Mine tunnel support intelligent decision-making method and system based on deep learning

Abstract

The present application relates to the technical field of mine safety decision, and more particularly to a mine roadway support intelligent decision method and system based on deep learning, comprising: constructing a mine multi-source sample set and calculating sample weights and sample confidence according to sample category frequency, feature space local density and category boundary neighborhood distribution; inputting segmented coding of continuous features and standardized coding of discrete features into a 3-branch deep network to extract features; fusing 3-branch features based on a channel interaction gate mechanism; constructing a joint objective function and establishing a support parameter feasible region constraint; performing feasibility checking on network output and projecting out-of-bound results into the feasible region. Through the technical scheme of the present application, rare geological working condition samples can obtain a higher weight proportion in training to alleviate sample imbalance bias, and safe support decision results can be output under engineering constraints.

Description

Technical Field

[0001] This invention relates to the field of mine safety decision-making technology. More specifically, this invention relates to a deep learning-based intelligent decision-making method and system for mine roadway support. Background Technology

[0002] Mineral resource extraction is fundamental to ensuring national energy and industrial development. In underground mine construction and production, roadway support is a core element in ensuring mine safety and maintaining normal operations. The geological conditions of the surrounding rock of a roadway are influenced by various factors, including lithology, in-situ stress, burial depth, and aquifer distribution. Different mining areas, and even different working faces within the same mining area, may exhibit differences in surrounding rock grade, cross-sectional dimensions, and in-situ stress distribution. When facing complex and variable local geological conditions, the configuration of support parameters directly affects the stability of the roadway and the safe operation of the mine. Insufficient support strength may lead to safety accidents such as roof collapse and spalling, while excessive support will result in material waste and reduced construction efficiency. How to scientifically and accurately select support types and configure parameters based on complex roadway geological environments has become an important issue in the field of mining engineering.

[0003] Currently, the determination of mine roadway support schemes mainly relies on engineering experience and engineering analogy. Technicians select parameters such as anchor bolt length and anchor cable spacing based on the surrounding rock classification results and by referring to existing support cases. This method is feasible in mining areas with simple geological conditions and small variations, but when geological conditions are complex or local abrupt changes occur, the experience-based method lacks real-time response capability, easily leading to unreasonable configuration of support parameters.

[0004] In the prior art, Chinese patent document CN117805938B discloses an intelligent prediction method for geomechanical parameters of surrounding rock based on deep learning. This method obtains drilling parameters through indoor drilling experiments and establishes a deep learning prediction model to rapidly predict the geomechanical parameters of the surrounding rock. However, this method focuses on the prediction of surrounding rock parameters and does not address the end-to-end decision mapping from surrounding rock parameters to support schemes, nor does it solve the problems of fusion of multi-source heterogeneous data and verification of engineering constraints on support parameters.

[0005] Leveraging the powerful nonlinear mapping capabilities of deep learning to uncover hidden patterns from vast amounts of historical exploration data, surrounding rock properties, construction disturbances, and monitoring information to achieve intelligent decision-making for roadway support schemes has become a development direction in this field. However, practical applications still face the following challenges: The actual data collected in mines suffers from severe sample imbalance, with normal, safe samples constituting the vast majority, while marginal samples representing special and complex geological conditions at the category boundaries are extremely scarce, leading to model training bias and inaccurate decision-making under critical and complex working conditions; existing methods typically involve simple concatenation of multi-source heterogeneous features, lacking the ability to combine coding processing with geological and engineering constraints with deep interactive fusion of multi-branch dimensional information, resulting in limitations in feature extraction; conventional deep learning models rely solely on pure data-driven approaches, failing to incorporate geological consistency, the inherent order of parameters, and the feasible domain boundaries of actual engineering into the model optimization and verification process, causing the network output decision results to often violate physical laws and engineering common sense, lacking constraint and boundary-forced correction mechanisms, leading to insufficient safety and reliability of support decision results in practical applications. Summary of the Invention

[0006] To address the technical problems of existing mine roadway support decision-making methods, such as sample imbalance leading to model bias, insufficient fusion of multi-source features, and lack of engineering constraint verification, this invention provides solutions in the following aspects.

[0007] In a first aspect, the present invention provides a deep learning-based intelligent decision-making method for mine roadway support, comprising: A multi-source sample set for the mine is constructed, which includes information on surrounding rock properties, construction disturbance and response monitoring. The sample weights and sample credibility are calculated based on the sample category frequency, local density of feature space and distribution of category boundary neighborhoods to obtain a reweighted training set. Continuous features are segmented and encoded, while discrete features are standardized and encoded. The encoded surrounding rock properties, construction disturbance, and response monitoring information are then input into a three-branch depth network to extract features. Based on a channel interaction gating mechanism, the features of the three branches are fused to obtain a representation of the mine roadway status. Using support type and support parameter combination as joint output, a joint objective function is constructed that includes category loss, parameter regression loss, parameter order constraint loss and geological consistency constraint loss. The network parameters are optimized by weighting each loss based on the sample weight and the sample confidence. Based on historical stable support samples, feasible domain constraints for support parameters related to surrounding rock grade, cross-sectional size, and burial depth are established; the feasibility of the network output is verified, and when the network output exceeds the feasible domain, the network output results exceeding the feasible domain are projected to the range that satisfies the feasible domain constraints of the support parameters, thus obtaining the intelligent decision-making results for mine roadway support.

[0008] This invention calculates sample weights and credibility based on sample category frequency, local density in the feature space, and neighborhood distribution of category boundaries. This ensures that rare geological condition samples and boundary area samples receive more attention during training, mitigating model bias caused by imbalanced mine data samples. By fusing multi-source features based on a channel interaction gating mechanism, the comprehensive perception capability of roadway state representation is improved. The joint objective function incorporates parameter order constraints and geological consistency constraints into the training process, guiding the network output to conform to engineering physics laws from the loss function level. The feasible region constraint projection mechanism enforces forced correction of out-of-bounds outputs during the inference phase, ensuring that all support parameters in the final output fall within the historically validated and safe parameter space. Simultaneously, the upper boundary surface model of the feasible region constraint sets an engineering-reasonable upper bound for the support parameters. When the support parameters output by the network exceed the economically reasonable range, projection correction is also triggered, avoiding material waste and reduced construction efficiency caused by excessive support.

[0009] Preferably, the step of calculating sample weights and sample confidence based on sample category frequency, local density in the feature space, and distribution of the category boundary neighborhood includes: The total number of samples in each category in the training set is used as the sample category frequency; Using the coordinates of each sample in the feature space as the center, the number of similar samples within a preset radius is counted as the local density of the feature space; Search the feature space for the nearest preset number of neighboring samples to the sample, and calculate the proportion of out-of-class samples among the preset number of neighboring samples as the distribution of the class boundary neighborhood; The initial sample weights are obtained by linearly weighting and summing the reciprocal of the sample category frequency, the reciprocal of the local density of the feature space, and the neighborhood distribution of the category boundary. The percentage of similar samples among the preset number of neighboring samples is used as the sample confidence level, and the initial sample weights are multiplied by the sample confidence level to obtain the sample weights used for training.

[0010] By weighting the training samples according to three dimensions—the frequency of the sample category, the local density of the feature space, and the distribution of the category boundary neighborhood—samples of lower-frequency categories and samples in sparse regions of the feature space can receive more attention during training. At the same time, the initial weights are adjusted a second time based on the sample credibility to suppress the interference of label noise samples on model parameter updates and improve the model's decision-making accuracy under complex geological conditions.

[0011] Preferably, the step of segmenting and encoding continuous features and standardizing and encoding discrete features includes: The rock mass compressive strength and roadway surface displacement measurements from the multi-source sample set are extracted as continuous features. The continuous features are then input into the K-means clustering model and clustered into a preset number of clusters. The original continuous feature values ​​are replaced with the cluster center coordinates of the cluster to which each continuous feature belongs, thus completing the segmented encoding of the continuous features. Discrete features containing rock grade and lithology codes are extracted from the multi-source sample set. Using one-hot coding rules, each discrete feature is transformed into a standard orthogonal vector composed of 0s and 1s, thus completing the standardized coding of the discrete features. Segmented coding eliminates individual noise fluctuations in continuous features, and one-hot coding eliminates pseudo-order relationships in discrete features. After the above coding processing, the numerical distribution of the input features of the three-branch network is more balanced.

[0012] Preferably, the step of inputting the encoded surrounding rock properties, construction disturbance, and response monitoring information into the three-branch depth network to extract features includes: A parallel three-branch deep network model is constructed, comprising a first-branch residual network, a second-branch convolutional neural network, and a third-branch long short-term memory network. The surrounding rock properties include stiffness assessment data. The surrounding rock properties containing stiffness assessment data are input into the first-branch residual network, and deep static spatial structure features are extracted through stacked residual modules. The construction disturbance includes a blasting impact change sequence, which is processed by the sliding convolution window of the second branch convolutional neural network to obtain local response features; the response monitoring information includes a continuous monitoring time series of displacement of the surrounding rock around the tunnel, and the third branch long short-term memory network models the time series and extracts the temporal features of geomechanical evolution over time. The three-branch architecture enables each type of feature to be extracted in a targeted manner while retaining its own information structure.

[0013] Preferably, the fusion of three-branch features based on the channel interaction gating mechanism includes: aligning the features output by the three-branch deep network in terms of dimensions and then concatenating them in the channel dimension to obtain a multi-type comprehensive feature tensor; performing global average pooling on the multi-type comprehensive feature tensor to obtain a channel feature vector, and calculating a channel attention weight vector through the channel interaction gating network; multiplying the channel attention weight vector element-wise with the multi-type comprehensive feature tensor to obtain a mine roadway state representation after multi-dimensional information aggregation processing.

[0014] After compressing spatial information through global average pooling, channel attention weights are calculated using two fully connected layers and a Sigmoid activation function. Channel-wise weighting is then performed on the spliced ​​multi-type integrated feature tensor, which suppresses background noise channels and strengthens key feature channels, making the fused lane state representation contain more compact multi-physics integrated information.

[0015] Preferably, constructing the joint objective function includes: calculating the cross-entropy between the predicted probability distribution of the support type and the true label as the category loss; calculating the mean square error between the predicted value of the support parameter and the true reference value as the parameter regression loss; calculating the difference between the predicted value of the anchor bolt length and the predicted value of the anchor cable length in the support parameter combination, and using the difference as the parameter order constraint loss when the difference is positive, and the parameter order constraint loss is zero when the difference is zero or negative; calculating the theoretical support demand strength based on the unit weight of the surrounding rock, the lateral pressure coefficient corresponding to the surrounding rock level, and the burial depth, and calculating the predicted support strength based on the support parameter combination, multiplying the difference between the two by a preset penalty coefficient as the geological consistency constraint loss when the predicted support strength is less than the theoretical support demand strength, otherwise the geological consistency constraint loss is zero; and linearly weighting and summing the four losses to obtain the joint objective function.

[0016] By linearly weighting and summing the category loss, parameter regression loss, parameter order constraint loss, and geological consistency constraint loss to construct a joint objective function, the deep network can simultaneously consider the accuracy of support type classification, parameter prediction accuracy, and satisfaction of engineering physical constraints during training. This avoids the network output violating engineering logic rules such as anchor cables being longer than anchor rods, and ensures that the predicted support strength is not lower than the theoretical safety requirements.

[0017] Preferably, the step of establishing feasible domain constraints for support parameters related to surrounding rock grade, cross-sectional dimensions, and burial depth based on historical stable support samples includes: Collect historical safety support engineering sample data without collapse failure records, extract the surrounding rock level, cross-sectional dimensions and burial depth of each sample as input variables, and extract the corresponding support parameters as target variables; The quantile polynomial surface regression algorithm is used to fit and generate the upper limit boundary surface model and lower limit boundary surface model of the safety support parameters of the target variable, with the input variable as the independent variable. The upper boundary surface model and the lower boundary surface model form inequality interval constraints in the multidimensional joint parameter space composed of the input variables and multidimensional objective variables, forming a continuous and closed set of multidimensional parameter spaces, which serves as the feasible region constraint for the support parameters.

[0018] By using the quantile polynomial surface regression algorithm to fit the upper and lower boundary surface models of the support parameters, a continuous closed multidimensional parameter space constraint related to the surrounding rock grade, cross-sectional size and burial depth is formed, providing a parameter boundary benchmark with safe redundancy for feasibility verification in the reasoning stage.

[0019] Preferably, projecting the excess network output results to a range that satisfies the feasible domain constraint of the support parameters includes: substituting the coordinate points of the support parameters output by the network, together with the surrounding rock environmental variables, into the feasible domain constraint for verification; if the coordinate points exceed the boundary range, keeping the surrounding rock grade, cross-sectional size, and burial depth unchanged, finding the projection point on the boundary surface that minimizes the Euclidean distance from the coordinate points through gradient descent search; and replacing the original network output values ​​with the parameter values ​​of the projection points.

[0020] By detecting whether the network output falls within the feasible region during the inference phase and performing gradient descent search projection on outputs that exceed the boundary, the out-of-bounds coordinates are mapped to the nearest compliant point on the boundary surface, ensuring that the support parameters of each output meet the security constraints verified in history.

[0021] Secondly, the present invention provides a deep learning-based intelligent decision-making system for mine roadway support, comprising: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned deep learning-based intelligent decision-making method for mine roadway support is implemented.

[0022] Preferably, when the processor executes the computer program instructions, the step of constructing a multi-source sample set for the mine specifically includes: counting the total number of samples of each category in the training set as the sample category frequency; taking the coordinate point of each sample in the feature space as the center, counting the number of similar samples within a preset radius as the local density of the feature space; searching for a preset number of neighboring samples closest to the sample in the feature space, calculating the proportion of dissimilar samples among the preset number of neighboring samples as the category boundary neighborhood distribution; linearly weighting and summing the inverse of the sample category frequency, the inverse of the local density of the feature space, and the category boundary neighborhood distribution to obtain the initial sample weight; taking the proportion of similar samples among the preset number of neighboring samples as the sample confidence, and multiplying the initial sample weight by the sample confidence to obtain the sample weight used for training.

[0023] The beneficial effects of this invention are as follows: This invention calculates sample weights and sample credibility based on sample category frequency, local density in the feature space, and neighborhood distribution at category boundaries. This allows rare geological condition samples to contribute a higher percentage of loss during training, mitigating model bias caused by sample imbalance. Multi-source features are segmented and standardized before being input into a three-branch deep network for feature extraction. Attention weighting is applied at the channel dimension via a channel interaction gating mechanism, forming a roadway state representation that integrates multi-physics information. A joint objective function incorporates parameter order constraint loss and geological consistency constraint loss, ensuring that the network output conforms to engineering physics principles during the training phase. A feasible region constraint projection mechanism performs forced correction on out-of-bounds outputs during the inference phase, ensuring that the output support parameters fall within a historically validated and safe parameter space, preventing insufficient or excessive support.

[0024] Furthermore, this invention eliminates noise interference from continuous features through K-means clustering segmented encoding and eliminates pseudo-order relationships from discrete features through one-hot encoding. This enables the three-branch depth network to extract high-quality features from three dimensions: surrounding rock properties, construction disturbance, and response monitoring. After selective aggregation through a channel interaction gating mechanism, a multi-physics integrated roadway state representation is formed. Combined with the feasible region constraint and gradient descent projection correction mechanism established by quantile polynomial surface regression, a safety boundary review is performed on each network output during the inference stage, correcting non-compliant results to the safety parameter space and reducing the risk of the support scheme deviating from the engineering safety range. Attached Figure Description

[0025] Figure 1 A flowchart of a deep learning-based intelligent decision-making method for mine roadway support; Figure 2 A schematic diagram of K-means clustering results for continuous features; Figure 3 A schematic diagram of the feasible region projection for support parameters; Figure 4 This is a schematic diagram of the convergence curve of the joint target loss. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0027] This invention discloses a deep learning-based intelligent decision-making method for mine roadway support, referring to... Figure 1 This includes steps S1-S4: S1. Construct a multi-source sample set and calculate the sample weights.

[0028] The system reads historical exploration reports, borehole logging data, and sensor logs from the mine via a data acquisition interface, and performs data cleaning to remove formatting errors and redundant records. Missing values ​​are filled using the K-nearest neighbor interpolation algorithm. The system connects to the mine's production database to extract discrete and continuous features such as surrounding rock grade, cross-sectional dimensions, burial depth, and geostress. Data from different sources are aligned and stitched together according to sample identifiers to construct a multi-source sample set containing three main categories of features: surrounding rock properties, construction disturbance, and response monitoring.

[0029] The total number of samples in each category in the training set is used as the sample category frequency. Centered on the coordinates of each sample in the feature space, the number of samples of the same category within a predetermined radius is counted, and this result is used as the local density of the feature space. When the number of samples of the same category within the predetermined radius is zero, the local density of the feature space is set to 1 to avoid division by zero anomalies when taking the reciprocal later. The five nearest neighbors of a sample in the feature space are searched, and the proportion of out-of-class samples among these five neighbors is calculated. This result is used as the category boundary neighborhood distribution. The reciprocal of the sample category frequency, the reciprocal of the local density of the feature space, and the category boundary neighborhood distribution are linearly weighted and summed to obtain the initial sample weights. Samples with lower frequency, lower density, and higher boundary mixing require higher attention during training. These three components measure the attention requirement of samples from three dimensions: category distribution, spatial distribution, and boundary attributes, respectively.

[0030] Initial sample weights Calculate according to the following formula: ; In the formula, For the first The frequency of the sample category to which each sample belongs. For the first Local density of the feature space of each sample For the first The distribution value of the class boundary neighborhood of each sample. , , These are the linear weighting coefficients.

[0031] The percentage of similar samples among the five nearest neighbors is used as the sample confidence level. The initial sample weights are multiplied by the sample confidence level to obtain the sample weights used for training.

[0032] Taking a specific mine's combined support category as an example, the total sample size is 10,000, with 500 samples in this combined support category, resulting in a category frequency of 500. In the preprocessed 128-dimensional multi-source feature space, with a preset radius of 0.5, 10 samples of the same category were found, resulting in a local density of 10 in the feature space. Retrieving the five nearest neighbors of this sample, and verifying the labels, found that three belonged to non-combined support categories, with a category boundary neighborhood distribution value of 0.6. For 1000, For 10, The initial sample weights are 5, which are then substituted into the relational expression to obtain the initial sample weights. The result is 1000×0.002+10×0.1+5×0.6=6. Since there are 2 samples of the same type among the 5 neighboring samples, the sample confidence is 0.4, and the final sample weight is 6×0.4=2.4.

[0033] The preset radius ranges from 0.3 to 0.8; in this embodiment, it is set to 0.5. A radius that is too small results in too few similar samples, leading to unstable density estimation. A radius that is too large includes distant, irrelevant samples, reducing the discriminative power of the density measure. Since the reciprocal of the sample category frequency, the reciprocal of the local density in the feature space, and the distribution of the category boundary neighborhood are not on the same order of magnitude, they are respectively measured using coefficients before linear weighted summation. , , Adjust the three sub-items to similar numerical ranges to make the contribution ratio of each dimension to the final weight controllable. Linear weighting coefficients. , , The grid search method determines that, in other implementations, the distribution of the actual samples can be adjusted within a reasonable range.

[0034] S2, Multi-source feature coding and three-branch feature fusion.

[0035] Rock mass compressive strength and tunnel surface displacement measurements from a multi-source sample set were extracted as continuous features. The rock mass compressive strength data ranged from 10 to 150 MPa, while the tunnel surface displacement measurements ranged from 0 to 200 mm, both exhibiting a long-tailed distribution effect. These continuous features were input into a K-means clustering model, resulting in five clusters. The K-means++ strategy was used to generate initial centroids, selecting five initial cluster centers sequentially based on the probability of the squared distance between each sample and the selected centroids. Multiple iterations were then performed, and after clustering converged, five cluster center coordinates representing the central tendency of typical data sets were obtained. (Refer to...) Figure 2This figure shows the K-means clustering results of continuous geological features. With rock mass compressive strength and tunnel surface displacement measurements on the horizontal and vertical axes respectively, it illustrates the distribution of the original monitoring data into five typical geomechanical grade clusters. Each cluster center corresponds to a different surrounding rock strength and deformation grade. The original continuous feature values ​​were replaced with the coordinates of the cluster centers of each continuous feature's respective clusters to complete the segmented coding of the continuous features. Taking rock mass compressive strength as an example, the five cluster centers were labeled as 25.5 MPa, 55.2 MPa, 85 MPa, 112.4 MPa, and 140.8 MPa, respectively. When a newly acquired monitoring value was 32 MPa, the Euclidean distance between this value and each cluster center was calculated, and it was assigned to the nearest cluster, with the center value of 25.5 MPa replacing the original data. Segmented coding reduces the interference of individual extreme high-frequency noise caused by vibrations from large underground electromechanical equipment on the data mean.

[0036] Discrete features containing surrounding rock grade codes and lithology codes are extracted from multi-source sample sets. Surrounding rock grade codes are classified into five levels, from I to V; lithology codes include sandstone, mudstone, limestone, tuff, and coal seam. Using one-hot encoding rules, each discrete feature is transformed into a standard orthogonal vector composed of 0s and 1s, completing the standardized encoding of the discrete features. Taking lithology codes as an example, a 5-dimensional orthogonal vector representation is constructed for the five lithologies; the encoding result for mudstone is 0, 1, 0, 0, 0. One-hot encoding eliminates the magnitude differences that may be introduced by numerical encoding, preventing deep networks from misinterpreting category attributes as numerical ranking relationships.

[0037] After encoding, the data is divided into feature subsets according to three dimensions: surrounding rock properties, construction disturbance, and response monitoring. A parallel three-branch deep network model is constructed, consisting of a first-branch residual network, a second-branch convolutional neural network, and a third-branch long short-term memory network.

[0038] The surrounding rock attributes, including segmented encoded values ​​of rock mass compressive strength and standardized encoded vectors of surrounding rock grade, are input into the first branch residual network. The first branch residual network contains two layers of residual modules with convolutional operations and identity fast mapping. Deep static spatial structure features are extracted through stacked residual modules, and a feature tensor with a dimension of 256 is output.

[0039] The construction disturbance information includes a blasting impact change sequence, with a sampling frequency of 500Hz and a single recorded sequence length of 1000 one-dimensional time-series impact vibration signals. The second branch convolutional neural network has a convolutional window length of 15 and a stride of 2. After multi-layer cascaded processing, local response features are extracted, and after batch normalization, a feature tensor with a dimension of 256 is output. A convolutional window length of 15 covers a sampling interval of nearly 30ms, sufficient to capture the main energy release characteristics of the blasting impact waveform.

[0040] The response monitoring information, including the continuous monitoring time series of displacement of the surrounding rock around the tunnel, is input into the third branch long short-term memory network. The response monitoring information consists of displacement and creep monitoring data of the roof and two side walls in three directions over the past 30 days, with a structure of a 30-day × 3-direction time series. The third branch long short-term memory network contains two layers of long short-term memory units with a hidden layer dimension of 128. This time series is modeled to extract the temporal characteristics of geomechanical evolution over time. The hidden vector at the end of the sequence is linearly mapped to output a feature tensor with a dimension of 256.

[0041] The three feature tensors, each with a dimension of 256, output by the three-branch deep network are aligned in dimension and then concatenated along the channel dimension to obtain a multi-type comprehensive feature tensor with a channel dimension of 768. Global average pooling is then performed on the multi-type comprehensive feature tensor to compress the features of each channel into a global statistical scalar, resulting in a 768-dimensional channel feature vector.

[0042] The channel feature vectors are input into a channel interaction gating network comprising two fully connected layers and a sigmoid activation function. The first fully connected layer performs dimensionality reduction according to a preset shrinkage ratio and performs nonlinear mapping using the ReLU activation function. The second fully connected layer restores the original 768-dimensional channel width. The output values ​​of each channel are normalized to between 0 and 1 using the sigmoid activation function to obtain the channel attention weight vectors. The specific value of the shrinkage ratio is determined by those skilled in the art through conventional experiments based on the network structure.

[0043] Each element of the channel attention weight vector is element-wise multiplied with the corresponding channel's feature map in the multi-type integrated feature tensor. Channel features with weight values ​​close to 1 are preserved and enhanced, while channel features with weight values ​​close to 0 are suppressed. This weighted enhancement of key features yields a mine roadway state representation after multi-dimensional information aggregation.

[0044] S3. Construct a joint objective function and train the network.

[0045] A dual-branch output layer is constructed at the end of the deep network: the first branch uses the Softmax function to output the probability distribution of support type prediction, and the second branch uses the linear activation function to output the specific values ​​of support parameters such as anchor bolt length, anchor cable length and anchor bolt spacing.

[0046] The cross-entropy between the predicted support type probability distribution output by the network and the true label is calculated and used as the category loss. The cross-entropy loss function is a well-known technique and will not be elaborated here.

[0047] The mean squared error between the predicted and actual reference values ​​of the support parameters output by the network is calculated as the parametric regression loss. This loss is obtained by averaging the squared differences between the predicted and actual values ​​for each support parameter dimension.

[0048] Extract the predicted anchor cable length and anchor bolt length from the network output. In mine roadway support engineering, anchor cables are anchored deep into the surrounding rock, while anchor bolts are anchored in shallower surrounding rock. The anchor cable length should be greater than the anchor bolt length. By calculating the predicted anchor bolt length minus the predicted anchor cable length, a positive difference indicates that the network output violates the logical rule that the anchor cable is longer than the anchor bolt; this positive difference is used as the parameter order constraint loss. When the difference is zero or negative, this loss is 0.

[0049] Parameter order constraint loss Calculate according to the following formula: ; In the formula, This is the predicted value for the anchor bolt length. This is the predicted value for the anchor cable length.

[0050] The theoretical support strength requirement is calculated based on the pre-set empirical formula of rock mass loosening zone pressure mechanics, according to the surrounding rock grade and burial depth.

[0051] The calculation was performed using a typical working condition with Class IV surrounding rock and a burial depth of 500m. The unit weight of the surrounding rock was... Take 25 kN / m³ as the lateral pressure coefficient corresponding to Class IV surrounding rock. Take 0.08. Based on the empirical formula of rock mass loosening zone pressure mechanics, the theoretical support strength required is... Calculate according to the following formula: ; In the formula, The unit weight of the surrounding rock is expressed in kN / m³ and is determined based on actual lithological measurements. The depth of the tunnel is expressed in meters (m). The lateral pressure coefficient corresponds to the surrounding rock grade, and its value ranges from 0.02 to 0.15. The higher the surrounding rock grade, the greater the lateral pressure coefficient. The larger it is. Substituting the above parameters, we get... The value is 25 × 500 × 0.08, and the calculated result is 1000 kN / m². The above relationship treats the surrounding rock of the roadway as an elastoplastic medium and establishes a mapping relationship between support requirements and burial depth based on the theory of loosened rock.

[0052] The support parameters output by the network are combined and converted into predicted support strength. The three parameters output by the network—anchor bolt length, anchor cable length, and anchor bolt spacing—constitute the support parameter vector, which is used to predict the support strength. Calculate according to the following formula: ; In the formula, This is the predicted value for the anchor cable length. This is the predicted value for the anchor bolt length. This is the predicted value for the anchor bolt spacing. The linear density of the ultimate pull-out bearing capacity of a single anchor body ranges from 50 to 150 kN / m; in this embodiment, it is taken as 80 kN / m. This parameter is determined by the interfacial shear strength between the anchor body and the surrounding rock, and is calibrated through pull-out tests in engineering practice. Assuming the above-mentioned Class IV surrounding rock condition, the network outputs an anchor length of 2.4 m, an anchor cable length of 6 m, and an anchor spacing of 0.8 m. Taking 80 kN / m, substituting it into the relationship yields... The result of dividing 80×6 by 0.64 and adding 80×2.4 by 0.64 is 1050kN / m².

[0053] The intensity of theoretical support requirements With predicted support strength A comparison is made. When the predicted support strength is less than the theoretical support requirement, it indicates that the support scheme output by the network is insufficient to meet the safety requirements of the geological conditions. The absolute difference between the two is calculated and multiplied by a preset penalty coefficient as the geological consistency constraint loss. When the predicted support strength is not less than the theoretical support requirement, this loss is 0. The preset penalty coefficient ranges from 5 to 50. In this embodiment, it is set to 20. When the penalty coefficient is too small, the driving force of the geological consistency constraint on the network parameter update is insufficient, and the support strength output by the network is likely to be lower than the theoretical safety requirement. When the penalty coefficient is too large, the network tends to output conservatively, and the parameter regression accuracy decreases.

[0054] The class loss, parametric regression loss, parameter order constraint loss, and geological consistency constraint loss are linearly weighted and summed according to a preset weight coefficient vector to construct a joint objective function.

[0055] The four losses are linearly combined using weighted coefficients to form a joint objective function. Calculated according to the following formula: ; In the formula, For category loss, For parametric regression loss, For parameter order constraint loss, Loss due to geological consistency constraints, , , , These are the weighting coefficients for each loss. , , , The range of values ​​for are respectively Between 0.5 and 2, Between 0.5 and 2, Between 0.1 and 1, Between 1 and 10. In this embodiment Take 1, Take 1, Take 0.5, Take 5. A larger value is chosen to ensure that the geological safety constraint receives sufficient optimization impetus during training. The smaller value is because the parameter order violation occurs less frequently in actual samples.

[0056] By utilizing sample weights and sample confidence, the loss of each sample in the joint objective function is weighted sample by sample, and the gradient is calculated and the deep network parameters are updated using the backpropagation algorithm.

[0057] S4. Establish feasible region constraints and out-of-bounds projection correction.

[0058] Historical safety support engineering sample data with no collapse failure records were collected. The screening criteria were that the tunnel had been in service for more than 3 years and the maximum displacement of the surrounding rock deformation monitoring data did not exceed 80% of the design allowable value. This embodiment collected 1200 sets of historical sample data that met the above conditions. The surrounding rock grade, cross-sectional dimensions, and burial depth of each sample were extracted as input variables, and the corresponding support parameters were extracted as target variables. A quantile polynomial surface regression algorithm was adopted, with the input variables as independent variables. A lower limit quantile was set to fit and generate a second-order surface as the lower limit boundary surface model of the safety support parameters, and an upper limit quantile was set to fit and generate an upper limit boundary surface model. The upper limit boundary surface model and the lower limit boundary surface model constituted inequality interval constraints in the multidimensional joint parameter space composed of input variables and multidimensional target variables, forming a continuous and closed multidimensional parameter space set, which served as the feasible region constraint of the support parameters.

[0059] Substitute the combined coordinate points of the multidimensional support parameters output by the network, along with the currently input surrounding rock environmental variables, into the feasible domain constraint of the support parameters to verify whether the combined coordinate points are located inside a continuous and closed set of multidimensional parameter spaces.

[0060] If a coordinate point is detected to be outside the boundary of the feasible region constraint, the boundary optimization projection algorithm is triggered. The surrounding rock grade, cross-sectional dimensions, and burial depth corresponding to the out-of-bounds coordinate point are kept unchanged, freezing the objective input state. In the dimension of the standardized support parameters, the objective function is to minimize the square of the Euclidean distance from the out-of-bounds coordinate point to the feasible region constraint boundary surface, with the constraint that the solution point lies on the boundary surface of the feasible region constraint. Using the out-of-bounds coordinate point as the initial point for gradient descent search, the gradient descent search algorithm is used to solve the objective function with a preset number of iterations and learning rate, finding the projection point on the boundary surface that minimizes the Euclidean distance to the out-of-bounds coordinate point. (Refer to...) Figure 3The figure is a schematic diagram of the feasible domain projection of support parameters. With burial depth and anchor length as variables, it shows the safety decision boundary constructed by historical stable samples. The out-of-bounds anomalies in the original network output are mapped to the nearest compliant point on the boundary after projection optimization.

[0061] Replace the original outbound network output values ​​with the compliant support parameter values ​​of the projection points to complete the feasible domain constraint projection correction operation and obtain the intelligent decision-making results for mine roadway support.

[0062] In the experimental verification, a dataset containing 10,000 historical mine roadway support samples was used, which was randomly divided proportionally into a training set containing 8,000 samples and a test set containing 2,000 samples. The evaluation metrics were set as support type classification accuracy, mean absolute error of anchor bolt length prediction, and geological compliance rate of parameter output. Figure 4 The loss reduction trends of the baseline model and various ablation models during training were compared.

[0063] To verify the contributions of each technical module, the following comparative experiments were conducted. The first group used only a conventional multi-task fully connected network as the baseline, achieving a classification accuracy of 82.5%, a mean absolute error in anchor length prediction of 0.35m, and a geological compliance rate of 81.2%. The second group added sample weight credibility calculation and segmented standard coding to the baseline network, improving the classification accuracy to 86.4%, reducing the prediction error to 0.28m, and achieving a compliance rate of 85.3%. The third group further replaced the network with a three-branch deep network and added a channel interaction gating fusion mechanism, achieving a classification accuracy of 91.7%, a prediction error of 0.21m, and a compliance rate of 89.6%. The fourth group, the complete solution of this embodiment, added a joint objective function and feasible region constraint projection verification to the third group, achieving a classification accuracy of 95.8%, a prediction error of 0.12m, and a 100% compliance rate after forced boundary constraints.

[0064] This invention also discloses a deep learning-based intelligent decision-making system for mine roadway support, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the aforementioned deep learning-based intelligent decision-making method for mine roadway support. At the software functional level, when the processor executes the computer program instructions, it sequentially completes the processing flow of the following functional modules.

[0065] The sample set construction module is used to build a multi-source sample set for the mine. The processor reads historical exploration reports, borehole logging data, and sensor logs from the mine through a data acquisition interface. After performing data cleaning and missing value imputation, it connects to the mine production database to extract features such as surrounding rock grade, cross-sectional dimensions, burial depth, and geostress. These features are then aligned and stitched together to form a multi-source sample set. The processor calculates sample weights and sample confidence based on sample category frequency, local density in the feature space, and the distribution of the category boundary neighborhood. The initial sample weights are obtained by linearly weighting and summing the inverse of the sample category frequency, the inverse of the local density in the feature space, and the distribution of the category boundary neighborhood. These initial sample weights are then multiplied by the sample confidence to obtain the sample weights used for training, thus forming a reweighted training set for deep learning.

[0066] The feature fusion module is used to segment and encode continuous features, standardize and encode discrete features, and input the encoded surrounding rock properties, construction disturbance, and response monitoring information into three-branch depth networks to extract features. The processor inputs continuous features such as rock mass compressive strength and roadway surface displacement measurements into a K-means clustering model, clustering them into a preset number of clusters, and replacing the original values ​​with cluster center coordinates to complete segmented encoding. The encoded surrounding rock properties are input into the first-branch residual network to extract deep static spatial structure features, construction disturbance information is input into the second-branch convolutional neural network to extract local response features, and response monitoring information is input into the third-branch long short-term memory network to extract temporal features. Based on a channel interaction gating mechanism, the processor concatenates the three-branch features along the channel dimension, calculates the channel attention weight vector through global average pooling and a gating network, and performs element-wise weighting on the multi-type comprehensive feature tensor to obtain a mine roadway state representation after multi-dimensional information aggregation processing.

[0067] The objective function and constraint establishment module is used to construct a joint objective function and establish feasible region constraints for support parameters, using support type and support parameter combination as joint outputs. The processor sets up a dual-branch output layer at the end of the deep network, outputting the predicted probability distribution of support type and the numerical values ​​of support parameters respectively. It calculates the category loss, parameter regression loss, parameter order constraint loss, and geological consistency constraint loss, and linearly weights and sums these four losses to construct the joint objective function. Then, based on sample weights and sample confidence, it performs backpropagation to optimize the deep network parameters. The processor collects historical safe support engineering sample data without collapse failure records and uses a quantile polynomial surface regression algorithm to fit and generate upper and lower boundary surface models of safe support parameters. These two boundary models constitute inequality interval constraints within the multidimensional joint parameter space, forming a continuous and closed multidimensional parameter space set, which serves as the feasible region constraint for support parameters.

[0068] The feasibility verification module is used to verify the feasibility of the network output. When the network output exceeds the feasible region, the excess network output is forcibly projected to the range that satisfies the feasible region constraints of the support parameters, thus obtaining a safe and reliable intelligent decision-making result for mine roadway support. The processor substitutes the coordinate points of the multi-dimensional support parameters combined with the surrounding rock environmental variables from the network output into the feasible region constraints of the support parameters for verification. If the coordinate points are located within a continuous and closed set of multi-dimensional parameter spaces, they are directly output; if the coordinate points exceed the boundary range, the boundary optimization projection algorithm is triggered. During the projection process, the surrounding rock level, cross-sectional dimensions, and burial depth remain unchanged. The Euclidean distance from the boundary coordinate points to the feasible region constraint boundary surface in the dimension of standardized support parameters is calculated. The projection point on the boundary surface that minimizes the Euclidean distance to the boundary coordinate points is found through gradient descent search. The original network output values ​​of the boundary points are replaced with the compliant support parameter values ​​of the projected points, completing the feasible region constraint projection correction operation.

[0069] The system also includes components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0070] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

Claims

1. A deep learning-based intelligent decision-making method for mine roadway support, characterized in that, include: S1. Construct a multi-source sample set for the mine, which includes information on surrounding rock properties, construction disturbance, and response monitoring. Calculate sample weights and sample confidence based on sample category frequency, local density in the feature space, and category boundary neighborhood distribution. This includes: counting the total number of samples of each category in the training set as the sample category frequency; counting the number of similar samples within a preset radius centered on the coordinates of each sample in the feature space as the local density in the feature space; searching for a preset number of nearest neighboring samples in the feature space, calculating the proportion of dissimilar samples among these nearest neighboring samples as the category boundary neighborhood distribution; linearly weighting and summing the inverse of the sample category frequency, the inverse of the local density in the feature space, and the category boundary neighborhood distribution to obtain the initial sample weights; using the proportion of similar samples among the preset number of neighboring samples as the sample confidence, and multiplying the initial sample weights by the sample confidence to obtain the sample weights used for training, thus obtaining a reweighted training set. S2. Continuous features are segmented and encoded, and discrete features are standardized and encoded. The encoded surrounding rock properties, construction disturbance, and response monitoring information are input into a three-branch deep network to extract features. This includes: constructing a parallel three-branch deep network model containing a first-branch residual network, a second-branch convolutional neural network, and a third-branch long short-term memory network; the surrounding rock properties include stiffness assessment data, which are input into the first-branch residual network, and deep static spatial structure features are extracted through stacked residual modules; the construction disturbance includes blasting impact change sequences, which are processed by the sliding convolution window of the second-branch convolutional neural network to obtain local response features; the response monitoring information includes a continuous monitoring time series of surrounding rock displacement around the roadway, which is modeled by the third-branch long short-term memory network to extract the temporal features of geomechanical evolution over time; and fusing the three-branch features based on a channel interaction gating mechanism to obtain a representation of the mine roadway state. S3. Calculate the cross-entropy between the predicted probability distribution of the support type and the true label as the category loss; calculate the mean square error between the predicted value of the support parameter and the true reference value as the parameter regression loss; calculate the difference between the predicted value of the anchor bolt length and the predicted value of the anchor cable length in the support parameter combination. When the difference is positive, use the difference as the parameter order constraint loss; when the difference is zero or negative, the parameter order constraint loss is zero; calculate the theoretical support demand strength based on the unit weight of the surrounding rock, the lateral pressure coefficient corresponding to the surrounding rock level, and the burial depth; calculate the predicted support strength based on the support parameter combination; when the predicted support strength is less than the theoretical support demand strength, multiply the difference between the two by a preset penalty coefficient as the geological consistency constraint loss; otherwise, the geological consistency constraint loss is zero; linearly weight and sum the four losses to obtain the joint objective function; train the network parameters by weighting each loss based on the sample weight and the sample credibility. S4. Establish feasible domain constraints for support parameters related to surrounding rock level, cross-sectional dimensions, and burial depth based on historical stable support samples. This includes: collecting historical safe support engineering sample data without collapse failure records; extracting the surrounding rock level, cross-sectional dimensions, and burial depth of each sample as input variables; and extracting the corresponding support parameters as target variables. Using a quantile polynomial surface regression algorithm, with the input variables as independent variables, fit and generate upper and lower boundary surface models for the safe support parameters of the target variables. The upper and lower boundary surface models constitute inequality interval constraints within the multidimensional joint parameter space composed of the input variables and multidimensional target variables, forming a continuous and closed multidimensional parameter space set as the feasible domain constraint for support parameters. Perform feasibility verification on the network output. When the network output exceeds the feasible domain, project the excess network output results to the range that satisfies the feasible domain constraint for support parameters to obtain the intelligent decision-making result for mine roadway support.

2. The method according to claim 1, characterized in that, The segmented encoding of continuous features and the standardized encoding of discrete features include: The rock mass compressive strength and roadway surface displacement measurements from the multi-source sample set are extracted as continuous features. The continuous features are then input into the K-means clustering model and clustered into a preset number of clusters. The original continuous feature values ​​are replaced with the cluster center coordinates of the cluster to which each continuous feature belongs, thus completing the segmented encoding of the continuous features. Discrete features containing surrounding rock grade codes and lithology codes are extracted from the multi-source sample set. Using the unique thermal coding rule, each discrete feature is transformed into a standard orthogonal vector composed of 0 and 1, thus completing the standardized coding of the discrete features.

3. The method according to claim 1, characterized in that, The method of fusing three-branch features based on channel interaction gating mechanism includes: aligning the features output by the three-branch deep network in terms of dimensions and then concatenating them in the channel dimension to obtain a multi-type comprehensive feature tensor; performing global average pooling on the multi-type comprehensive feature tensor to obtain a channel feature vector, and calculating a channel attention weight vector through the channel interaction gating network; and multiplying the channel attention weight vector element-wise with the multi-type comprehensive feature tensor to obtain a mine roadway state representation after multi-dimensional information aggregation processing.

4. The method according to claim 1, characterized in that, The step of projecting the excess network output results to a range that satisfies the feasible domain constraint of the support parameters includes: substituting the coordinate points of the support parameters output by the network, together with the surrounding rock environmental variables, into the feasible domain constraint for verification; if the coordinate points exceed the boundary range, keeping the surrounding rock grade, cross-sectional size, and burial depth unchanged, finding the projection point on the boundary surface that minimizes the Euclidean distance from the coordinate points through gradient descent search; and replacing the original network output values ​​with the parameter values ​​of the projection points.

5. A deep learning-based intelligent decision-making system for mine roadway support, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the intelligent decision-making method for mine roadway support based on deep learning as described in any one of claims 1 to 4.

Images