Roadway surrounding rock danger identification model construction method

Abstract

The invention relates to the technical field of roadway surrounding rock danger identification, and discloses a roadway surrounding rock danger identification model construction method, which comprises the steps of collecting multi-modal data, and generating preprocessed data through synchronous calibration and denoising; extracting a seismic wave frequency domain and image texture features, and generating a multi-modal feature matrix; in combination with a geological prior clustering mining abnormal mode, generating a labeled sample data set; generating a danger identification model based on a transfer learning and feature fusion training network; and the edge deployment model performs real-time reasoning, and generates an early warning result through an adaptive algorithm. According to the method, the frequency domain features of the seismic fluctuation signals and the depth texture features of the surrounding rock images are fused, the multi-modal feature matrix is constructed, abnormal mode mining is carried out in combination with geological prior knowledge, and early weak abnormal signals such as hidden fault slippage or asymmetric microfracture extension which are difficult to find by a single monitoring means can be effectively recognized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of roadway surrounding rock danger identification, in particular to a roadway surrounding rock danger identification model construction method. BACKGROUND

[0002] The depth of coal resource exploitation extends downward year by year, and the deep roadway surrounding rock bears higher ground stress, more complex tectonic stress and strong mining disturbance, which leads to the stability of the deep roadway surrounding rock changing easily. In the deep well high stress environment, the precursors of surrounding rock instability or disaster are often hidden and weak abnormal signals.

[0003] In the existing roadway surrounding rock monitoring field, independent data acquisition and analysis are mainly relied on by multiple physical sensors. Seismic wave monitoring or microseismic monitoring systems are widely used to locate potential fracture sources or high stress concentration areas by recording elastic waves generated by stress release in rock mass. This method focuses on the mechanical behavior of the surrounding rock, and engineers use high-definition cameras or industrial video systems for manual observation or simple image processing to capture macroscopic cracks, spalling or rib spalling on the surface of the surrounding rock as an auxiliary visual judgment. In terms of data processing, the existing technology mainly includes statistical analysis of single sensor data, such as calculating the energy and frequency of seismic waves, or using basic unsupervised clustering algorithms to attempt to preliminarily classify monitoring data in order to find possible abnormal aggregation points.

[0004] However, the existing detection technology focuses on the internal mechanical state of the seismic wave data, and the image texture data reflects the surface deformation information. The independent processing of heterogeneous data in time and space fails to effectively establish the coupling relationship between the stress evolution of the surrounding rock and the external surface characteristics. This single information source analysis mode is difficult to fully represent the dangerous state of the surrounding rock, resulting in a lack of sensitivity to some hidden and complex early instability precursors. Geological disasters are small probability events, resulting in extremely scarce abnormal sample data available for supervised training, and a serious class imbalance problem. Although some technologies use unsupervised clustering, they lack the guidance and verification of geological prior knowledge and are difficult to distinguish between real geological anomalies and environmental noise or working condition interference from a physical mechanism, resulting in insufficient physical interpretability and accuracy of the model. Therefore, the present application provides a roadway surrounding rock danger identification model construction method to solve the problems in the prior art. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a roadway surrounding rock danger identification model construction method, which solves the problems of insufficient deep fusion of multi-modal data, lack of model training labeled samples and poor robustness and adaptability of early warning criteria in the prior art.

[0006] To achieve the above object, the present application is realized by the following technical solutions: The first aspect of the present application provides a roadway surrounding rock danger identification model construction method, comprising the following steps: S1, collecting original data including acoustic emission signals, seismic wave signals and image data through a plurality of types of sensor networks arranged in the roadway surrounding rock area, performing synchronous time calibration and denoising processing to generate multi-modal preprocessed data with time alignment and denoising; S2, using wavelet transform to extract frequency domain features of the seismic wave signals, and combining deep convolutional neural network to extract local and global texture features of the surrounding rock images, to generate a multi-modal feature matrix; S3, using an improved clustering algorithm combined with prior geological knowledge to mine abnormal patterns from the multi-modal feature matrix, identifying abnormal signal patterns, and marking feature data corresponding to the abnormal signal patterns as dangerous samples to generate a labeled sample dataset; S4, using a deep learning algorithm based on transfer learning, integrating heterogeneous features in the multi-modal feature matrix through a feature fusion strategy, training a deep learning network using the labeled sample dataset, and generating a roadway surrounding rock danger identification model; S5, setting the roadway surrounding rock danger identification model in an edge computing device, performing inference calculation on real-time collected sensor data to obtain a danger score, adjusting an alarm strategy according to the danger score through an adaptive warning algorithm, and generating a surrounding rock stability determination result.

[0007] Preferably, in step S1, the step of performing synchronous time calibration and denoising processing further comprises: Providing a unified time reference for all sensors through a GPS timestamp or a high-precision clock synchronization mechanism, recording the time offset value of each type of sensor, and reordering the data stream according to the timestamp; Using specific denoising algorithms for different data types: for acoustic emission signals and seismic wave signals, using frequency domain filtering and wavelet transform decomposition to remove high-frequency interference and low-frequency baseline drift; for image data, improving image clarity through a non-local mean algorithm or a convolutional neural network denoising model; After completing the denoising processing, performing integrity check and consistency check on the data, if abnormal or mismatched data is found, supplementing through data interpolation or reacquisition mechanism, and outputting multi-modal preprocessed data.

[0008] Preferably, the step S2 specifically comprises: Segmenting the seismic wave signals in the multi-modal preprocessed data and selecting a time window, decomposing the signals into multi-scale time-frequency components using wavelet transform, and extracting frequency domain features including energy density spectrum, high-frequency signal abrupt change point and low-frequency trend; inputting the surrounding rock image in the multi-modal pre-processed data into a pre-trained deep convolutional neural network, extracting local texture features including micro-fissure edge characteristics and surrounding rock surface roughness changes through multi-layer convolution operations, and analyzing the distribution pattern of the overall surface texture of the surrounding rock through a fully connected layer to extract global texture features; combining the extracted frequency domain features with the local and global texture features to construct the multi-modal feature matrix.

[0009] Preferably, in step S3, the step of identifying the abnormal signal pattern further comprises: According to the data distribution characteristics of the multi-modal feature matrix, a density-based clustering or Gaussian mixture model is selected as an improved clustering algorithm, and prior geological knowledge is introduced to optimize the initial parameter settings of the clustering algorithm: for the Gaussian mixture model, the initial mean vector and covariance matrix are set according to prior knowledge; for the density-based clustering, the radius parameter and the minimum number of sample points are adjusted; Performing the clustering algorithm, assigning the feature points in the multi-modal feature matrix to corresponding clustering clusters, and identifying the abnormal signal pattern pointing to the hidden fault or micro-fissure expansion area by judging the local density of the data points and the number of neighbors within a certain range; Comparing and verifying the clustering results with prior geological knowledge, taking the feature data corresponding to the verified matching clustering results as positive samples, and taking the rest as negative samples to generate a labeled sample dataset.

[0010] Preferably, in step S4, the step of integrating heterogeneous features in the multi-modal feature matrix through a feature fusion strategy further comprises: Selecting a deep residual network to process image data and a deep recurrent neural network to process seismic wave signal features, obtaining pre-trained model parameters of the deep residual network, and generating a set of randomly initialized parameters of the deep residual network; Based on the pre-trained model parameters and the set of randomly initialized parameters, performing weighted summation according to a weight factor to obtain an initialization parameter set of the model; Performing feature splicing operations on the seismic wave signal frequency domain features and surrounding rock image data in the multi-modal feature matrix to generate a fusion feature matrix containing multi-modal features.

[0011] Preferably, in step S4, the step of generating a roadway surrounding rock danger identification model further comprises: Dividing the labeled sample dataset into a training set and a validation set; Constructing a loss function for model training, and calculating the value of the loss function based on the training set, wherein the value of the loss function is calculated based on the weight of the sample category, the true label, and the predicted value of the model output, and is used to represent the gap between the current prediction and the true label; During the training process, a hybrid data augmentation technique and a dynamic dropout mechanism are introduced to randomly zero the dimensions of the fusion feature matrix, and the validation set is used to evaluate the model to obtain the precision, recall, and F1 value to dynamically adjust the learning rate.

[0012] Preferably, in step S4, the step of training the deep learning network further includes introducing an adversarial training technique to enhance the robustness of the model. An adversarial sample is generated by selecting a sample from the labeled sample dataset as an input sample, and the adversarial sample is obtained by superimposing a perturbation term on the input sample. The direction of the perturbation term is determined by the sign of the gradient of the loss function with respect to the input sample during the training process, and the size of the perturbation term is determined by a pre-set perturbation control coefficient.

[0013] Preferably, in step S5, the step of generating the surrounding rock stability determination result further includes: Using a Sigmoid function to map the linear combination value output by the surrounding rock danger identification model into a probability form of danger score, and when the danger score exceeds a threshold value, determining that the surrounding rock stability result is unstable; The danger score threshold is dynamically adjusted by an adaptive warning algorithm, and the adjustment strategy is calculated based on the current danger score threshold, combined with the change rate of the false alarm rate over time, the change rate of the false alarm rate over time, and the acceleration of the danger score trend, to balance the false alarm rate and the false alarm rate.

[0014] The second aspect of the present application provides a surrounding rock danger identification model construction device, comprising: A data acquisition and processing module is used to acquire raw data including acoustic emission signals, seismic wave signals, and image data through a multi-type sensor network arranged in the surrounding rock area of the roadway, perform synchronous time calibration and denoising processing, and generate multi-modal preprocessed data that is time-aligned and denoised. A feature extraction module is used to extract frequency domain features of seismic wave signals using wavelet transform, and combine deep convolutional neural networks to extract local and global texture features of surrounding rock images, and generate a multi-modal feature matrix. An anomaly mining module is used to mine abnormal patterns from the multi-modal feature matrix through an improved clustering algorithm combined with prior geological knowledge, identify abnormal signal patterns, and mark feature data corresponding to the abnormal signal patterns as dangerous samples to generate a labeled sample dataset. A model training module is used to adopt a deep learning algorithm based on transfer learning, integrate heterogeneous features in the multi-modal feature matrix through a feature fusion strategy, train a deep learning network using the labeled sample dataset, and generate a surrounding rock danger identification model. The early warning application module is used for setting the roadway surrounding rock danger identification model to an edge computing device, performing inference calculation on real-time collected sensing data to obtain a danger score, adjusting an alarm strategy according to the danger score through an adaptive early warning algorithm, and generating a surrounding rock stability judgment result.

[0015] The third aspect of the present application provides an electronic device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the roadway surrounding rock danger identification model construction method according to any one of the preceding aspects when executing the computer program.

[0016] The present application provides a roadway surrounding rock danger identification model construction method. 1、The present application fuses the frequency domain features of the seismic wave signal and the depth texture features of the surrounding rock image, constructs a multi-modal feature matrix, and can capture the correlation information of the internal structure changes and surface deformation trends of the surrounding rock from two dimensions of time and space. After feature depth extraction, abnormal pattern mining is performed in combination with geological prior knowledge, so that the model can effectively identify early and weak abnormal signals such as concealed fault slip or asymmetric microcrack expansion that are difficult to discover by a single monitoring means, and the sensitivity and accuracy of early warning are improved.

[0017] 2、The present application adopts an abnormal pattern mining step based on geological prior knowledge, uses an improved clustering algorithm in unsupervised learning to integrate the prior knowledge of geologists into clustering parameter setting, successfully converts the original monitoring data which is difficult to obtain labels into a labeled sample data set which can be used for supervised training, introduces a transfer learning and an adversarial training mechanism in the model construction link, the transfer learning solves the problem of insufficient underground data, and the adversarial training forces the model to learn a more robust decision boundary by adding a small disturbance in the sample, thereby enhancing the anti-interference ability and generalization performance of the model in a complex and high-noise roadway environment.

[0018] 3、The present application deploys the trained model to an edge computing device, realizes real-time collection of data, feature extraction and danger score calculation, ensures the timeliness of early warning, and the early warning application module adopts an adaptive early warning algorithm, comprehensively considers the change trend of the false alarm rate and the missed alarm rate in history statistics and the second derivative of the current danger score, so that the system can automatically adapt to the change of the surrounding rock environment, capture sudden geological risks in advance, and avoid the lag and mistakes commonly seen in traditional fixed threshold early warning. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The present application provides a method step flowchart; Figure 2 The present application provides a multi-modal feature extraction and fusion flowchart; Figure 3 A schematic diagram of an electronic device according to the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the specification of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0021] Referring to the accompanying Figure 1 , Figure 1 is a flow chart of method steps according to an embodiment of the present application. The present application provides a roadway surrounding rock danger identification model construction method, which is applied to a roadway surrounding rock danger identification model construction device. The device comprises a data acquisition and processing module, a feature extraction module, an anomaly mining module, a model training module and a warning application module.

[0022] The multi-type sensor network comprises an acoustic emission sensor, a seismic wave recorder and a high-definition camera, which are respectively used to collect acoustic emission signals, seismic wave signals and image data of the surrounding rock. These sensors are connected to the data acquisition and processing module through a wired or wireless network.

[0023] The data acquisition and processing module performs data receiving and preprocessing work. The data acquisition and processing module provides a unified time reference for heterogeneous sensor data through a GPS timestamp or a high-precision clock synchronization mechanism, and records a time offset value to realize data stream reordering. For different types of data, frequency domain filtering, wavelet transform decomposition or image denoising algorithms are called to remove environmental noise and interference, and multi-modal preprocessed data after time alignment and denoising are output.

[0024] The feature extraction module is used to analyze the feature dimension of the preprocessed data. The wavelet transform algorithm is used to process the seismic wave signals to extract frequency domain features. At the same time, the deep convolutional neural network is used to process the surrounding rock images to extract local texture features and global texture features. The feature extraction module combines the extracted heterogeneous features to generate a multi-modal feature matrix.

[0025] The anomaly mining module is used to solve the problem of missing label data in an industrial scene. The anomaly mining module runs an improved clustering algorithm and combines pre-set prior geological knowledge to perform unsupervised or semi-supervised mining on the multi-modal feature matrix. The anomaly mining module identifies abnormal signal patterns pointing to hidden faults or micro-crack expansion areas, and marks the feature data corresponding to these patterns as dangerous samples, thereby constructing a labeled sample data set with labels.

[0026] The model training module is used for generating the identification model, the model training module adopts a deep learning algorithm based on transfer learning, integrates heterogeneous features in a multi-modal feature matrix through a feature fusion strategy, trains a deep learning network by using a labeled sample data set, and introduces an adaptive learning rate optimization algorithm and an adversarial training technique in the training process, and finally generates a roadway surrounding rock danger identification model.

[0027] The early warning application module is arranged in the edge computing device, the early warning application module loads the trained roadway surrounding rock danger identification model, performs inference calculation on the real-time collected sensing data, and outputs a danger score, the early warning application module is internally provided with an adaptive early warning algorithm, dynamically adjusts an alarm threshold according to a historical false alarm rate, a missed alarm rate and a score trend, and generates a final surrounding rock stability determination result.

[0028] The present application provides a kind of roadway surrounding rock danger identification model construction method, comprising the following steps: S1, the original data including acoustic emission signal, seismic fluctuation signal and image data are collected by the multi-type sensor network arranged in the surrounding rock area of roadway, and the multi-modal preprocessed data after time alignment and denoising are generated by time calibration and denoising processing;In this step, the data acquisition and processing module cleans the original data, uses GPS time stamp or high-precision clock synchronization mechanism to unify time reference, and applies frequency domain filtering or denoising algorithm to different modal data, to ensure the integrity and consistency of input data; S2, the frequency domain features of seismic fluctuation signal are extracted using wavelet transform, and the local and global texture features of surrounding rock image are extracted combining deep convolutional neural network, to generate multi-modal feature matrix;Feature extraction module analyzes one-dimensional time series signal and two-dimensional image signal respectively, excavates frequency energy feature and spatial texture feature, and constructs multi-modal feature matrix that can comprehensively represent surrounding rock state; S3, abnormal pattern mining is performed on multi-modal feature matrix by improved clustering algorithm combined with prior geological knowledge, abnormal signal pattern is identified, and feature data corresponding to abnormal signal pattern is marked as dangerous sample to generate labeled sample data set;Abnormal mining module identifies abnormal clusters pointing to hidden fault or microcrack through unsupervised or semi-supervised clustering, and marks them as dangerous samples in reverse, to realize the conversion from unsupervised data to supervised training sample; S4, a deep learning algorithm based on transfer learning is adopted, heterogeneous features in multi-modal feature matrix are integrated through feature fusion strategy, deep learning network is trained by using labeled sample data set, and roadway surrounding rock danger identification model is generated;The model training module integrates heterogeneous features using dual-flow network architecture, trains by using labeled data set, and introduces adversarial training to strengthen model boundary, to generate a model with high-precision identification capability; S5, setting the roadway surrounding rock danger identification model in the edge computing device, performing inference calculation on the real-time collected sensor data to obtain a danger score, adjusting an alarm strategy according to the danger score through an adaptive early warning algorithm, and generating a surrounding rock stability determination result; the early warning application module dynamically adjusts an early warning threshold according to a real-time false alarm rate, a missed alarm rate and a score trend, balances false alarms and missed alarms, and outputs a reliable stability determination result.

[0029] The specific implementation process of each step in the above method flow will be described in detail below.

[0030] Referring to the accompanying Figure 2 In step S1, in order to ensure that the data basis for subsequent model construction has high integrity and consistency, the data collection and multi-modal preprocessing process mainly includes four logical links of sensor network deployment, time synchronization calibration, specific denoising processing and data verification.

[0031] In view of the comprehensive collection requirement of roadway surrounding rock data, the sensor layout points in the roadway are planned. According to the geometric structure of the roadway, the rock layer distribution and the danger area prediction result, the installation positions of the acoustic emission sensor, the seismic wave recorder and the high-definition camera are determined respectively. Among them, the acoustic emission sensor is configured to capture the dynamic change of the surrounding rock crack, the seismic wave recorder is configured to monitor the underground microseismic vibration, and the high-definition camera is configured to record the surface change and texture characteristics of the surrounding rock. The above-mentioned sensors are connected to the data collection terminal through wireless or wired network. According to the noise characteristics in the roadway environment, the corresponding collection frequency and sensitivity parameters are configured. In specific implementation, the acoustic emission sensor is set to a high sampling rate mode to capture the energy release in a short time, and the high-definition camera is adjusted to a mode suitable for low light conditions. The specific connection mode and signal transmission protocol of the sensor hardware belong to the known technology of those skilled in the art, and will not be described here.

[0032] Considering the differences in sampling frequency and start-up time of multiple types of sensors, the system performs synchronous time calibration on the collected original data, and provides a unified time reference for all connected sensors through a GPS timestamp or a high-precision clock synchronization mechanism. The data processing unit records the time offset value of each type of sensor relative to the reference, and reorders the data stream according to the recorded time offset value and the timestamp of the original data. Through the above operation, it is ensured that the data of all modalities are mapped to the same time axis. This strict time alignment processing can reveal the time sequence correlation between different physical quantities, for example, accurately matching the microseismic event in the seismic wave signal with the energy burst in the acoustic emission signal in the time dimension, providing data support for revealing the internal relationship between crack propagation and microseismic events.

[0033] According to the characteristics of complex roadway environment and a large amount of noise, the system adopts specific denoising algorithms according to different data types. For one-dimensional time series data such as acoustic emission signals and seismic wave signals, frequency domain filtering and wavelet transform decomposition techniques are used. Through frequency domain filtering, background noise in a specific frequency band is filtered out, and using the multi-scale characteristics of wavelet transform, high-frequency interference signals and low-frequency baseline drift are separated and removed, so as to retain the effective geological signal characteristics. For two-dimensional data such as surrounding rock images, non-local mean (NL-Means) algorithm or denoising model based on convolutional neural network (CNN) is used. The non-local mean algorithm uses the redundant information within the image to remove noise points, or uses a pre-trained CNN denoising model to infer the image to improve image clarity and enhance texture details. After denoising, the signal-to-noise ratio of the signal is improved, which can meet the requirements of data accuracy for geological anomaly detection.

[0034] After completing the denoising process, the system performs data integrity check and consistency check. The integrity check detects whether there is data interruption caused by data packet loss or sensor failure by statistical analysis of data records in consecutive time periods, ensuring signal continuity within each set time window. Consistency check focuses on verifying the logical matching between multi-modal data, detecting whether there is a situation where one type of modal data is missing while other modal data exists. If data anomalies, missing or mismatching are found during the inspection process, the system starts the data interpolation algorithm to digitally fit and supplement the missing points, or triggers the re-acquisition mechanism to obtain the data of that time period.

[0035] In step S2, in order to construct a multi-modal feature matrix that can comprehensively represent the state of surrounding rock, the feature extraction process mainly involves frequency domain analysis of one-dimensional time series signals and deep texture mining of two-dimensional image data, and the specific execution process is as follows: For seismic wave signals in multi-modal preprocessed data, the feature extraction module performs segmentation processing and selects a specific time window. In this time window, wavelet transform algorithm is used for time-frequency analysis of the signal. Specifically, Gaussian wavelet or Morlet wavelet is selected as the basis function to decompose the seismic wave signal into multi-scale time-frequency components. This process aims to capture the dynamic characteristics of frequency change over time, and uses the localization ability of wavelet transform in time-frequency domain to detect the high-frequency characteristics of microseismic events in a short time window, and analyze the distribution of low-frequency energy in a long time window.

[0036] Based on the above decomposition results, frequency domain features including energy density spectrum, high frequency signal mutation point and low frequency trend are extracted. Among them, the energy density spectrum reflects the distribution of signal energy in different frequency bands; the high frequency signal mutation point indicates the specific time and intensity of the microseismic event; the low frequency trend represents the background benchmark of the energy evolution of the surrounding rock. By extracting the above features, the dynamic change pattern of crack propagation or hidden fault in the surrounding rock of the roadway can be identified. The specific mathematical implementation of wavelet transform and the selection of basis functions belong to the common technical knowledge of those skilled in the art, and will not be described here.

[0037] For the surrounding rock image in the multi-modal pre-processed data, it is input into a pre-trained deep convolutional neural network (CNN) model. The network structure includes multiple convolutional layers, pooling layers and fully connected layers, which are used to analyze the image information layer by layer.

[0038] In the local feature extraction stage, the network processes the input image through multiple convolutional operations. Each layer of convolutional operation cooperates with the activation function (such as ReLU function) to perform nonlinear mapping and amplification of texture features, and reduces the feature dimension through the pooling layer to improve the calculation efficiency. The shallow part of the convolutional layer mainly extracts local texture features such as the edge characteristics of micro-cracks on the surface of the surrounding rock and the roughness change of the surrounding rock surface; the deep part extracts features with semantic information, such as the distribution pattern of crack network or the potential fault trend.

[0039] In the global feature extraction stage, the spatial information of the image is summarized by using the subsequent fully connected layers of the deep convolutional neural network, and the global texture features are extracted by analyzing the distribution pattern of the overall surface texture of the surrounding rock, such as the consistency of the surface crack trend and the regional color change. The network gives dynamic weights to different feature channels during the extraction process to highlight the texture signals that are strongly related to the stability of the surrounding rock and suppress the background interference. These global features can provide data support for the estimation of crack penetration length or the evaluation of the weathering degree of the surrounding rock.

[0040] After the frequency domain features of the seismic wave signal and the local and global texture features of the surrounding rock image are extracted, the system performs feature combination operation to align and splice the extracted wavelet features of the seismic wave signal and the image depth features according to the time dimension. The feature fusion process is realized by feature-level splicing, which ensures the information complementation between different data sources. The spliced features are processed by dimension reduction to remove redundant information and adapt to the input dimension requirements of the subsequent model. After the above processing, a feature matrix containing multi-modal information is finally generated, which is stored in the database and bound with the original data index, serving as the input basis for subsequent anomaly pattern mining and model training.

[0041] In step S3, in order to solve the problem of lack of geological anomaly label data in industrial scenarios, an unsupervised or semi-supervised learning strategy is used, combined with prior geological knowledge to deeply mine the multi-modal feature matrix, to identify potential dangerous signals and generate labeled sample data sets for model training. The process is implemented as follows: For the extracted seismic wave signal frequency domain features and surrounding rock image texture features, considering that the feature dimensions and physical quantities of different sensors are significantly different (for example, the energy spectrum value range of the seismic wave is usually much larger than the numerical range of the image texture features), the anomaly mining module first performs data normalization. Z-score standardization or Min-Max normalization algorithm is used to map all feature values in the feature matrix to a unified numerical interval to eliminate dimensional differences. The integrity of the feature matrix is scanned, and if missing values or obvious outlier values are found, linear interpolation is used for filling or direct elimination to ensure that the data input into the clustering algorithm has mathematical completeness.

[0042] According to the distribution characteristics of geological data, an improved clustering algorithm is selected as the core tool for anomaly pattern mining. In specific implementation, a density-based clustering algorithm (such as DBSCAN) or a Gaussian mixture model (GMM) based on probability distribution can be selected. In order to overcome the blindness of traditional clustering algorithms in the parameter initialization stage, prior geological knowledge is introduced to optimize the algorithm parameters.

[0043] If a Gaussian mixture model is selected, the surrounding rock lithology characteristics and historical geological exploration data are used to infer the possible fault distribution area in the initialization stage, and the initial mean vector and covariance matrix of the model are set accordingly to guide the clustering process to converge to the actual geological structure. If a density-based clustering algorithm (DBSCAN) is selected, the algorithm parameters are adjusted according to the scale characteristics of the target detection object; for example, in order to improve the sensitivity to small crack propagation signals, the neighborhood radius parameter and the minimum number of sample points of the core object are adjusted according to prior knowledge.

[0044] The clustering algorithm is executed, and Euclidean distance or cosine similarity is used as the similarity measure in the feature space to assign data points in the multi-modal feature matrix to different clusters. The algorithm divides data points into core points, boundary points and noise points according to the local density of the data points and the number of neighbors within a certain range. Core points in high-density areas usually represent normal signal distribution of stable surrounding rock, while noise points and isolated small clusters that cannot be classified or are located in low-density areas indicate potential geological anomalies.

[0045] If the characteristics of some cluster are found to be concentrated in the frequency domain, showing low-frequency energy anomalies, or in the texture space, showing specific abnormal patterns, it is determined that the area may have hidden fault slip; if there are small clusters of asymmetric distribution, it is determined that they correspond to the asymmetric expansion area of micro-cracks. Through the above analysis, the mathematical clustering anomaly is mapped to the specific potential risk point of the surrounding rock of the roadway.

[0046] After completing the preliminary clustering identification, the identified abnormal pattern results are compared and verified with the prior knowledge base of geologists. For the cluster determined to be a hidden fault, compare the core sampling records or ultrasonic detection data to confirm the authenticity of its location and shape; for the area determined to be a micro-crack distribution, compare the high-energy release records of acoustic emission signals.

[0047] After verification and matching, the feature data corresponding to the confirmed abnormal signal pattern is marked as a positive sample (dangerous sample), and the feature data corresponding to the high-density cluster where the core point is located is marked as a negative sample (safe sample). Thus, a labeled sample dataset containing feature data and its corresponding class labels is constructed. This labeled sample dataset realizes the transformation from unsupervised raw data to supervised training samples, providing a reliable data foundation for the training of subsequent deep learning models. The specific mathematical principles of the DBSCAN algorithm and the GMM model belong to the common technical knowledge of those skilled in the art, and will not be repeated here.

[0048] In step S4, in order to generate a high-precision roadway surrounding rock danger identification model, the system uses a deep learning algorithm based on transfer learning to train the model using the labeled sample dataset generated in step S3. The specific execution process is as follows: The labeled sample dataset generated by the abnormal mining module includes a seismic wave frequency domain feature matrix and a surrounding rock image feature matrix. Define the seismic wave frequency domain feature matrix as , where represents the number of samples, represents the seismic wave feature dimension; define the surrounding rock image feature matrix as , where represents the image feature dimension. Each sample corresponds to a label , which is used to indicate whether the surrounding rock has hidden faults or asymmetric micro-cracks and other dangerous anomalies.

[0049] To enhance the generalization ability of the model to complex downhole environments, data augmentation is performed on the sample data. For image features, random cropping, rotation transformation, and brightness adjustment operations are applied to simulate different shooting angles and lighting conditions; for seismic wave features, Gaussian noise is added or random masking operations are performed to simulate interference during signal transmission. The processed data is divided into training and test sets in the ratio of 80% and 20%. The training set is used to iteratively optimize the model parameters to minimize the loss function, and the test set is used to independently verify the model performance. By this division, it is detected whether the model has overfitting phenomenon that performs well in the training set but performs poorly in the test set, so as to evaluate the robustness of the model in actual application scenarios.

[0050] The model adopts a dual-stream network architecture to process heterogeneous features, and selects a pre-trained ResNet model on a large dataset (such as ImageNet) as the backbone network to process image features, while selecting a deep recurrent neural network (RNN) to process seismic wave time series features. In the parameter initialization stage, the pre-trained weights and random initialization strategy are combined, and the expression is as follows: ; In the formula, is the initialization parameter set of the model, and is the final model parameter used for the current task, is the parameter set of the pre-trained model, which is the weight obtained by training on a large dataset (such as ImageNet), and contains strong generalization ability for general features, is the parameter set of random initialization, which is usually generated by initialization or initialization method, is the weight factor, whose value range is , which controls the proportion of pre-trained parameters and randomly initialized parameters.

[0051] A fusion module is set before the output layer of the model to jointly process seismic wave and image features extracted through different network branches. The feature splicing strategy is adopted, and the expression is as follows: ; In the formula, represents the fused feature matrix, which contains multi-modal features from seismic wave frequency domain and surrounding rock image, is the space of the feature matrix, which indicates that the fused data has rows (sample number) and columns (total feature dimension after splicing).

[0052] The model is trained using an adaptive learning rate optimization algorithm (such as AdamW) to maximize the identification accuracy of abnormal samples. The loss function is in the form of weighted binary cross-entropy, and the calculation expression is as follows: ; In the formula, is the value of the loss function, indicating the difference between the current prediction of the model and the true label, is the weight of the sample category, used to balance the influence of positive and negative samples, is the true label of the th sample, is the prediction value of the th sample, which is the probability output by the model, ranging from .

[0053] During the training iteration process, the mixed data augmentation technique and dynamic dropout mechanism are introduced to randomly set zero to part of the dimensions of the feature matrix , so as to reduce the co-adaptation between features and prevent overfitting. After completing one round of training (Epoch), the system calculates the performance indicators on the validation set and dynamically adjusts the learning rate according to the following formula: ; In the formula, is the learning rate of the th iteration, is the initial learning rate, the learning rate value at the beginning of training, is the learning rate decay factor, is the current iteration number.

[0054] After training, the evaluation indicators include precision, recall and F1 value, and the calculation expressions are as follows: ; In the formula, is the precision, is the true positive, that is, the number of samples correctly predicted as positive by the model (the number of times the model correctly identifies dangerous anomalies), is the false positive, that is, the number of samples incorrectly predicted as positive by the model (the number of times the model incorrectly judges stable surrounding rock as abnormal); ; In the formula, is the recall, is the number of false negatives, indicating the number of samples that actually exist dangerous anomalies but are incorrectly judged as normal by the model.

[0055] ; In the formula, is a balance index.

[0056] To solve the problem of rare geological anomaly samples in roadway surrounding rock, the adversarial training technique is introduced. By superimposing a small perturbation on the original input sample, an adversarial sample is generated, forcing the model to learn a more robust decision boundary. The generation formula of the adversarial sample is as follows: ; In the formula, represents the generated adversarial sample; represents the original input sample; represents the control coefficient of the perturbation size, which is used to limit the perturbation amplitude to maintain the semantic invariance of the sample; represents the loss function about the gradient of the input sample ; represents the sign function, which is used to extract the direction of the gradient. The generated adversarial sample is added to the training set for mixed training, and the final output roadway surrounding rock danger recognition model can output the danger probability map of rare anomalies and has the ability to resist environmental noise interference.

[0057] In step S5, in order to realize real-time perception and dynamic response to the risk of roadway surrounding rock, the trained model is deployed to the edge side, and an adaptive warning algorithm is introduced to dynamically calibrate the output results. The specific execution process is as follows: The multi-modal danger recognition model trained by transfer learning is deployed to the edge computing device on the roadway site. The edge computing device is equipped with a high-performance computing unit (such as a GPU accelerator or a TPU tensor processing unit) to meet the real-time requirements of matrix operations of deep learning models.

[0058] Sensing data flows into the edge device in real time through a unified communication interface, and the system module divides the continuous data stream into discrete batches for processing according to the preset time window. The real-time input data received by the model includes the seismic wave signal feature matrix and the image texture feature matrix . In order to eliminate the differences in numerical dimensions of different modal features, the model first performs feature normalization and weighted fusion operations, and the calculation expression is as follows: ; In the formula, is the fused multi-modal feature vector, which is used to comprehensively represent the seismic wave signal feature and the image texture feature , is a feature normalization function, which aims to map the features to a unified scale range to eliminate the dimensional differences between the features, and are feature weights, respectively reflecting the relative contribution of seismic signal feature and image texture feature to model determination.

[0059] Based on the fused features , the model outputs the hazard score at current time through the classifier layer. The hazard score is a quantitative indicator that converts the linear output of the model into a probabilistic form, and the calculation expression is as follows: ; In the formula, is the hazard score at the current time, ranging from [0, 1], and the higher the score, the more unstable the surrounding rock, is a function for mapping linear output to the range of [0, 1], and are weight matrices that control the influence of input features and nonlinear activation features on the output score, and the weight values are optimized through back propagation during model training, is a bias term that adjusts the baseline level of the score output, allowing the model to more flexibly adapt to different data distributions.

[0060] The function is a mathematical activation function. Its output range is (0, 1), which can smoothly map any input value (positive or negative real number) to a probability value between 0 and 1. The function has an "S" shaped curve characteristic. When the input value is small (tends to negative infinity), the output approaches 0; when the input value is large (tends to positive infinity), the output approaches 1; when the input value approaches 0, the function is most sensitive to changes, and the output is near 0.5.

[0061] Here the function is to convert the linear output calculated by the model (linear combination of weight matrices , and input features ) into a hazard score so as to represent it as a probability, which is convenient for further judging the risk of surrounding rock. This mapping method not only captures linear relationships, but also strengthens the representation of nonlinear features, while providing standardized output for subsequent threshold comparison. Through the function, the system can directly determine the risk level of the score (such as whether it exceeds the threshold), achieving sensitive and robust risk identification.

[0062] To solve the problem that fixed threshold is prone to false positive or false negative in complex geological environment, the early warning application module is built-in adaptive early warning algorithm. The algorithm dynamically calibrates the dangerous score threshold by comprehensively considering the change rate of historical statistical indicators and the evolution trend of current score, and the calculation expression is as follows: ; In the formula, is the adjusted dangerous score threshold, which is used to dynamically adapt to the current environmental characteristics, is the initially set fixed threshold, which is generally determined based on experience value or statistical historical data, is the false positive and false negative adjustment coefficient, which is used to balance the weight of false positive rate and false negative rate , is the change rate of false positive rate over time, if the false positive rate decreases rapidly, the model may need to appropriately improve the sensitivity, is the change rate of false negative rate over time, if the false negative rate increases, the threshold should be appropriately reduced to increase the early warning range, is the score trend adjustment coefficient, which reflects the sensitivity of score trend to threshold adjustment, is the score trend acceleration, the expression is as follows: , which is used to capture the acceleration of the score curve and reflect the potential of sudden change of risk.

[0063] The real-time calculated dangerous score is compared with the dynamic threshold . If , it is determined that the surrounding rock is in an unstable state, and the corresponding level of alarm (such as prompting monitoring, detailed inspection or emergency evacuation) is triggered according to the amplitude of the score exceeding, and the surrounding rock stability determination result is generated for engineering decision reference. The specific hardware selection and communication protocol implementation of the edge computing device belong to the common technical knowledge of those skilled in the art, and will not be described here.

[0064] Referring to the accompanying drawings, Figure 3 the electronic device includes at least one processor, memory, communication interface and communication bus. The communication bus is used to realize the connection communication between these components.

[0065] The processor is the control center of the electronic device, which connects various parts of the entire electronic device through various interfaces and lines. The processor can be a general central processing unit (CPU), a microprocessor, an application specific integrated circuit (ASIC), a graphics processor (GPU), or a tensor processing unit (TPU). In view of the high computing power requirement of deep convolutional neural networks and transfer learning algorithms for matrix operations, the processor preferably integrates a hardware acceleration unit to support real-time feature extraction and model inference calculation. The processor performs various functions and processes data of the electronic device by running or executing software programs or modules stored in the memory and calling data stored in the memory, thereby monitoring the dangerous state of the surrounding rock of the roadway.

[0066] The memory can be used to store software programs and modules, and the processor executes various function applications and data processing by running the software programs and modules stored in the memory. The memory mainly includes a program storage area and a data storage area. The program storage area can store an operating system and at least one application program required by a function (such as a data acquisition function, a feature extraction function, etc.); the data storage area can store data created according to the use of the electronic device (such as multi-modal pre-processing data, feature matrix, dangerous score history record, etc.). In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device.

[0067] The communication interface is used for data communication between the above-mentioned electronic device and a multi-type sensor network (including acoustic emission sensors, seismic wave recorders, high-definition cameras) or other network elements. In the edge deployment scenario, the communication interface receives real-time raw data streams from the sensor network and transmits them to the processor for time synchronization and denoising processing.

[0068] The embodiment of the present application also provides a storage medium, and a computer readable storage medium stores a computer program, which is executed by a processor to implement the above-mentioned roadway surrounding rock danger identification model construction method. The storage medium can be a tangible device that maintains and stores instructions for use by an instruction execution device. The computer readable storage medium can be, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any appropriate combination of the above. More specific examples (non-exhaustive list) include a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanical encoding device, and any appropriate combination of the above.

[0069] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A roadway surrounding rock danger identification model construction method, characterized in that, The method comprises the following steps: S1, collecting original data including acoustic emission signals, seismic wave signals and image data through a multi-type sensor network arranged in the surrounding rock area of the roadway, performing synchronous time calibration and denoising processing to generate multi-modal preprocessed data with time alignment and denoising; S2, using wavelet transform to extract frequency domain features of the seismic wave signals, and combining deep convolutional neural network to extract local and global texture features of the surrounding rock images to generate a multi-modal feature matrix; S3, using an improved clustering algorithm combined with prior geological knowledge to mine abnormal patterns from the multi-modal feature matrix, identifying abnormal signal patterns, marking feature data corresponding to the abnormal signal patterns as dangerous samples, and generating a labeled sample dataset; S4, using a deep learning algorithm based on transfer learning, integrating heterogeneous features in the multi-modal feature matrix through a feature fusion strategy, training a deep learning network using the labeled sample dataset, and generating a roadway surrounding rock danger identification model; S5, setting the roadway surrounding rock danger identification model in an edge computing device, performing inference calculation on real-time collected sensor data to obtain a danger score, adjusting an alarm strategy according to the danger score through an adaptive warning algorithm, and generating a surrounding rock stability determination result.

2. The method according to claim 1, characterized in that, In step S1, the step of performing synchronous time calibration and denoising processing further comprises: providing a unified time reference for all sensors through a GPS timestamp or a high-precision clock synchronization mechanism, recording the time offset value of each type of sensor, and reordering the data stream according to the timestamp; applying specific denoising algorithms for different data types: for acoustic emission signals and seismic wave signals, using frequency domain filtering and wavelet transform decomposition to remove high-frequency interference and low-frequency baseline drift; for image data, improving image clarity through a non-local mean algorithm or a convolutional neural network denoising model; after completing the denoising processing, performing integrity check and consistency check on the data, if abnormal or mismatched data is found, supplementing through data interpolation or reacquisition mechanism, and outputting multi-modal preprocessed data.

3. The method according to claim 1, characterized in that, The step S2 specifically comprises: segmenting the seismic wave signals in the multi-modal preprocessed data and selecting a time window, decomposing the signals into multi-scale time-frequency components using wavelet transform, and extracting frequency domain features including energy density spectrum, high-frequency signal mutation point and low-frequency trend; inputting the surrounding rock images in the multi-modal preprocessed data into a pre-trained deep convolutional neural network, extracting local texture features including micro-fissure edge characteristics and surrounding rock surface roughness changes through multi-layer convolution operation, and analyzing the distribution pattern of the overall surface texture of the surrounding rock through a fully connected layer to extract global texture features; combining the extracted frequency domain features with local and global texture features to construct the multi-modal feature matrix.

4. The method according to claim 1, characterized in that, In step S3, the step of identifying abnormal signal patterns further comprises: According to the data distribution characteristics of the multi-modal feature matrix, a density-based clustering or Gaussian mixture model is selected as an improved clustering algorithm, and prior geological knowledge is introduced to optimize the initial parameter setting of the clustering algorithm: for the Gaussian mixture model, the initial mean vector and covariance matrix are set according to the prior knowledge; for the density-based clustering, the radius parameter and the minimum number of sample points are adjusted; The clustering algorithm is executed to assign the feature points in the multi-modal feature matrix to corresponding clustering clusters, and the abnormal signal pattern pointing to the hidden fault or micro-crack expansion area is identified by judging the local density of the data points and the number of neighbors within a certain range; The clustering results are compared and verified with the prior geological knowledge, the feature data corresponding to the verified matching clustering results are taken as positive samples, and the rest are taken as negative samples to generate a labeled sample dataset.

5. The method according to claim 1, characterized in that, In step S4, the heterogeneous features in the multi-modal feature matrix are integrated by a feature fusion strategy, which further includes: A deep residual network is selected to process image data, and a deep recurrent neural network is selected to process seismic wave signal features, a pre-trained model parameter set of the deep residual network is obtained, and a set of randomly initialized parameters of the deep residual network is generated; Based on the pre-trained model parameter set and the set of randomly initialized parameters, a weighted sum is performed according to a weight factor to obtain an initialization parameter set of the model; The seismic wave signal frequency domain features and surrounding rock image data in the multi-modal feature matrix are subjected to feature splicing operations to generate a fusion feature matrix containing multi-modal features.

6. The method according to claim 1, characterized in that, In step S4, the step of generating a roadway surrounding rock danger identification model further includes: The labeled sample dataset is divided into a training set and a verification set; A loss function for model training is constructed, and the value of the loss function is calculated based on the training set, the value of the loss function being calculated by weighting based on the weight of the sample category, the real label, and the predicted value output by the model, and being used to represent the gap between the current prediction and the real label; In the training process, a mixed data augmentation technique and a dynamic dropout mechanism are introduced to randomly set zero to the dimensions of the fusion feature matrix, and the verification set is used to evaluate the model to obtain precision, recall, and F1 value to dynamically adjust the learning rate.

7. The method according to claim 1, characterized in that, In step S4, the step of training the deep learning network further includes introducing an adversarial training technique to strengthen the robustness of the model: Samples are selected from the labeled sample dataset as input samples to generate adversarial samples, the adversarial samples being obtained by superimposing a perturbation term on the input samples; The direction of the perturbation term is determined by the sign of the gradient of the loss function with respect to the input sample in the training process, and the size of the perturbation term is determined by a pre-set perturbation control coefficient.

8. The method according to claim 1, characterized in that, In step S5, the step of generating a surrounding rock stability determination result further includes: The linear combination value output by the roadway surrounding rock danger identification model is mapped to a probability form of a danger score using a Sigmoid function, and when the danger score exceeds a threshold value, the surrounding rock stability result is determined to be unstable. The dangerous score threshold is dynamically adjusted by an adaptive early warning algorithm, and the adjustment strategy is calculated based on the current dangerous score threshold, in combination with the change rate of false alarm rate over time, the change rate of missed alarm rate over time, and the acceleration of dangerous score trend, to balance the false alarm rate and the missed alarm rate.

9. A roadway surrounding rock danger identification model construction device applied to the roadway surrounding rock danger identification model construction method of any one of claims 1-8, characterized in that, Comprise: A data acquisition and processing module for acquiring raw data including acoustic emission signals, seismic wave signals and image data through a multi-type sensor network arranged in the surrounding rock area of the roadway, performing synchronous time calibration and denoising processing to generate multi-modal preprocessed data aligned in time and denoised; A feature extraction module for extracting frequency domain features of seismic wave signals using wavelet transform, and combining deep convolutional neural network to extract local and global texture features of surrounding rock images, to generate a multi-modal feature matrix; An anomaly mining module for mining abnormal patterns from the multi-modal feature matrix through an improved clustering algorithm combined with prior geological knowledge, identifying abnormal signal patterns, and marking feature data corresponding to the abnormal signal patterns as dangerous samples to generate a labeled sample dataset; A model training module for using a deep learning algorithm based on transfer learning to integrate heterogeneous features in the multi-modal feature matrix through a feature fusion strategy, training a deep learning network using the labeled sample dataset, and generating a roadway surrounding rock danger identification model; An early warning application module for setting the roadway surrounding rock danger identification model on an edge computing device, performing inference calculation on real-time collected sensor data to obtain a dangerous score, adjusting an alarm strategy according to the dangerous score through an adaptive early warning algorithm, and generating a surrounding rock stability determination result.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement a roadway surrounding rock danger identification model construction method according to any one of claims 1-8. The processor executes the computer program to implement a roadway surrounding rock danger identification model construction method according to any one of claims 1-8.

Images