second attention results using a weight coefficient λ, resulting in the final denoised feature output, Out. .

[0032] Furthermore,

[0033] The multi-branch hybrid pooling convolutional neural network includes three feature extraction branches configured in parallel:

[0034] The first branch employs a max pooling strategy and positive interval linear units to extract strong response features from the image. ;

[0035] The second branch employs a minimum pooling strategy and negative interval linear units to extract weak response features from the image. ;

[0036] The third branch employs an interleaved feature fusion strategy, simultaneously capturing strong and weak response features through dual-interval linear units. The output of the dual-interval linear unit is formed by interleaving and splicing the outputs of the positive and negative interval linear units according to the feature dimensions, thus obtaining the strong and weak response features. ;

[0037] The outputs of the three branches are fused using a weighted average or arithmetic average to generate the final image features. : .

[0038] Furthermore, the cross-modal feature channel attention mechanism (CFCA) includes:

[0039] Feature mapping and compression: Adaptive average pooling (AAP) is used to compress each parameter feature into a channel feature vector;

[0040] Channel attention interaction: Perform activation-compression and compression-activation operations on the compressed channel feature vectors to generate channel attention vectors;

[0041] Cross-feature association modeling: Construct a cross-parameter channel correlation matrix, and achieve feature mapping and fusion through selective Softmax normalization and 1-modulus tensor product to generate a multi-parameter collaborative representation vector.

[0042] Furthermore, the channel attention interaction operation includes:

[0043] The channel feature vector of the first parameter is first subjected to an excitation transformation and then a compression transformation. At the same time, the channel feature vector of the second parameter is first subjected to the compression transformation and then the excitation transformation to generate their respective channel attention vectors.

[0044] Furthermore, the alarm generation module includes:

[0045] The anomaly detection unit calculates the multi-parameter collaborative representation vector based on the gradient boosting tree model and outputs anomaly probability values. ;

[0046] The alarm classification unit has at least three preset probability thresholds. The abnormal probability value The alarm signals are compared with the three probability thresholds to trigger different levels of alarm signals, specifically including:

[0047] when The system will trigger an alarm to indicate potential early signs of anomalies in the monitored area.

[0048] when When a warning is triggered, it indicates that the abnormal characteristics are obvious and that enhanced monitoring is needed;

[0049] when An emergency alarm is triggered, indicating that an impending geological anomaly requires immediate preventative measures.

[0050] Furthermore, the alarm generation module also includes:

[0051] The contribution analysis unit is used to calculate the contribution of each parameter to the alarm result based on the weights in the cross-modal feature channel attention mechanism.

[0052] The prevention and control strategy matching unit is used to associate alarm results with geological structures, match them with a preset prevention and control strategy library, and output targeted prevention and control action suggestions.

[0053] Compared with the prior art, the present invention achieves the following beneficial effects:

[0054] 1. This invention achieves noise suppression and anomaly feature preservation in multi-parameter data through a layout conditional diffusion model: by fusing layout embedding, layout mask attention and double cross attention mechanisms, it resolves the contradiction between noise suppression and feature preservation in traditional denoising methods, improves the purity and effectiveness of multi-parameter data, and provides a high-quality data foundation for subsequent analysis;

[0055] 2. This invention proposes a multi-parameter layout definition and structured embedding method: each parameter data is regarded as a "multimodal layout element", and the layout position of temporal data and image data is uniformly represented by an 8-coordinate bounding box. The layout control token is generated by combining Fourier coding and category coding to realize the structured integration of multi-parameter information and establish spatial-temporal correlation between parameters.

[0056] 3. This invention achieves precise control of noisy regions based on the layout mask attention (LMA) mechanism: by constructing a 0 / 1 layout mask matrix, attention weights are guided to be allocated to effective data regions during the diffusion model denoising process, suppressing feature updates in noisy regions, reducing the interference of noise on effective features, and improving the targeting of denoising;

[0057] 4. This invention is based on probabilistic pooling convolutional neural network (PP-CNN) to comprehensively extract deep visual features of images: it integrates three pooling strategies, MaxPooling, MinPooling and MaxMinPooling, and matches them with different ReLU activation functions to capture both strong and weak response features in image data. Through probabilistic fusion, it achieves full-spectrum feature preservation and improves the feature learning ability of unlabeled downhole images.

[0058] 5. This invention is based on the Cross-Modal Feature Channel Attention (CFCA) mechanism to achieve deep collaborative fusion of multimodal features: through feature mapping compression, channel attention interaction and cross-modal feature association modeling, it deeply learns the coupling patterns between multiple parameters, mines the potential correlation information between parameters, and generates collaborative features with strong representation capabilities, providing high-dimensional and effective input for anomaly detection and alarm. Attached Figure Description

[0059] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0060] Figure 1 This is a schematic diagram of the module structure of a multi-parameter collaborative sensing and intelligent alarm system for prominent main geological anomalies provided in an embodiment of the present invention;

[0061] Figure 2 This is a schematic diagram of the system architecture of a multi-parameter collaborative sensing and intelligent alarm system for prominent main geological anomalies provided in an embodiment of the present invention;

[0062] Figure 3 This is a schematic diagram of the layout condition diffusion model (denoising processing core) in an embodiment of the present invention;

[0063] Figure 4 This is a schematic diagram of the probabilistic pooling convolutional neural network (PP-CNN) structure according to an embodiment of the present invention;

[0064] Figure 5 This is a flowchart illustrating the cross-modal feature channel attention mechanism (CFCA) according to an embodiment of the present invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0066] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0067] The acquisition, storage, use, and processing of data in this application comply with relevant national laws and regulations. It should be noted that existing industry solutions such as software, components, or models may be mentioned in the embodiments of this application. These should be considered exemplary and intended only to illustrate the feasibility of implementing the technical solution of this application, but do not imply that the applicant has already used or necessarily used such a solution.

[0068] Figure 1 This is a schematic diagram of the module structure of a multi-parameter collaborative sensing and intelligent alarm system for prominent main geological anomalies provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system architecture of a multi-parameter collaborative sensing and intelligent alarm system for prominent primary geological anomalies provided in an embodiment of the present invention. Figure 1 and Figure 2 As shown, a multi-parameter collaborative sensing and intelligent alarm system 100 for prominent main geological anomalies includes a data acquisition and preprocessing module 110, a feature extraction module 120, a feature fusion module 130, and an alarm generation module 140 connected in sequence, wherein:

[0069] The data acquisition and preprocessing module 110 is used to synchronously acquire multi-dimensional parametric data, including microseismic waveforms, gas concentration, stress / strain data, acoustic emission signals, temperature and pressure data, and geological structure vector images, through a downhole sensor array. It performs noise reduction based on the layout condition diffusion model and outputs the multi-parameter data matrix to the feature extraction module 120.

[0070] Step S1: Multi-parameter acquisition and preprocessing

[0071] Step S1.1: Multi-parameter data acquisition

[0072] The downhole sensor array synchronously acquires microseismic waveforms (sampling rate 1kHz~5kHz), continuous gas concentration (sampling interval 1s), stress / strain gauge data (sampling interval 10s), acoustic emission signals (sampling rate 500Hz~2kHz), downhole temperature and pressure data (sampling interval 1min), and geological structure vector images (resolution 256×256 pixels), covering the multi-dimensional response of geological anomalies in terms of mechanics (microseismic, stress), chemistry (gas), physics (temperature and pressure), and structure (geological vector).

[0073] Step S1.2: Timing Synchronization

[0074] The IEEE 1588 PTPv2 high-precision time synchronization protocol is adopted to control the timestamp error of all sensors within ±500μs. For non-equal interval data (such as temperature, pressure and micro-vibration), linear interpolation and sliding time window alignment are used to generate a multi-parameter time series matrix with a uniform time step (1s) to ensure that the spatiotemporal coupling relationship is traceable.

[0075] Step S1.3: Noise Reduction Processing

[0076] Step S1.3 proposes a layout conditional diffusion model that integrates layout embedding, layout mask attention, and dual-cross attention mechanisms to achieve noise suppression and anomaly feature preservation for multi-parameter data. Furthermore, the denoising layout conditional diffusion model specifically includes:

[0077] The multi-parameter layout embedding unit 1131 is used to define each parameter data as a multimodal layout element, and uniformly represent the layout position of time series data and image data through normalized bounding box coordinate vectors, and generate layout control tokens by combining coordinate encoding and category encoding.

[0078] The layout mask attention unit 1132 is used to construct a 0 / 1 layout mask matrix, where 1 represents the effective data region and 0 represents the noise / invalid region. In the reverse denoising process of the diffusion model, the attention weight is forced to focus on the effective data region through the mask mechanism, thereby suppressing the feature update of the noise region.

[0079] The dual cross-attention unit 1133 is used to fuse the global text semantic conditions that represent prior knowledge of geological anomalies and the local layout conditions that represent the local layout relationship of parameters. The layout control token is used as the input of the local layout conditions to construct the local layout condition vector and outputs denoising features through weighted cross-attention.

[0080] Step S1.3 employs a layout-condition-based denoising diffusion probability model for denoising. This model, in the reverse denoising process of the standard diffusion model, introduces layout conditions to guide noise prediction at each step. Its core lies in a conditional denoising network, whose input includes noisy data. In addition to time step t, it also includes global text conditions. and local layout conditions . Figure 3 This is a schematic diagram of the layout condition diffusion model (denoising processing core) according to an embodiment of the present invention. Figure 3 As shown, the specific process of noise reduction is as follows:

[0081] Step S1.3.1: Define multi-parameter layout and embedding

[0082] Furthermore, the operations performed by the multi-parameter layout embedding unit 1131 include: defining each parameter data as a multimodal layout element, wherein time-series data is represented by a time-amplitude two-dimensional matrix, and image data is represented by a spatial-pixel value two-dimensional matrix; and defining the effective data region of each parameter using an 8-coordinate bounding box. Define its layout position to ensure consistent representation of layout information for different parameters; use Fourier encoding to convert the bounding box coordinates into a frequency domain vector. Simultaneously, the parameter category code c is obtained through the frozen CLIP text encoder; the coordinate code and category code are fused through a linear layer to generate a layout control token. Where Linear(·) is a linear transformation layer, and the layout control token Used to implement structured embedding of multi-parameter layout information.

[0083] Specifically, each parameter data is treated as a "multimodal layout element," where time-series data (microseismic, gas, stress, etc.) are represented by a two-dimensional "time-amplitude" matrix, and image data (geological structure vectors) are represented by a two-dimensional "space-pixel value" matrix. For the effective data region of each parameter type, an 8-coordinate bounding box is used. Define its layout location (such as the effective time period of microseismic data within the time series window, the spatial region of the geological image fault layer) to ensure that the layout information of different parameters is uniformly represented.

[0084] Fourier encoding is used to convert the bounding box coordinates into frequency domain vectors. Simultaneously, a frozen CLIP text encoder is used to obtain parameter category codes (such as "microseismic waveforms," ​​"gas concentration," and "geological faults"). Then, a linear layer is used to fuse the coordinate and category codes to generate the layout control input.

[0085]

[0086] in, The Fourier-encoded layout coordinate vector; Encode the parameter category; Linear ( () is a linear transformation layer.

[0087] Final output As a layout control token, it enables the structured embedding of multi-parameter layout information.

[0088] Step S1.3.2: Noise Region Modulation for Layout Mask Attention (LMA)

[0089] Furthermore, the operations performed by the layout mask attention unit 1132 include: in each step of the denoising process of the diffusion model, introducing the layout mask matrix M into the attention calculation: suppressing the attention weights of the noisy region through mask multiplication, and further strengthening the noise suppression through a learnable mask bias term, so that the attention weights are concentrated on the effective data region identified by the mask.

[0090] Specifically: For each type of parameter data, construct... Layout mask matrix (Valid data area is 1, noise / invalid area is 0), and the category attribute encoding is updated through CLIP encoder. In each denoising step of the diffusion model, the layout mask matrix is... Introducing attention computation for feature updating in noisy regions The constraints are applied, as shown in the following formula:

[0091]

[0092] in, : The output features of the layout mask attention layer, i.e., feature updates of the noisy region; : The feature query vector for the current denoising step, used to retrieve feature information related to the currently processed data; : The layout mask matrix for the i-th type of parameter, whose elements are 0 or 1, where 1 represents the valid data area and 0 represents the noise / invalid area; n: The total number of parameter categories that are monitored and processed synchronously in the system; An activation function is used to normalize the attention score of the input into a probability distribution; The core computation of scaling dot product attention. query vector With the Class parameter key vector The product of the transposes of the matrices (to obtain the attention score). Scaling factor ( (where the key vector dimension is used to alleviate the numerical inflation of attention scores as the dimension increases); : No. The key vector of the class parameter is used to calculate relevance by matching it with the query vector; : No. Class parameter key vector The transpose of the matrix; Key vector The dimension (feature dimension); : No. The learnable mask bias term matrix for the class parameters has the same dimension as the attention score matrix. During initialization, a mask-based differential initialization strategy is employed: for the mask matrix... The location marked as a noise area (value 0) corresponds to... The matrix elements are initialized to large negative values ​​(e.g., -10). For positions in the valid data region (value 1), the corresponding... Initializing matrix elements to 0 is an initialization strategy that helps suppress noisy regions in the early stages of training, while avoiding introducing unnecessary biases into effective data regions. : No. The value vector of the class parameter contains the feature information to be extracted, which will be multiplied by the attention weights to obtain the weighted features. This represents element-wise multiplication (Hadamard product).

[0093] This mechanism uses a mask. Learnable mask bias By guiding the allocation of attention weights, the model prioritizes updating features of effective data regions, reducing the interference of noise on the denoising process.

[0094] Step S1.3.3: Multi-parameter collaborative denoising using dual cross-attention (DCA)

[0095] Furthermore, the dual-cross attention unit 1133 performs operations including: inputting text representing the geological anomaly scene into a pre-trained language encoder, generating global text conditional features, and then obtaining a global key vector. and global value vector The layout control tokens are processed through a linear projection layer to generate local layout condition features, which in turn yield local key vectors. and local value vector The query vector Q and the global key-value pair are calculated using the cross-attention function Ψ. , The first attention result, and Q with local key-value pairs , The second attention result is obtained, and the first attention result and the second attention result are weighted and summed through the weight coefficient λ to obtain the final denoised feature output Out.

[0096] Specifically, to avoid the feature isolation problem caused by single-parameter denoising, this invention introduces a dual cross-attention mechanism to fuse "global text conditions" and "local layout conditions":

[0097] Global text conditions: Generate global control vectors based on text descriptions (such as "multi-parameter data related to downhole fault activity"). This guides the model to learn prior knowledge of geological anomalies;

[0098] Local layout conditions: Based on the layout control token generated in step S1.3.1 Construct layout control vectors Capture the local layout relationships of each parameter.

[0099] The final denoised feature output is fused using a dual-cross attention weighted fusion method:

[0100]

[0101] out: The final denoising feature output by the dual cross-attention mechanism, which integrates feature information from global text conditions and local layout conditions; Cross-attention calculation function, used to implement attention interaction between query vector and key-value vector pairs; : The feature query vector for the current denoising step, used to retrieve features related to global / local conditions; The key vector corresponding to the global text conditions is generated from the text description and is used to encode the feature identifiers of prior knowledge of geological anomalies. The value vector corresponding to the global text conditions contains feature information related to global prior knowledge; Layout condition weighting coefficient, used to balance the contribution of global text conditions and local layout conditions, with a value range of... (Determined through model training and optimization); : Key vectors corresponding to local layout conditions, based on layout control tokens Construct local layout association feature identifiers for encoding each parameter; : The value vector corresponding to the local layout conditions, which contains feature information related to the parameter local layout.

[0102] Step S1.3.4: Perform post-processing verification on the preliminary results output by the denoising network.

[0103] Meanwhile, the loss function of the diffusion model incorporates global and local conditional constraints:

[0104]

[0105] The loss function of the diffusion model is used to optimize model parameters to minimize the difference between predicted noise and actual noise. : Expectation operator, which calculates the overall average of the input variables; Real noise, i.e., the noise component mixed into noisy data (known quantity). : The noise function predicted by the model The model's learnable parameters are inputs. Noisy data at any given moment, time step, global text conditions, and local layout conditions; : Noisy data at any given time (intermediate data at a certain time step during the diffusion process); : Time step index during the diffusion process (used to identify the stage of noise addition / removal); : Global text conditions, a feature vector encoded from text describing geological anomaly scenarios (such as "multi-parameter data related to underground fault activity"); Local layout conditions, based on layout control tokens The generated feature vectors are used to characterize the local layout association information of each parameter; : The norm square operation is used to calculate the square of the Euclidean distance between the actual noise and the predicted noise.

[0106] The training objective of this model is to minimize the loss function. Through iterative training, the network learns to adapt to different layout conditions. Predict and remove noise .

[0107] Furthermore, some embodiments of the present invention also include: layout condition screening and noise reduction post-processing. Specifically: based on the distribution characteristics of real multi-parameter data (such as microseismic frequency range) Gas concentration threshold ), construct normal distribution constraints for physical consistency verification. For each parameter X, calculate the mean of its historical data. and standard deviation Set the filter criteria as Among them, the screening threshold The value is set according to the 3σ principle of normal distribution, with a range of 1.5 to 2.0. This range corresponds to a confidence level of approximately 87% to 95%, which can effectively remove outliers that do not conform to physical laws while preserving the normal fluctuation characteristics of data.

[0108] Step S1.4: Time-frequency transformation

[0109] For time-series data such as microseismic waveforms and acoustic emissions, a short-time Fourier transform (window length 256, overlap rate) is used. The data is converted into a two-dimensional time-frequency graph. For non-stationary stress / strain data, wavelet transform (db4 wavelet basis, decomposition level 5) is used to extract abrupt change features at different scales, providing input for subsequent image feature extraction.

[0110] Feature extraction module 120 is used to extract features from time-series data based on the preprocessed multi-parameter data matrix using a time-series feature extraction network, to obtain a fixed-dimensional time-series feature vector. , dimension Simultaneously, a multi-branch hybrid pooling convolutional neural network (PP-CNN) is used to extract features from the image data, resulting in a fixed-dimensional image feature vector. , dimension Where N is the batch size and D is the feature dimension. Two types of feature vectors. and They have the same feature dimension D to facilitate subsequent feature fusion operations.

[0111] Step S2: Multimodal Feature Extraction

[0112] For different types of monitoring data, a dedicated network is used for feature extraction, especially for unlabeled downhole data to achieve efficient feature learning. The specific steps are as follows:

[0113] Step S2.1: Time series data feature extraction

[0114] For time-series data such as gas concentration, stress / strain, temperature and pressure, a temporal convolutional network (TCN) is used to extract trend features, abrupt change features and periodic features, and key time-series information is preserved through multi-layer convolution and pooling operations.

[0115] Step S2.2: Feature extraction of image data

[0116] This invention proposes a probabilistic pooling convolutional neural network (PP-CNN) to extract deep visual features from image data such as time-frequency transformed two-dimensional feature maps and geological structure vector maps. This network integrates three pooling strategies: MaxPooling, MinPooling, and MaxMinPooling. Figure 4 This is a schematic diagram of the probabilistic pooling convolutional neural network (PP-CNN) structure according to an embodiment of the present invention. Figure 4 As shown, specifically: the probabilistic pooling convolutional neural network includes three parallel feature extraction branches to achieve feature enhancement:

[0117] 1. The first branch (MaxPooling branch) uses a max pooling strategy and positive interval linear units. Its function expression is: This is used to extract strong response features from images;

[0118] 2. The second branch (MinPooling branch) employs a minimum pooling strategy and linear units in the negative interval. Its function expression is as follows: This is used to extract weak response features from images;

[0119] 3. The third branch (MaxMinPooling branch) adopts a strategy that combines max and min pooling. It captures strong and weak response features simultaneously through dual-interval linear units. The output of the dual-interval linear units is generated by interleaving and splicing the outputs of the positive and negative interval linear units according to the feature dimension. ;

[0120] 4. Probabilistic Fusion: The output features of the three branches are probabilistically averaged to obtain the final image feature representation. The formula is as follows:

[0121]

[0122] in, The final feature representation of image data, which integrates feature information from three branches, is used for subsequent collaborative fusion and anomaly detection; The feature output of the MaxPooling branch includes strong response features in image data (such as high amplitude regions in time-frequency maps and obvious structural boundaries in geological maps). The feature output of the MinPooling branch includes weak response features in image data (such as low amplitude fluctuations in time-frequency maps and subtle texture changes in geological maps). The feature output of the MaxMinPooling branch is concatenated with the MaxPooling and MinPooling results through interleaving, preserving the full spectrum features (while also covering the correlation information between strong and weak response features).

[0123] This method effectively preserves positive and negative correlation features and improves the feature learning ability of unlabeled image data.

[0124] The feature fusion module 130 is used to receive temporal feature vectors and image feature vectors from the feature extraction module 120, construct a cross-parameter channel correlation matrix through the cross-modal feature channel attention mechanism (CFCA), align and fuse heterogeneous features, and output a multi-parameter collaborative representation vector to the alarm generation module 140.

[0125] Furthermore, the cross-modal feature channel attention mechanism includes: feature mapping and compression: compressing each parameter feature into a channel feature vector through adaptive average pooling (AAP); channel attention interaction: performing activation-compression and compression-activation operations on the compressed channel feature vectors to generate channel attention vectors; cross-feature association modeling: constructing a cross-parameter channel correlation matrix, and achieving feature mapping and fusion through selective softmax normalization and 1-modulus tensor product to generate a collaborative representation.

[0126] Step S3: Multi-parameter feature collaborative fusion

[0127] Step S3 proposes a cross-modal feature channel attention mechanism (CFCA) to learn the coupling patterns between multiple parameters (such as the time delay relationship between the rise in microseismic frequency and the sudden increase in gas, and the correlation between stress changes and the intensity of acoustic emission signals), thereby achieving deep collaborative fusion of multi-parameter features. Figure 5 This is a flowchart illustrating the Cross-Modal Feature Channel Attention (CFCA) mechanism according to an embodiment of the present invention, as shown below. Figure 5 As shown, the specific steps are as follows:

[0128] Step S3.1: Feature Mapping and Compression

[0129] Output the characteristics of each parameter (Time series characteristics) or (Image features) or Channel dimension compression is performed, and channel feature vectors are obtained through adaptive average pooling (AAP). and .in, These represent the number of channels with different parameter characteristics; For timing length, As a dimension of temporal features, The height of the image feature map. This represents the width of the image feature map.

[0130] Step S3.2: Channel Attention Interaction

[0131] The channel attention interaction operation includes: performing an activation transformation followed by a compression transformation on the channel feature vector of the first parameter, and simultaneously performing a compression transformation followed by the activation transformation on the channel feature vector of the second parameter, to generate their respective channel attention vectors. Specifically:

[0132] Perform activation-compression and compression-activation operations on the compressed channel feature vectors to generate channel attention vectors. and :

[0133]

[0134]

[0135] in, The weight matrix is ​​a linear transformation matrix; Use the Sigmoid activation function; The first type of parameter is the channel attention vector, which is used to characterize the importance weights of each channel feature; The Sigmoid activation function has the following formula: Used to map input values ​​to The interval is used to output attention weights; The linear transformation weight matrix for the first type of parameter "compression" stage, with dimension 1. This is used to compress the excited high-dimensional vector back to the original channel dimension; The linear transformation weight matrix for the first type of parameter, the "excitation" phase, has dimensions of... This is used to map the compressed channel vectors to a high-dimensional space to enhance expressive power. : Activation functions are used to introduce nonlinearity and enhance the model's feature learning ability; : The channel feature vector (input value) of the first type of parameter after AAP compression; The second type of parameter is the channel attention vector, which is used to characterize the importance weights of each channel feature; The linear transformation weight matrix for the second type of parameter, the "excitation" phase, has dimensions of... This is used to map the compressed channel vectors to a high-dimensional space; The linear transformation weight matrix for the second type of parameter "compression" stage, with dimension [missing information]. This is used to compress the excited high-dimensional vector back to the original channel dimension; The second type of parameter is the channel feature vector (input value) after AAP compression.

[0136] Step S3.3: Cross-feature association modeling

[0137] Constructing cross-parameter channel correlation matrix After selective softmax normalization, feature mapping and fusion are achieved using 1-modulus tensor product:

[0138]

[0139] Wherein, B: cross-parameter channel correlation matrix, used to characterize the correlation strength between the two types of parameter channel attention vectors; : Channel attention vector of the first type of parameter (dimension) This includes the importance weights of each channel for this type of parameter; Second type of parametric channel attention vector transpose (dimension) ), after transpose and Multiplication yields the dimension as The correlation matrix; : Selective Softmax normalization function, which normalizes the association matrix by columns (or rows, adaptively based on the parameter channel dimension) to convert the association strength into a probability distribution; The features of the first type of parameter after mapping to the second type of parameter, i.e., the features of the first type of parameter. The feature space of the second type of parameters is adapted after weighting by the correlation matrix. ; :1 - Modular tensor product operation, used to apply the weights of the correlation matrix to the original features of the first type of parameters. The channel dimension enables feature mapping; Transpose of the normalized incidence matrix (dimension) ), used to map the features of the first type of parameters to the channel space of the second type of parameters; The features of the second type of parameter after mapping to the first type of parameter, i.e., the features of the second type of parameter. The feature space of the first type of parameter is adapted by weighting the correlation matrix. .

[0140] The final fusion features are:

[0141]

[0142] The final fusion feature of the first type of parameter incorporates its original features. Features mapped from the second type of parameters ; The final fusion feature of the second type of parameter incorporates its original features. Features mapped from the first type of parameter .

[0143] A multi-parameter collaborative representation is obtained through iterative fusion (multiple cross-mapping superpositions). This provides high-dimensional and effective features for alarm identification.

[0144] The alarm generation module 140 is used to input the multi-parameter collaborative representation vector into the integrated learning anomaly discrimination model, output the anomaly probability value, and generate a multi-level alarm signal by comparing the anomaly probability value with a preset threshold.

[0145] Step S4: Anomaly Detection and Multi-Level Alarm Generation

[0146] Furthermore, the alarm generation module 140 includes:

[0147] Anomaly detection unit 141 calculates the multi-parameter collaborative representation vector based on the gradient boosting tree model and outputs anomaly probability values. Alarm classification unit 142 has at least three preset probability thresholds. , the abnormal probability value The alarm signal is triggered at different levels based on a comparison with three probability thresholds. Specifically:

[0148] Co-representation The input is fed into an anomaly detection model based on gradient boosting trees (XGBoost), which learns the boundary between normal and abnormal states through training data and outputs anomaly probability values. Based on the anomaly probability value and a preset threshold, a three-level alarm is generated:

[0149] 1. Alarm / Notification: When The system will trigger an alarm to indicate potential early signs of anomalies in the monitored area.

[0150] 2. Warning: When When a warning is triggered, it indicates that the abnormal characteristics are obvious and that enhanced monitoring is needed;

[0151] 3. Emergency Alarm: When An emergency alarm is triggered, indicating that an impending geological anomaly requires immediate preventative measures.

[0152] in For alarm thresholds determined through statistical learning (e.g.) ).

[0153] Furthermore, the alarm generation module 140 also includes: a contribution analysis unit 143, used to calculate the contribution of each parameter to the alarm result based on the weights in the cross-modal feature channel attention mechanism; and a prevention strategy matching unit 144, used to associate the alarm result with the geological structure, match it with a preset prevention strategy library, and output targeted prevention action suggestions. Specifically:

[0154] Based on the channel attention weights of the Cross-Modal Feature Channel Attention (CFCA) module, the contribution score of each parameter to the alarm result is calculated. (As a parameter type), output the core parameters that support the alarm; associate the alarm results with the geological structure vector data, map them to specific geological units, and mark the main geological body (such as a fault segment or fold core) that is most likely to cause the anomaly.

[0155] Based on the type of the main geological body and the alarm level, a preset prevention and control strategy library is matched to output targeted exploration / prevention actions, such as local borehole monitoring, rock mass reinforcement, isolation support, and enhanced gas extraction.

[0156] The multi-parameter collaborative sensing and intelligent alarm system for prominent primary geological anomalies provided by the above embodiments of the present invention achieves intelligent denoising of multi-parameter data through a layout conditional diffusion model, comprehensively extracts strong / weak response features of images using a probabilistic pooling convolutional neural network, and deeply mines multi-parameter coupling relationships based on a cross-modal feature channel attention mechanism. This effectively solves the problems of low signal-to-noise ratio, incomplete feature extraction, and poor multimodal fusion collaboration in traditional monitoring systems, and ultimately achieves accurate sensing and early warning of primary geological anomalies, significantly improving alarm accuracy and timeliness.

[0157] Test results: By deploying the system in a real downhole monitoring environment and conducting continuous testing for 6 months, more than 10TB of multi-parameter monitoring data were collected. In comparative tests with traditional methods, the system of this invention exhibits significant advantages: Denoising effect: Under the same noise environment, after processing by the layout condition diffusion model, the average signal-to-noise ratio (SNR) of multi-parameter data is improved by 15-25 dB compared with the traditional wavelet threshold denoising method. At the same time, through verification of known anomaly markers, the retention rate of anomaly waveform features exceeds 95%; Feature extraction: Tested on a dataset containing 5000 labeled downhole images, the probabilistic pooling convolutional neural network (PP-CNN) achieves a recall rate of 92% for weak response features (such as subtle crack textures in images), which is about 30 percentage points higher than the 62% of the traditional VGG16 network; Alarm performance: Based on the collaborative representation fusion of cross-modal feature channel attention mechanism, in the backtesting of historical disaster data, the accuracy of the anomaly detection model reaches 98.5%, and the false alarm rate is reduced to below 2.1%. Compared with traditional methods based on single-parameter analysis, it issues effective alarms on average 10-15 minutes earlier, improving alarm timeliness by about 40%-60%.

[0158] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0159] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A multi-parameter collaborative sensing and intelligent alarm system for prominent controlling geological anomalies, characterized in that, It includes a data acquisition and preprocessing module, a feature extraction module, a feature fusion module, and an alarm generation module connected in sequence, wherein: The data acquisition and preprocessing module is used to synchronously acquire multi-dimensional parametric data, including microseismic waveforms, gas concentration, stress / strain data, acoustic emission signals, temperature and pressure data, and geological structure vector images, through a downhole sensor array, and perform noise reduction based on a layout condition diffusion model to output a preprocessed multi-parameter data matrix. The denoising process employs a layout-conditional diffusion model, specifically including: The multi-parameter layout embedding unit is used to define each parameter data as a multimodal layout element. It uniformly represents the layout position of time series data and image data through normalized bounding box coordinate vectors, and generates layout control tokens by combining coordinate encoding and category encoding. The layout mask attention unit is used to construct a 0 / 1 layout mask matrix, where 1 represents the valid data region and 0 represents the noise / invalid region. In the reverse denoising process of the diffusion model, the attention weight is forced to focus on the valid data region through the mask mechanism, thereby suppressing the feature update of the noise region. The dual cross-attention unit is used to fuse global textual semantic conditions that represent prior knowledge of geological anomalies with local layout conditions that represent local layout relationships of parameters. The layout control token is used as input to the local layout conditions to construct a local layout condition vector and outputs denoising features through weighted cross-attention. The operations performed by the layout mask attention unit include: In each denoising step of the diffusion model, the layout mask matrix M is introduced into the attention calculation: attention weights in noisy regions are suppressed through mask multiplication, and noise suppression is further enhanced through a learnable mask bias term matrix, so that the attention weights are concentrated on the effective data regions identified by the mask. The specific calculation formula is as follows: in, : Layout mask output features of the attention layer; : Feature query vector for the current denoising step; : The layout mask matrix for the i-th type of parameter, whose elements are 0 or 1; Activation function; : No. Key vectors of class parameters; : No. Class parameter key vector The transpose of the matrix; Key vector The dimension; : No. Learnable mask bias term matrix of class parameters; : No. The value vector of the class parameter contains the feature information to be extracted; This indicates element-wise multiplication; the feature extraction module is used to extract features from the time-series data using a time-series feature extraction network based on the preprocessed multi-parameter data matrix, to obtain a time-series feature vector; and simultaneously uses a multi-branch hybrid pooling convolutional neural network to extract features from the image data, to obtain an image feature vector; The feature fusion module is used to receive temporal feature vectors and image feature vectors from the feature extraction module, construct a cross-parameter channel correlation matrix through a cross-modal feature channel attention mechanism, align and fuse heterogeneous features, and output a multi-parameter collaborative representation vector. The alarm generation module is used to input the multi-parameter collaborative representation vector into the integrated learning anomaly discrimination model, output anomaly probability values, and generate multi-level alarm signals by comparing the anomaly probability values ​​with preset thresholds.

2. The system according to claim 1, characterized in that, The operations performed by the multi-parameter layout embedding unit include: Each parameter data is defined as a multimodal layout element, where time-series data is represented by a time-amplitude two-dimensional matrix, and image data is represented by a space-pixel value two-dimensional matrix. For the effective data region of each type of parameter, an 8-axis bounding box is used. Define its layout position to ensure that the layout information of different parameters is represented in a consistent manner; The bounding box coordinates are converted into frequency domain vectors using Fourier coding. Meanwhile, the parameter category code c is obtained through the frozen CLIP text encoder; Layout control tokens are generated by fusing coordinate encoding and category encoding through a linear layer. Where Linear(·) is a linear transformation layer, and the layout control token Used to implement structured embedding of multi-parameter layout information.

3. The system according to claim 2, characterized in that, The dual-cross attention unit performs the following operations: Text representing geological anomalies is input into a pre-trained language encoder to generate global textual conditional features, which in turn yields global key vectors. and global value vector ; The layout control token is processed through a linear projection layer to generate local layout condition features, thereby obtaining local key vectors. and local value vector ; The query vector Q and the global key-value pair are calculated using the cross-attention function Ψ. , The first attention result, and Q with local key-value pairs , The second attention result is then used to weight and sum the first and second attention results using a weight coefficient λ, resulting in the final denoised feature output, Out. .

4. The system according to claim 1, characterized in that, The multi-branch hybrid pooling convolutional neural network includes three feature extraction branches configured in parallel: The first branch employs a max pooling strategy and positive interval linear units to extract strong response features from the image. ; The second branch employs a minimum pooling strategy and negative interval linear units to extract weak response features from the image. ; The third branch employs an interleaved feature fusion strategy, simultaneously capturing strong and weak response features through dual-interval linear units. The output of the dual-interval linear unit is formed by interleaving and splicing the outputs of the positive and negative interval linear units according to the feature dimensions, thus obtaining the strong and weak response features. ; The outputs of the three branches are fused using a weighted average or arithmetic average to generate the final image features. : .

5. The system according to claim 1, characterized in that, The cross-modal feature channel attention mechanism includes: Feature mapping and compression: The features of each parameter are compressed into channel feature vectors through adaptive average pooling; Channel attention interaction: Perform activation-compression and compression-activation operations on the compressed channel feature vectors to generate channel attention vectors; Cross-feature association modeling: Construct a cross-parameter channel correlation matrix, and achieve feature mapping and fusion through selective Softmax normalization and 1-modulus tensor product to generate a multi-parameter collaborative representation vector.

6. The system according to claim 5, characterized in that, The channel attention interaction operation includes: The channel feature vector of the first parameter is first subjected to an excitation transformation and then a compression transformation. At the same time, the channel feature vector of the second parameter is first subjected to the compression transformation and then the excitation transformation to generate their respective channel attention vectors.

7. The system according to claim 1, 5, or 6, characterized in that, The alarm generation module includes: The anomaly detection unit calculates the multi-parameter collaborative representation vector based on the gradient boosting tree model and outputs anomaly probability values. ; The alarm classification unit has at least three preset probability thresholds. The abnormal probability value The alarm signals are compared with the three probability thresholds to trigger different levels of alarm signals, specifically including: when The system will trigger an alarm to indicate potential early signs of anomalies in the monitored area. when When a warning is triggered, it indicates that the abnormal characteristics are obvious and that enhanced monitoring is needed; when An emergency alarm is triggered, indicating that an impending geological anomaly requires immediate preventative measures.

8. The system according to claim 1 or 5, characterized in that, The alarm generation module also includes: The contribution analysis unit is used to calculate the contribution of each parameter to the alarm result based on the weights in the cross-modal feature channel attention mechanism. The prevention and control strategy matching unit is used to associate alarm results with geological structures, match them with a preset prevention and control strategy library, and output targeted prevention and control action suggestions.

Images