A Smart Grouting Sensing Anchor Bolt Control Method and System Applicable to Layered Soft Rock
By acquiring and processing anchor bolt measurement data in real time through an intelligent grouting sensing anchor bolt system, and optimizing grouting decisions using generative adversarial filling networks and unsupervised clustering algorithms, the blindness and passivity of traditional grouting methods are solved, achieving precise reinforcement and material optimization, and improving the safety and economy of the project.
Patent Information
- Application Number
- CN202611264434.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-10
AI Technical Summary
Traditional anchor grouting methods lack real-time sensing and feedback, making it impossible to accurately deliver grout to critical weak areas. This results in poor reinforcement effects, material waste, and the inability to dynamically adjust construction plans, leading to blindness and passivity, and uncertainty in project safety and economy.
By acquiring the radial stress and axial strain values at the measuring points on the anchor bolts, a generative adversarial filling network is used to fill in the missing data. Combined with an unsupervised clustering algorithm, the rock mass state is classified, priority scores are calculated, grouting decisions are made, and the grouting effect is fed back to form a closed-loop control.
It enables real-time monitoring of stress distribution and deformation within the surrounding rock, precise delivery of grout to weak areas, improved reinforcement effect and material utilization, suppression of bias pressure effect, and enhanced engineering safety and economy.
Smart Images

Figure CN123067259A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent control, and in particular relates to an intelligent grouting sensing anchor bolt control method and system suitable for layered soft rock. Background Technology
[0002] With the rapid development of tunnel engineering and underground space development technology, anchor grouting support technology for layered soft rock masses has emerged. This technology uses hollow grouting anchors or prestressed anchors to inject grout into the borehole, so that the anchors are bonded to the rock mass, thereby reinforcing the surrounding rock. It has the advantages of simple construction and relatively low cost, which leads to the currently widely used empirical grouting methods or traditional methods.
[0003] In traditional techniques, full-length or segmented pressure grouting is usually carried out by pre-designed fixed grouting parameters. The grout flows freely along the path of least resistance. Construction personnel cannot grasp the stress distribution inside the surrounding rock and the stress state of the rod in real time. The evaluation of the support effect mainly relies on the later surface displacement monitoring or manual experience judgment, lacking effective perception and feedback during the construction process.
[0004] Current empirical grouting methods or traditional approaches suffer from high levels of blindness, passivity, and lack of feedback loops. Specifically, they cannot detect the exact location of weak rock surfaces or the severity of eccentric loads before or after anchor installation. The grouting process cannot accurately deliver grout to critical weak areas, resulting in poor reinforcement effects and material waste. Furthermore, the support effect cannot be quantitatively assessed after grouting, and the construction plan cannot be dynamically adjusted according to the actual response of the surrounding rock, leading to significant uncertainties in both project safety and economy. Summary of the Invention
[0005] Therefore, it is necessary to provide an intelligent grouting sensing anchor control method and system suitable for layered soft rock that can monitor the stress distribution and deformation of the surrounding rock in real time, improve the reinforcement effect and material utilization, and suppress the bias effect, in order to address the above-mentioned technical problems.
[0006] In a first aspect, this application provides an intelligent grouting sensing anchor control method suitable for layered soft rock, comprising:
[0007] The radial stress and axial strain values at each measuring point on the anchor bolt are obtained to obtain the original sensing data matrix. Based on the generative adversarial filling network, the missing values of the original sensing data matrix are filled to obtain the sensing data matrix.
[0008] Based on the sensor data matrix, cluster analysis is performed using an unsupervised clustering algorithm to obtain a rock mass state classification map; the rock mass state classification map includes high-risk weak areas, stable transition areas, and high-stress bearing areas;
[0009] Based on the rock mass condition classification map and sensor data matrix, the priority score of each measuring point is calculated, and based on the priority score, the grouting priority decision is made to obtain the grouting priority control command; among which, the grouting priority control command is used to instruct the anchor to perform the grouting action and to feed back the sensor data after grouting.
[0010] Based on the sensor data after grouting, the grouting effect is evaluated, and an evaluation conclusion is obtained; the evaluation conclusion is used to indicate the next round of grouting.
[0011] Furthermore, based on a generative adversarial imputation network, missing values are imputed in the original sensor data matrix to obtain a sensor data matrix, including:
[0012] Traverse the original sensor data matrix and mark the mask matrix positions corresponding to elements with empty values or values exceeding the preset physical range as to be filled, thus obtaining the missing value mask matrix.
[0013] The original sensor data matrix and the missing value mask matrix are input into the generator of the generative adversarial imputation network to generate missing values and obtain the adversarial generative data matrix.
[0014] Based on the missing value mask matrix, the missing location features are extracted from the original sensor data matrix to obtain the missing features;
[0015] The missing features are input into the random forest model to predict the missing values, and the random forest imputation data matrix is obtained.
[0016] Based on the missing value mask matrix, the original sensor data matrix, the adversarial generative data matrix, and the random forest imputation data matrix are weighted and summed to obtain the sensor data matrix.
[0017] Furthermore, the generative adversarial filling network is trained using the following method:
[0018] Based on historical complete sensor data, a training missing mask matrix is randomly generated, and based on the training missing mask matrix, the historical complete sensor data is replaced with the true values to obtain the missing sample matrix.
[0019] Based on the missing sample matrix and the training missing mask matrix, a cue matrix is constructed. Then, based on the cue matrix, the missing sample matrix, the training missing mask matrix, and historical complete sensor data, the generator and discriminator are iteratively trained through iterative training steps until the loss value converges, thus obtaining the generative adversarial incomplete network.
[0020] The iterative training steps include:
[0021] S1: Input the missing sample matrix and the training missing mask matrix into the generator to generate the imputation data and obtain the generated data matrix;
[0022] S2: Concatenate the cue matrix and historical complete sensor data into positive samples, and concatenate the generated data matrix and cue matrix into negative samples. Train the discriminator based on binary cross-entropy loss to obtain the updated discriminator.
[0023] S3: Input the generated data matrix and the prompt matrix into the updated discriminator to predict the original observations, obtain the prediction probability, and train the generator based on the prediction probability with the goal of minimizing the generator loss value, thus obtaining the updated generator.
[0024] Furthermore, based on the predicted probabilities, the generator is trained with the goal of minimizing the generator loss value, resulting in an updated generator, including:
[0025] The formula for calculating the generator loss value is:
[0026] ;
[0027] in, Let be the generator loss function. To help the generator combat loss, For the trade-off function, Reconstruct the loss for the generator.
[0028] Furthermore, based on the sensor data matrix, unsupervised clustering algorithms are used for cluster analysis to obtain a rock mass state classification map, including:
[0029] Traverse the sensor data matrix, extract the spatial coordinates, radial stress values and axial strain values of each measuring point, and stitch them together to obtain the feature dataset;
[0030] Based on the feature dataset, the K-Means++ algorithm is used to initialize the cluster centers, resulting in initial cluster centers.
[0031] Based on the Euclidean distance from each measurement point in the feature dataset to each initial cluster center, each measurement point is assigned to each initial cluster center to obtain a list of category labels;
[0032] Based on the list of category labels, the cluster centers of the feature dataset are updated to obtain the updated cluster centers. Then, based on the Euclidean distance between the updated cluster centers and the initial cluster centers, a convergence check is performed to obtain the convergence flag.
[0033] Based on the convergence criteria, the rock mass state is classified according to the updated cluster centers to obtain classification labels. Based on the spatial coordinates, each measuring point is assigned to a classification label to obtain a rock mass state classification map.
[0034] Furthermore, based on the post-grouting sensor data, the grouting effect is evaluated, and the evaluation conclusions are as follows:
[0035] The radial stress values of each earth pressure cell measuring point are extracted from the sensor data after grouting, and the standard deviation of the radial stress values is calculated to obtain the pressure uniformity value.
[0036] The axial strain values of each fiber optic strain sensor measurement point are extracted from the grouting post-sensing data, and linear regression calculation is performed based on the axial strain values to obtain the strain stability rate value.
[0037] The pressure equalization value and strain stability rate value are compared with the preset stability threshold to obtain the evaluation conclusion.
[0038] Furthermore, based on the rock mass condition classification map and the sensor data matrix, the priority score for each measuring point was calculated, including:
[0039] The formula for calculating priority scores is:
[0040] ;
[0041] in, For the first The first anchor bolt Priority score for each measurement point For the first The first anchor bolt Normalized fractions of radial stress values at each measuring point For the first The first anchor bolt Normalized fractions of axial strain values at each measuring point For the first The first anchor bolt The normalized score of the risk clustering results for each measurement point. , and These are the weighting coefficients.
[0042] Secondly, this application also provides an intelligent grouting sensing anchor control system suitable for layered soft rock, comprising:
[0043] The filling module is used to obtain the radial stress and axial strain values of each measuring point on the anchor bolt, obtain the original sensing data matrix, and fill in the missing values of the original sensing data matrix based on the generative adversarial filling network to obtain the sensing data matrix.
[0044] The clustering module is used to perform cluster analysis based on the sensor data matrix using an unsupervised clustering algorithm to obtain a rock mass state classification map. The rock mass state classification map includes high-risk weak areas, stable transition areas, and high-stress bearing areas.
[0045] The decision module is used to calculate the priority score of each measuring point based on the rock mass condition classification map and the sensor data matrix, and to make grouting priority decisions based on the priority scores to obtain grouting priority control instructions. The grouting priority control instructions are used to instruct the anchor bolts to perform grouting actions and to feed back the sensor data after grouting.
[0046] The evaluation module is used to evaluate the grouting effect based on the sensor data after grouting and obtain the evaluation conclusion; the evaluation conclusion is used to indicate the next round of grouting.
[0047] Thirdly, this application also provides a computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement any step of the method provided in the first aspect of this application.
[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any step of the method provided in the first aspect of this application.
[0049] The aforementioned intelligent grouting sensing anchor control method and system applicable to layered soft rock acquires the radial stress and axial strain values at each measuring point on the anchor to obtain an original sensing data matrix. Based on a generative adversarial network (GAN), missing values in the original sensing data matrix are filled to obtain another sensing data matrix. Based on the sensing data matrix, an unsupervised clustering algorithm is used for cluster analysis to obtain a rock mass state classification map. This map includes high-risk weak zones, stable transition zones, and high-stress bearing zones. Based on the rock mass state classification map and the sensing data matrix, priority scores for each measuring point are calculated. Based on these priority scores, grouting priority decisions are made to obtain grouting priority control instructions. These instructions instruct the anchor to perform grouting actions and provide feedback on post-grouting sensing data. Based on the post-grouting sensing data, the grouting effect is evaluated to obtain an evaluation conclusion. This evaluation conclusion is used to indicate the next round of grouting. It can monitor the stress distribution and deformation of the surrounding rock in real time, and know the specific location of the weak surface of the bedding and the severity of the eccentric load; the grout resources can be accurately delivered to the key weak areas that need reinforcement, which greatly improves the reinforcement effect and material utilization, effectively suppresses the eccentric effect, and significantly improves the economy; it enables the support system to have self-optimization capabilities, and improves the safety and reliability of the project. Attached Figure Description
[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0051] Figure 1 This is a schematic diagram of the process of an intelligent grouting sensing anchor control method for layered soft rock provided in an embodiment of the present invention;
[0052] Figure 2 This is a schematic diagram of the structure of an intelligent grouting sensing anchor control system for layered soft rock, provided in an embodiment of the present invention. Detailed Implementation
[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0054] In one embodiment, such as Figure 1 As shown, an intelligent grouting sensing anchor control method suitable for layered soft rock is provided. This embodiment illustrates the method's application to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:
[0055] Step 101: Obtain the radial stress and axial strain values at each measuring point on the anchor rod to obtain the original sensing data matrix. Then, based on the generative adversarial filling network, fill in the missing values in the original sensing data matrix to obtain the sensing data matrix.
[0056] Among them, the anchor bolt is a rod-shaped component used to reinforce the rock mass. It integrates a multi-source sensor array inside, which can sense the mechanical state of the surrounding rock and perform grouting.
[0057] The measuring point is the specific location where the sensor is placed on the anchor bolt, and each measuring point corresponds to a set of sensor readings.
[0058] The radial stress value is the contact stress perpendicular to the anchor bolt axis, measured by a micro earth pressure cell, reflecting the degree of compression of the anchor bolt by the surrounding rock.
[0059] The axial strain value is the strain along the anchor bolt axis measured by a fiber optic strain sensor, reflecting the tensile or compressive deformation of the anchor bolt itself.
[0060] The raw sensor data matrix is a two-dimensional data structure consisting of radial stress and axial strain values of all measuring points. Rows represent measuring point numbers, and columns represent different types of sensing variables, which may contain missing or outlier values.
[0061] Generative adversarial imputation networks (GANs) are deep learning models consisting of a generator and a discriminator, specifically designed for intelligently imputing missing values in data. This network learns the distribution patterns of the data through adversarial training, thereby generating appropriate imputation values.
[0062] The sensor data matrix is a complete data matrix obtained by applying a generative adversarial imputation network to fill in missing values on the original sensor data matrix. This matrix eliminates null values and values outside the physical range, while retaining the effective radial stress and axial strain information of all measuring points.
[0063] The terminal synchronously collects radial stress and axial strain values at each measuring point according to a preset sampling frequency using a miniature earth pressure cell array and fiber optic strain sensors installed inside the anchor bolt, and summarizes them to form a raw sensor data matrix. Due to potential electromagnetic interference or sensor transient failures in the field environment, the raw data may contain anomalous elements with empty values or values exceeding the preset physical range. Therefore, a pre-trained generative adversarial incomplete (GAI) network is used to process this matrix. The GAI traverses the raw sensor data matrix, identifies all anomalous or missing locations, and generates a corresponding missing value mask matrix, where normal locations are marked as 0 and locations to be filled are marked as 1. The terminal inputs the raw sensor data matrix and the missing value mask matrix into the generator of the GAI. The generator infers reasonable values for the locations to be filled based on the known data distribution patterns and outputs a generative adversarial data matrix. Features of the missing locations are extracted from the raw sensor data matrix and input into a random forest model for missing value prediction, resulting in a random forest-filled data matrix. Based on the missing value mask matrix, the observed values in the original sensor data matrix, the generated values in the adversarial generative data matrix, and the predicted values in the random forest imputation data matrix are weighted and summed to obtain the sensor data matrix. In this matrix, the radial stress and axial strain values at all measuring points are valid and continuous values.
[0064] Step 102: Based on the sensor data matrix, cluster analysis is performed using an unsupervised clustering algorithm to obtain a rock mass state classification map; the rock mass state classification map includes high-risk weak areas, stable transition areas, and high-stress bearing areas.
[0065] Specifically, unsupervised clustering algorithms are machine learning algorithms that do not require pre-labeling of categories. They automatically divide data points into several groups based on the similarity of the data's features. In this embodiment, the K-Means++ algorithm is used, which initializes cluster centers using a roulette wheel algorithm to improve cluster stability.
[0066] A rock mass condition classification map is a visual map drawn on a tunnel cross-section or anchor bolt distribution map by combining cluster analysis results with the spatial coordinates of measuring points. Each region in the map is assigned a rock mass condition category label, intuitively displaying the mechanical zoning within the surrounding rock.
[0067] High-risk weak areas are regions with low radial stress and high axial strain in the clustering results. These correspond to open cracks, loose fragmentation, or critical bedding planes and are areas that need to be reinforced first.
[0068] The stable transition zone is a region where both radial stress and axial strain are at intermediate levels, indicating that the rock mass in this area is relatively stable and requires less reinforcement.
[0069] The high-stress bearing zone is a region with large radial stress and small axial strain. It represents the area directly affected by the eccentric load or a complete and hard rock mass. This part bears high pressure but has small deformation.
[0070] The terminal traverses the sensor data matrix, extracting the spatial coordinates, radial stress value, and axial strain value of each measuring point. This information is then concatenated into a feature dataset, where each measuring point corresponds to a data point containing multiple feature dimensions. The K-Means++ unsupervised clustering algorithm is used to perform cluster analysis on this feature dataset. The algorithm first selects initial cluster centers using a roulette wheel algorithm, randomly choosing a data point as the first cluster center. It calculates the squared Euclidean distance from each data point to its nearest existing cluster center, and uses the proportion of this squared distance to the sum of all squared distances as a probability to select the next cluster center using the roulette wheel algorithm. This process is repeated until a predetermined number of initial cluster centers are selected. The Euclidean distance from each data point to each initial cluster center is calculated, and the data points are assigned to the cluster containing the nearest cluster center, forming an initial list of category labels. The positions of the cluster centers are updated based on the mean of all data points within each cluster, resulting in the updated cluster centers. The algorithm repeatedly assigns data points and updates cluster centers until the Euclidean distance between the updated cluster centers and the previous cluster centers is less than a set convergence threshold. At this point, the algorithm converges, and the clustering result is obtained. Based on the mean radial stress and mean axial strain of the data points within each cluster, the clusters are mapped to three rock mass state categories: high-risk weak zone, stable transition zone, and high-stress bearing zone. Combining the spatial coordinates of each measuring point, the category label of each measuring point is marked on its spatial location, generating a rock mass state classification map.
[0071] Step 103: Based on the rock mass condition classification map and the sensor data matrix, calculate the priority score of each measuring point, and make a grouting priority decision based on the priority score to obtain the grouting priority control command; wherein, the grouting priority control command is used to instruct the anchor to perform the grouting action and to feed back the sensor data after grouting.
[0072] Specifically, the priority score is a comprehensive rating calculated for each measuring point, used to quantify the urgency of grouting at that point. This score is obtained by summing the radial stress normalized score, axial strain normalized score, and risk clustering result normalized score, each multiplied by its respective weighting coefficient. A higher score indicates that the measuring point requires priority grouting.
[0073] Grouting priority decision-making is the process of sorting all measuring points from highest to lowest priority score to determine the grouting order and the grouting parameters required for each measuring point.
[0074] The grouting priority control command is an execution signal generated and sent to the smart anchor by the ground analysis platform. This command contains specific parameters such as the measurement point number requiring grouting, the grouting sequence, the grouting pressure setting value, and the grouting duration, which are used to guide the anchor to perform directional grouting actions.
[0075] Post-grouting sensing data refers to the radial stress and axial strain values re-collected by the sensor array after the anchor bolt has completed the grouting process, which are used for subsequent effect evaluation.
[0076] The terminal extracts the radial stress and axial strain values of each measuring point from the sensor data matrix, and performs normalization processing to obtain the radial stress normalization score and axial strain normalization score. It obtains the risk category of each measuring point from the rock mass condition classification map, and normalizes and assigns a value to this category to obtain the risk clustering result normalization score. The three normalization scores are multiplied by preset weighting coefficients and summed to obtain the priority score for each measuring point. The terminal sorts all the priority scores of all measuring points in descending order to generate a grouting sequence. For the higher-ranked measuring points, especially those located in high-risk weak areas, the terminal allocates higher grouting pressure and more grout volume to achieve targeted reinforcement. The decision information is encoded as grouting priority control instructions and transmitted to the corresponding smart anchor bolts via a wireless communication network. Upon receiving the instructions, the anchor bolts activate their built-in, independently controllable directional grouting mechanism, injecting grout into the target measuring point areas sequentially according to the parameters specified in the instructions. After the grouting action is completed, the sensors inside the anchor bolts immediately resume data acquisition, feeding back the post-grouting sensor data to the terminal.
[0077] Step 104: Based on the sensor data after grouting, evaluate the grouting effect and obtain the evaluation conclusion; the evaluation conclusion is used to indicate the next round of grouting.
[0078] Specifically, grouting effect evaluation is a process of quantitatively analyzing post-grouting sensor data to determine whether the grouting has achieved the expected reinforcement effect. The evaluation involves pressure uniformity and strain stability.
[0079] The assessment conclusion is the result of evaluating the grouting effect, divided into successful grouting and the need for another round of grouting. Successful grouting indicates that the surrounding rock condition has met the stability requirements, and the current round of operation can be ended or transitioned to long-term monitoring; the need for another round of grouting indicates that the surrounding rock has not yet reached the stability standard, and a new cycle needs to be started.
[0080] The next round of grouting is when the evaluation conclusion is that the standard is not met. The terminal automatically returns to the data preprocessing stage, and performs missing value filling, cluster analysis and priority decision based on the latest post-grouting sensor data, and then performs the grouting action again until the evaluation conclusion is that the grouting is successful.
[0081] The terminal extracts the radial stress values of all earth pressure cell measuring points from the post-grouting sensor data, calculates the standard deviation, and obtains the pressure uniformity index. The smaller the standard deviation, the more uniform the pressure distribution of the surrounding rock on the anchor bolt after grouting, and the better the grouting filling effect. The terminal extracts the axial strain values of all fiber optic strain sensor measuring points from the post-grouting sensor data, performs linear regression calculation on the time-varying sequence, and obtains the slope of the regression line, which is the strain stability index. The smaller the absolute value of the slope, the more stable the axial deformation of the anchor bolt, and the more effectively the convergence deformation of the surrounding rock is controlled. The terminal compares the calculated pressure uniformity and strain stability with preset stability thresholds. If the pressure uniformity is lower than the threshold and the absolute value of the strain stability is less than the threshold, the grouting effect is deemed qualified, the evaluation conclusion is successful, and the terminal ends this round of operation or switches to long-term monitoring mode. If either index fails to reach the threshold, the grouting effect is deemed insufficient, and the evaluation conclusion is that another round of grouting is required. At this point, the terminal automatically uses the post-grouting sensor data as new input, jumps to the data preprocessing and cluster analysis stage, and re-executes the entire process of missing value filling, rock mass state classification, priority decision and grouting execution, forming a closed-loop feedback control until the surrounding rock state meets the stability requirements.
[0082] This embodiment provides an intelligent grouting sensing anchor control method and system suitable for layered soft rock. It acquires radial stress and axial strain values at various measuring points on the anchor to obtain an original sensing data matrix. Based on a generative adversarial network (GAN), missing values in the original sensing data matrix are filled to obtain another sensing data matrix. Based on the sensing data matrix, an unsupervised clustering algorithm is used to perform cluster analysis, resulting in a rock mass state classification map. This map includes high-risk weak zones, stable transition zones, and high-stress bearing zones. Based on the rock mass state classification map and the sensing data matrix, priority scores for each measuring point are calculated. Based on these priority scores, grouting priority decisions are made, resulting in grouting priority control instructions. These instructions instruct the anchor to perform grouting actions and provide feedback on post-grouting sensing data. Based on the post-grouting sensing data, the grouting effect is evaluated, yielding an evaluation conclusion. This evaluation conclusion is used to indicate the next round of grouting. Through the above methods, the stress distribution and deformation of the surrounding rock can be monitored in real time, the specific location of the weak bedding planes and the severity of the eccentric load can be determined, the grout resources can be accurately delivered to the key weak areas that need the most reinforcement, greatly improving the reinforcement effect and material utilization, effectively suppressing the eccentric effect and significantly improving the economy, and enabling the support system to have self-optimization capabilities, thereby improving the safety and reliability of the project.
[0083] In one embodiment, a generative adversarial imputation network is used to impute missing values in the original sensing data matrix to obtain a sensing data matrix, including:
[0084] Step 201: Traverse the original sensor data matrix and mark the mask matrix positions corresponding to elements with empty values or values exceeding the preset physical range as to be filled, thus obtaining the missing value mask matrix.
[0085] Specifically, the mask matrix is a binary matrix with the same shape as the original sensor data matrix, used to mark whether each element needs to be filled. Each position in the matrix takes a value of 0 or 1, where 0 indicates that the data at that position is normal, and 1 indicates that the position needs to be filled.
[0086] The preset physical range is a reasonable upper and lower limit for radial stress and axial strain values that are pre-set based on the sensor range and rock mechanics characteristics. Values exceeding this range are considered abnormal values.
[0087] The missing value mask matrix is a mask matrix generated after traversal marking. All elements with empty values or values exceeding a preset physical range are marked as 1, and the rest are marked as 0.
[0088] The terminal reads the original sensor data matrix and creates a blank mask matrix of the same size, initially with all elements set to 0. It then iterates through each element of the original sensor data matrix row by row and column by column. For each element, the terminal performs two checks: first, whether the element's value is empty; and second, whether the element's value exceeds a preset physical range. If either condition is met, the value at the corresponding position in the mask matrix is changed from 0 to 1, indicating that the data at that position needs to be filled. After iterating through all elements, all positions marked as 1 in the mask matrix constitute a missing value mask matrix. This matrix precisely records all abnormal or missing positions in the original sensor data matrix that need to be processed.
[0089] Step 202: Input the original sensor data matrix and the missing value mask matrix into the generator of the generative adversarial imputation network to generate missing values and obtain the adversarial generative data matrix.
[0090] Specifically, the generator in a Generative Adversarial Completion Network (GAN) is a subnetwork within the GAN. Its function is to infer and generate reasonable values for missing positions based on partial observation data and mask information. This generator consists of multiple layers of neural networks, and its training objective is to make the generated data as close as possible to the distribution of real data.
[0091] Missing value generation is the process by which a generator uses information from known data and the missing locations indicated by a mask to calculate the corresponding filler value for each missing location through forward propagation.
[0092] The adversarial generative data matrix is a complete data matrix output by the generator after the original sensor data matrix and the missing value mask matrix are input into the generator. This matrix retains the original values where the original observation data remains unchanged, while filling in the values predicted by the generator at the positions marked as 1 in the mask.
[0093] The terminal takes the original sensor data matrix and the missing value mask matrix as two independent inputs, both fed into the generator of a pre-trained generative adversarial imputation network. The generator extracts features from known locations in the original data matrix using convolutional or fully connected layers, and combines this with the missing value location information provided by the missing value mask matrix to learn the overall spatial distribution pattern and correlations between variables. For each location marked 1 in the mask matrix, the generator generates an imputation value that matches its context. For normal locations marked 0 in the mask matrix, the generator makes no modifications and directly retains the original observation value. The generator outputs a generative adversarial data matrix with the exact same shape as the original sensor data matrix, where all originally missing or abnormal locations have been replaced with generated values, while normal locations remain unchanged.
[0094] Step 203: Based on the missing value mask matrix, extract the missing location features from the original sensor data matrix to obtain the missing features.
[0095] The missing location feature extraction process extracts only the feature information corresponding to the positions marked as 1 by the missing value mask matrix from the original sensor data matrix. The features include the neighborhood statistics around the position, the historical data of the measuring point at other times, and the correlation data of other measuring points on the same anchor.
[0096] Missing features are feature vectors or feature matrices obtained after extracting features at missing locations. They are specifically used to describe the data context of each missing location so that the random forest model can make predictions.
[0097] The terminal uses a missing value mask matrix as an index to filter out all locations marked as 1 from the original sensor data matrix. For each such missing location, the terminal constructs a set of auxiliary features around that location. For example, features may include, but are not limited to, the spatial coordinates of the measuring point at that location, the mean and variance of the non-missing data for the same measuring point within adjacent time windows, the radial stress and axial strain values of neighboring measuring points around that location, and the overall stress-strain trend of the anchor bolt to which that measuring point belongs. The terminal organizes these features into a feature matrix, where each row corresponds to a missing location and each column corresponds to a feature dimension. The missing features condense the available information around each missing location.
[0098] Step 204: Input the missing features into the random forest model to predict missing values and obtain the random forest imputation data matrix.
[0099] The random forest model is a machine learning model that integrates multiple decision trees. It improves prediction accuracy and stability by aggregating the prediction results of multiple trees. This model has been trained on complete historical sensor data and can predict the corresponding radial stress or axial strain values based on the missing features of the input.
[0100] Missing value prediction is the process of inputting missing features into a random forest model, and the model outputs a predicted value for each missing position.
[0101] The random forest imputation matrix is a complete data matrix formed by filling in the missing locations with the predictions of the random forest model according to the index positions of the original sensor data matrix. In this matrix, the original observation values are retained for normal locations, and the predicted values of the random forest are filled in for missing locations.
[0102] The terminal inputs the missing features into a pre-trained random forest model. The random forest model consists of multiple decision trees, each independently providing a prediction value based on the input features. The model takes the average of all tree predictions as its output. For each row in the missing feature matrix, the model outputs a specific numerical value, which is an estimate of the radial stress or axial strain at that location. The terminal collects the prediction results for all missing locations and combines them with the observations of normal locations in the original sensor data matrix to construct a random forest imputation data matrix with the exact same shape as the original sensor data matrix. In this matrix, all locations originally masked as 1 are replaced with the random forest's prediction value, while normal locations remain unchanged.
[0103] Step 205: Based on the missing value mask matrix, perform a weighted summation of the original sensor data matrix, the adversarial generation data matrix, and the random forest imputation data matrix to obtain the sensor data matrix.
[0104] Specifically, weighted summation is an operation that combines elements at the same position in the three data matrices according to the instructions of the mask matrix. For normal positions with a mask of 0, only the observations in the original sensor data matrix are used; for positions to be filled with a mask of 1, the corresponding values of the adversarial generative data matrix and the random forest filling data matrix are linearly combined according to preset weights.
[0105] The sensor data matrix is a complete, continuous, and unmissing data matrix, where the value at each location is determined by a weighted summation. This matrix will serve as the basis for subsequent cluster analysis and decision-making.
[0106] The terminal iterates element-wise through the original sensor data matrix, the adversarial generative data matrix, and the random forest imputation data matrix, as well as a missing value mask matrix of the same size. For any position marked 0 in the mask matrix, the terminal directly uses the value of that position in the original sensor data matrix as the result. For any position marked 1 in the mask matrix, the terminal reads the values from the adversarial generative data matrix and the random forest imputation data matrix at that position, and performs a weighted sum according to a preset fusion weight coefficient to obtain the imputation value at that position. The terminal stores the results of all positions into a new data matrix, obtaining the sensor data matrix. All missing and outlier values in the original sensor data matrix have been properly imputed, and the sensor data matrix can be directly used for subsequent unsupervised clustering analysis.
[0107] In one embodiment, the generative adversarial filling network is trained using the following method:
[0108] Step 301: Based on historical complete sensor data, a training missing mask matrix is randomly generated, and based on the training missing mask matrix, the historical complete sensor data is replaced with real values to obtain a missing sample matrix.
[0109] Among them, the historical complete sensor data is a complete sensor data matrix that was collected in the past without any missing or abnormal data. It contains the radial stress value and axial strain value of all measuring points and is used as training data.
[0110] The training missing mask matrix is a binary matrix with the same shape as the historical complete sensor data, where some positions are randomly set to 1 and the rest to 0, to simulate data missing patterns in real-world scenarios.
[0111] True value replacement involves removing the values from the historical complete sensor data that correspond to the positions marked as 1 in the training missing mask matrix, in order to simulate the operation of the missing state.
[0112] The missing sample matrix is a data matrix obtained after replacing the true values. The original complete data becomes missing at the marked positions, while the original values are retained at the other positions. It is used to train the generative adversarial imputation network.
[0113] The terminal randomly generates a training missing data mask matrix with the exact same size as the historical complete sensor data matrix. According to a preset missing proportion, a corresponding proportion of locations are randomly selected from all positions and marked as 1 in the mask matrix, with the remaining locations marked as 0. The terminal processes the historical complete sensor data based on this mask matrix. For each location marked as 1 in the mask matrix, the corresponding value in the historical complete sensor data is replaced with a symbol indicating a missing value; for locations marked as 0 in the mask matrix, the original value is retained. The resulting data matrix after this replacement operation is the missing sample matrix. This matrix contains some known true values and some deliberately hidden missing values for subsequent training.
[0114] Step 302: Based on the missing sample matrix and the training missing mask matrix, a cue matrix is constructed. Based on the cue matrix, the missing sample matrix, the training missing mask matrix, and historical complete sensor data, the generator and discriminator are iteratively trained through iterative training steps until the loss value converges, thus obtaining the generative adversarial incomplete network.
[0115] Specifically, the cue matrix is an auxiliary information matrix used to provide additional clues about missing patterns to the generator and discriminator. The terminal generates a random binary matrix of the same size as the training missing mask matrix, where each element is set to 1 with a preset probability of 0.9 and 0 with a probability of 0.1. For each position, if the cue mask is 1, the value at that position in the cue matrix is equal to the value at the corresponding position in the training missing mask matrix; if the cue mask is 0, the value at that position in the cue matrix is set to neutral. The cue matrix partially reveals information about the missing positions, but does not completely expose them.
[0116] The generator is a subnetwork in the generative adversarial imputation network. It consists of multiple layers of neural networks and is responsible for generating imputation data based on the missing sample matrix and the training missing mask matrix.
[0117] The discriminator is another subnetwork in the generative adversarial imputation network. It consists of multiple layers of neural networks and is responsible for distinguishing whether the input data is real, complete data or data imputed by the generator.
[0118] Iterative training is a process of repeatedly executing S1, S2, and S3, updating the parameters of the generator and discriminator each time, until the loss function value no longer decreases significantly.
[0119] The terminal constructs a cue matrix based on the missing sample matrix and the training missing mask matrix. The cue matrix provides the model with partial information about which locations are missing, but does not fully expose them, thus enhancing the robustness of training. The terminal then enters an iterative training loop, executing the three sub-steps S1, S2, and S3 sequentially until the loss value converges.
[0120] The iterative training steps include:
[0121] S1: Input the missing sample matrix and the training missing mask matrix into the generator to generate the imputation data and obtain the generated data matrix.
[0122] Specifically, data imputation generation is the process by which a generator infers reasonable values for missing locations based on the input missing sample matrix and mask matrix through network computation.
[0123] The generated data matrix is a data matrix with the same size as the input output from the generator. In this matrix, all positions marked as 1 in the trained missing mask matrix are filled with generated values, while positions marked as 0 retain the original observations from the missing sample matrix.
[0124] The terminal takes the missing sample matrix and the training missing mask matrix as two inputs and feeds them into the generator. The generator extracts features from known locations in the missing sample matrix using its internal convolutional and fully connected layers, and combines this with the missing location information provided by the training missing mask matrix to learn the overall distribution pattern of the data and the correlation between variables. For each location marked 1 in the training missing mask matrix, the generator generates a filler value that matches its context; for locations marked 0, the generator makes no modification and directly retains the original observations in the missing sample matrix. The generator outputs a generated data matrix with the exact same size as the input.
[0125] S2: Concatenate the cue matrix and historical complete sensor data into positive samples, and concatenate the generated data matrix and cue matrix into negative samples. Train the discriminator based on binary cross-entropy loss to obtain the updated discriminator.
[0126] Specifically, a positive sample is a composite matrix formed by splicing a cue matrix and historical complete sensor data along a certain dimension, representing a combination of real complete data and cue information.
[0127] Negative samples are composite matrices formed by concatenating the generated data matrix and the cue matrix, representing the combination of generated data and cue information.
[0128] Binary cross-entropy loss is a loss function commonly used in binary classification tasks, which measures the difference between the discriminator's output probability and the true label.
[0129] The updated discriminator is a discriminator that has undergone one gradient descent parameter update, and its classification ability is improved compared to the original discriminator.
[0130] The terminal constructs two sets of samples for training the discriminator. The cue matrix and historical complete sensor data are concatenated along a certain dimension to form a composite matrix containing cue information and real complete data, serving as the positive sample. The generated data matrix and cue matrix are concatenated along the same dimension to form a composite matrix containing cue information and generator output data, serving as the negative sample. The terminal inputs the positive and negative samples sequentially into the discriminator. The discriminator outputs a scalar probability value for each input sample, representing the likelihood that the sample belongs to the real data. Ideally, the output should be close to 1 for positive samples and close to 0 for negative samples. The terminal calculates the binary cross-entropy loss between the discriminator's output and the real labels, and updates the discriminator's network parameters using a backpropagation algorithm to enhance its classification ability.
[0131] S3: Input the generated data matrix and the prompt matrix into the updated discriminator to predict the original observations, obtain the prediction probability, and train the generator based on the prediction probability with the goal of minimizing the generator loss value, thus obtaining the updated generator.
[0132] Specifically, the original observation prediction involves concatenating the generated data matrix and the cue matrix and inputting them into the updated discriminator. The discriminator then outputs a probability value, representing its confidence level that the input is real data.
[0133] The predicted probability is the probability value output by the discriminator, used to calculate the generator loss.
[0134] The generator loss value is a loss function that measures the performance of the generator. It is usually designed to make the discriminator output a high probability for the generated data, so that the generated data is difficult for the discriminator to recognize.
[0135] The terminal concatenates the generated data matrix and the prompt matrix along the same dimension and inputs them into the newly updated discriminator. The discriminator judges this input and outputs a predicted probability value, representing its confidence that the input is real data. The terminal calculates the generator's loss based on this predicted probability. The training objective of the generator is to make it difficult for the discriminator to distinguish between generated and real data, i.e., to have the discriminator output the highest possible probability for generated data. Therefore, the generator loss is usually designed to be based on the negative log-likelihood of the discriminator's output or a similar form, causing the generator parameters to be updated in the direction that makes the discriminator err. The terminal updates the generator's network parameters using the backpropagation algorithm, aiming to minimize the generator loss value. The updated generator is denoted as the updated generator.
[0136] After completing S3, one iteration ends. The terminal repeats S1, S2, and S3, using the latest generator and discriminator each time, until the loss values of the generator and discriminator stabilize. At this point, the trained generator and discriminator together form a generative adversarial incomplete network, which can be used to fill in missing values in the original sensor data matrix in real-world scenarios.
[0137] In one embodiment, based on the predicted probability, the generator is trained with the goal of minimizing the generator loss value to obtain an updated generator, including:
[0138] The formula for calculating the generator loss value is:
[0139] ;
[0140] in, Let be the generator loss function. To help the generator combat loss, For the trade-off function, Reconstruct the loss for the generator.
[0141] Specifically, the generator loss function is the overall objective function that the generator needs to minimize during training. The generator loss function comprehensively reflects the generator's performance in the current state; a lower value indicates better generator performance. During training, the generator's network parameters are continuously adjusted using the backpropagation algorithm to keep this loss value decreasing.
[0142] The generator adversarial loss measures the generator's ability to deceive the discriminator. Its calculation depends on the discriminator's predicted probability of the generated data output. The generator aims to make the discriminator misclassify the generated data as real data; therefore, the adversarial loss is typically designed to make the discriminator output a component with a high probability. Specifically, the adversarial loss value approaches its minimum when the discriminator's probability of outputting the generated data is close to 1. This term drives the generator to produce imputation data whose distribution characteristics are as close as possible to the real, complete sensor data, thus possessing high realism and plausibility.
[0143] The tradeoff function is a pre-defined positive weighting coefficient used to adjust the relative importance of adversarial loss and reconstruction loss in the total loss. The value of this coefficient determines whether the model prioritizes the realism of the generated data or the consistency between the generated data and the original observations during training. By adjusting the tradeoff function, the contribution ratio of the two loss terms can be controlled, preventing one from dominating while the other is neglected.
[0144] The generator reconstruction loss is a loss term that measures the difference between the generator's output at known observation locations in the original sensor data matrix and the true observations. It is obtained by calculating the error between the generated data matrix and the true complete data at known locations. This term constrains the generator to accurately retain the existing true observations in the original data while performing missing value imputation, preventing the generator from tampering with data at known locations to deceive the discriminator, thus ensuring the reliability and physical consistency of the imputation results.
[0145] This dual constraint forces the generator to use adversarial loss during training, ensuring that the generated imputed data approximates the real data in terms of overall distribution and possesses statistical reasonableness. Conversely, the reconstruction loss ensures that the real data is strictly reproduced at the known locations of the observations, possessing physical accuracy. This dual constraint avoids deviations that might occur with a single loss function.
[0146] In one embodiment, based on the sensor data matrix, an unsupervised clustering algorithm is used to perform cluster analysis to obtain a rock mass state classification map, including:
[0147] Step 501: Traverse the sensor data matrix, extract the spatial coordinates, radial stress values, and axial strain values of each measuring point, and stitch them together to obtain the feature dataset.
[0148] Among them, the measuring point is the specific location where the sensor is placed on the anchor bolt, and each measuring point corresponds to a set of spatial coordinates and sensor readings.
[0149] Spatial coordinates are three-dimensional coordinates that describe the position of a measuring point on a tunnel cross-section or anchor bolt, and are used for subsequent spatial positioning and visualization.
[0150] The radial stress value is the contact stress perpendicular to the anchor axis, measured by a micro earth pressure cell.
[0151] The axial strain value is the strain along the anchor bolt axis measured by a fiber optic strain sensor.
[0152] The feature dataset is a multidimensional data set formed by concatenating the spatial coordinates, radial stress values, and axial strain values of each measuring point. Each row represents a measuring point, and each column represents a feature dimension.
[0153] Step 502: Based on the feature dataset, the cluster centers are initialized using the K-Means++ algorithm to obtain the initial cluster centers.
[0154] Specifically, the K-Means++ algorithm is an improved K-Means clustering initialization method that improves the stability and quality of clustering results by making the initial cluster centers as dispersed as possible.
[0155] Cluster center initialization is the process of determining the initial cluster center locations before running the formal clustering iteration.
[0156] The initial cluster centers are the first batch of cluster centers selected by the K-Means++ algorithm, and each center point corresponds to a data point in the feature dataset.
[0157] The terminal randomly selects a data point from the feature dataset as the first initial cluster center. For each remaining data point in the feature dataset, the terminal calculates the square of the minimum Euclidean distance from the remaining data point to all currently selected cluster centers. The terminal also calculates the probability that each data point will be selected as the next cluster center, which is equal to the square of that point divided by the sum of the squares of all data points. The terminal generates a random number uniformly distributed between 0 and 1, then iterates through all data points, accumulating the probabilities. When the accumulated probability is greater than or equal to the random number for the first time, the currently traversed data point is selected as the next initial cluster center. This roulette wheel selection process is repeated until a preset number of initial cluster centers are selected.
[0158] Step 503: Based on the Euclidean distance from each measurement point in the feature dataset to each initial cluster center, assign each measurement point to each initial cluster center to obtain a list of category labels.
[0159] Specifically, Euclidean distance is the straight-line distance between two multidimensional vectors, used to measure the similarity between data points and cluster centers.
[0160] The category label list is a one-dimensional array with a length equal to the total number of test points. Each element records the cluster center number to which the test point is assigned.
[0161] The terminal iterates through every data point in the feature dataset. For the current data point, the terminal calculates the Euclidean distance between the current data point and each initial cluster center, which is calculated as the square root of the sum of the squares of the differences between the components of the two vectors. The terminal finds the minimum distance among all distances and assigns the data point to the initial cluster center that produced the minimum distance, recording the cluster center's number. After iterating through all data points, the terminal obtains a list of category labels, where each data point is assigned to the cluster containing the nearest initial cluster center. At this point, each cluster contains a group of data points with similar features.
[0162] Step 504: Based on the list of category labels, update the cluster centers of the feature dataset to obtain the updated cluster centers, and perform a convergence check based on the Euclidean distance between the updated cluster centers and the initial cluster centers to obtain the convergence flag.
[0163] Cluster center update is the process of recalculating the center position of a cluster based on the mean of all data points in each cluster.
[0164] The updated cluster centers are the new cluster center coordinates obtained after an update, with each cluster corresponding to an updated center point.
[0165] Convergence checking is the process of comparing the offset between the updated cluster centers and the previous cluster centers to determine whether the clustering has reached a stable state.
[0166] The convergence flag is a Boolean value or a flag indicating whether the clustering has converged. If the sum of the Euclidean distances between the updated cluster centers and the previous cluster centers is less than a preset threshold, the convergence flag is true; otherwise, it is false.
[0167] The terminal groups the data points in the feature dataset into clusters based on the category label list. For each cluster, the terminal calculates the average value of all data points within that cluster across each feature dimension, obtaining a new coordinate vector, which serves as the updated cluster center for that cluster. The updated cluster centers of all clusters constitute the updated cluster center set. The terminal performs a convergence check, calculating the Euclidean distance between each updated cluster center and its corresponding previous cluster center, and summing the distances of all clusters. If this sum is less than a preset convergence threshold, the clustering is considered converged, and the convergence flag is set to true; otherwise, the convergence flag is set to false, and the terminal uses the updated cluster centers as new cluster centers, reallocates data points, and updates the centers again until the convergence flag is true.
[0168] Step 505: Based on the convergence flag, classify the rock mass state of the updated cluster centers to obtain classification labels, and based on the spatial coordinates, assign each measuring point to the classification label to obtain the rock mass state classification map.
[0169] Among them, rock mass state classification is the process of mapping each updated cluster center to one of three rock mass states based on the characteristics represented by it.
[0170] The classification label is the name of the rock mass state category assigned to each updated cluster center, including high-risk weak zone, stable transition zone, and high-stress bearing zone.
[0171] A rock mass condition classification map is a visualization drawn on a tunnel cross-section or anchor bolt distribution map by combining the classification label of each measuring point with its spatial coordinates. It intuitively displays the spatial distribution of different rock mass conditions.
[0172] After the terminal confirms the convergence flag is true, it obtains the final updated cluster centers. For each updated cluster center, the terminal examines its corresponding radial stress and axial strain values. Based on the relative magnitudes of these two values, the rock mass condition is classified: if the radial stress is small and the axial strain is large, the cluster corresponding to that center is marked as a high-risk weak zone; if the radial stress and axial strain are moderate, it is marked as a stable transition zone; if the radial stress is large and the axial strain is small, it is marked as a high-stress bearing zone. Each updated cluster center receives a classification label. The terminal assigns each measuring point to the classification label of its respective cluster according to the category label list. The terminal reads the spatial coordinates of each measuring point, attaches the classification label to the corresponding spatial location, and generates a rock mass condition classification map. The rock mass condition classification map distinguishes different rock mass condition regions by color or pattern, clearly showing the spatial distribution pattern of high-risk weak zones, stable transition zones, and high-stress bearing zones.
[0173] In one embodiment, the grouting effect is evaluated based on post-grouting sensing data, and the evaluation conclusions are obtained, including:
[0174] Step 601: Extract the radial stress values of each earth pressure cell measuring point from the post-grouting sensor data, and calculate the standard deviation of the radial stress values to obtain the pressure uniformity value.
[0175] Among them, the post-grouting sensing data is a sensing data matrix that is re-collected and fed back by the sensor array after the anchor bolt has completed the grouting action. It includes the latest radial stress value and axial strain value of each measuring point.
[0176] Earth pressure cell measuring points are specific locations on anchor bolts where miniature earth pressure cells are installed. These sensors are used to measure contact stress in a direction perpendicular to the anchor bolt axis.
[0177] The radial stress value is the contact stress value perpendicular to the anchor bolt axis measured by the earth pressure cell, reflecting the degree of compression of the anchor bolt by the surrounding rock.
[0178] Standard deviation is a statistical measure of the dispersion of a set of values. It is calculated as the square root of the sum of the squares of the differences between each value and the mean. A smaller standard deviation indicates that the data is more concentrated; a larger standard deviation indicates that the data is more dispersed.
[0179] The pressure uniformity value is a single numerical value obtained by calculating the standard deviation of the radial stress values at all earth pressure cell measuring points. It is used to quantitatively describe the uniformity of the pressure exerted by the surrounding rock on the anchor bolts after grouting. The smaller the value, the more uniform the pressure distribution and the better the grouting filling effect.
[0180] The terminal filters the radial stress values corresponding to all earth pressure cell measuring points from the post-grouting sensor data, forming a list of radial stress values. The terminal calculates the arithmetic mean of this list. The terminal iterates through each radial stress value in the list, calculates the difference between it and the mean, and squares the difference. The sum of all squared values is divided by the total number of measuring points to obtain the variance. The square root of the variance is then taken to obtain the standard deviation. This standard deviation is the pressure equalization value.
[0181] Step 602: Extract the axial strain values of each fiber optic strain sensor measurement point from the post-grouting sensing data, and perform linear regression calculation based on the axial strain values to obtain the strain stability rate value.
[0182] Specifically, the fiber optic strain sensor measuring point is a specific location on the anchor bolt where a fiber optic strain sensor is installed. This sensor is used to measure the strain along the anchor bolt axis.
[0183] Axial strain is the strain value measured by a fiber optic strain sensor along the axis of the anchor bolt, reflecting the tensile or compressive deformation of the anchor bolt.
[0184] Linear regression is a statistical analysis method used to establish a linear relationship model between two variables. Here, time is used as the independent variable and axial strain is used as the dependent variable to fit a straight line, and the slope of the line represents the rate of change of strain with time.
[0185] The strain stability rate is the absolute value of the slope of the regression line obtained through linear regression. It is used to quantitatively describe the stability of the axial deformation of the anchor bolt over time after grouting. The closer this value is to zero, the more stable the deformation tends to be.
[0186] The terminal extracts the axial strain value sequence of all fiber optic strain sensor measuring points over a period of time after grouting from the post-grouting sensing data. For each measuring point, the terminal fits a straight line using the least squares method with time as the independent variable and axial strain value as the dependent variable, and calculates the slope of the line. The slope represents the change in axial strain per unit time. The terminal averages the slopes of all measuring points to obtain the strain stability rate. The closer this value is to zero, the smaller the axial deformation rate of the anchor bolt after grouting, and the more effectively the surrounding rock convergence deformation is controlled.
[0187] Step 603: Compare the pressure equalization value and strain stability rate value with the preset stability thresholds to obtain the evaluation conclusion.
[0188] Specifically, the preset stability thresholds are two critical values set based on engineering experience and safety standards, corresponding to the qualified standards for pressure equalization and strain stability rate, respectively. For example, the pressure equalization threshold can be set to 50% of the initial value, and the strain stability rate threshold can be set to 1 microstrain per minute.
[0189] The assessment conclusion is the final judgment obtained by comparing the pressure isostatic ratio and strain stability rate values with their respective thresholds. The assessment conclusion includes whether grouting was successful or whether another round of grouting is needed. Successful grouting indicates that both indicators have reached or exceeded the threshold, and the surrounding rock condition has met the stability requirements; the need for another round of grouting indicates that at least one indicator has not reached the threshold, and a new round of perception, analysis, decision-making, and execution needs to be initiated.
[0190] The terminal compares the pressure equalization value with the pressure equalization threshold. If the pressure equalization value is less than or equal to the threshold, the pressure equalization is considered to be up to standard; otherwise, it is considered to be down to standard. The terminal compares the absolute value of the strain stability rate with the strain stability rate threshold. If the absolute value is less than or equal to the threshold, the strain stability rate is considered to be up to standard; otherwise, it is considered to be down to standard. Only when both indicators meet the standards is the terminal determined that the grouting effect is qualified, the evaluation conclusion is that the grouting is successful, and the current round of operation ends or switches to long-term monitoring mode. If any indicator fails to meet the standards, the terminal determines that the grouting effect is insufficient, the evaluation conclusion is that the next round of grouting is required, and the terminal automatically uses the post-grouting sensor data as new input, jumps to the data preprocessing and cluster analysis stage, and re-executes the complete closed-loop control process.
[0191] In one embodiment, based on the rock mass condition classification map and the sensor data matrix, the priority score of each measuring point is calculated, including:
[0192] The formula for calculating priority scores is:
[0193] ;
[0194] in, For the first The first anchor bolt Priority score for each measurement point For the first The first anchor bolt Normalized fractions of radial stress values at each measuring point For the first The first anchor bolt Normalized fractions of axial strain values at each measuring point For the first The first anchor bolt The normalized score of the risk clustering results for each measurement point. , and These are the weighting coefficients.
[0195] The system simultaneously considers the stress state, deformation state, and rock mass risk level of the measuring points, avoiding the one-sidedness of judging based on a single indicator. Through weighted summation, three indicators with different physical meanings are integrated into a unified quantitative score, making the decision-making basis more comprehensive and scientific. All measuring points are sorted from highest to lowest priority score to generate a grouting sequence. The measuring points with the highest scores are prioritized for grouting, thus achieving precise execution and concentrating limited grouting resources on the areas most in need of reinforcement, improving grouting efficiency and support effectiveness.
[0196] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0197] Based on the same inventive concept, this application also provides an intelligent grouting sensing anchor bolt control system for layered soft rock, which implements the aforementioned intelligent grouting sensing anchor bolt control method for layered soft rock. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the intelligent grouting sensing anchor bolt control system for layered soft rock provided below can be found in the limitations of the intelligent grouting sensing anchor bolt control method for layered soft rock described above, and will not be repeated here.
[0198] In one exemplary embodiment, such as Figure 2 As shown, an intelligent grouting sensing anchor control system 800 suitable for layered soft rock is provided, comprising:
[0199] The filling module 801 is used to obtain the radial stress value and axial strain value of each measuring point on the anchor rod, obtain the original sensing data matrix, and fill the missing value of the original sensing data matrix based on the generative adversarial filling network to obtain the sensing data matrix.
[0200] Clustering module 802 is used to perform cluster analysis based on the sensor data matrix using an unsupervised clustering algorithm to obtain a rock mass state classification map; the rock mass state classification map includes high-risk weak areas, stable transition areas, and high-stress bearing areas;
[0201] The decision module 803 is used to calculate the priority score of each measuring point based on the rock mass condition classification map and the sensor data matrix, and to make a grouting priority decision based on the priority score to obtain the grouting priority control command; wherein, the grouting priority control command is used to instruct the anchor to perform the grouting action and to feed back the sensor data after grouting.
[0202] The evaluation module 804 is used to evaluate the grouting effect based on the sensor data after grouting and obtain the evaluation conclusion; the evaluation conclusion is used to indicate the next round of grouting.
[0203] Furthermore, the filling module 801 is also used for:
[0204] Traverse the original sensor data matrix and mark the mask matrix positions corresponding to elements with empty values or values exceeding the preset physical range as to be filled, thus obtaining the missing value mask matrix.
[0205] The original sensor data matrix and the missing value mask matrix are input into the generator of the generative adversarial imputation network to generate missing values and obtain the adversarial generative data matrix.
[0206] Based on the missing value mask matrix, the missing location features are extracted from the original sensor data matrix to obtain the missing features;
[0207] The missing features are input into the random forest model to predict the missing values, and the random forest imputation data matrix is obtained.
[0208] Based on the missing value mask matrix, the original sensor data matrix, the adversarial generative data matrix, and the random forest imputation data matrix are weighted and summed to obtain the sensor data matrix.
[0209] Furthermore, the intelligent grouting sensing anchor control system 800, suitable for layered soft rock, also includes a training module for:
[0210] Based on historical complete sensor data, a training missing mask matrix is randomly generated, and based on the training missing mask matrix, the historical complete sensor data is replaced with the true values to obtain the missing sample matrix.
[0211] Based on the missing sample matrix and the training missing mask matrix, a cue matrix is constructed. Then, based on the cue matrix, the missing sample matrix, the training missing mask matrix, and historical complete sensor data, the generator and discriminator are iteratively trained through iterative training steps until the loss value converges, thus obtaining the generative adversarial incomplete network.
[0212] The iterative training steps include:
[0213] S1: Input the missing sample matrix and the training missing mask matrix into the generator to generate the imputation data and obtain the generated data matrix;
[0214] S2: Concatenate the cue matrix and historical complete sensor data into positive samples, and concatenate the generated data matrix and cue matrix into negative samples. Train the discriminator based on binary cross-entropy loss to obtain the updated discriminator.
[0215] S3: Input the generated data matrix and the prompt matrix into the updated discriminator to predict the original observations, obtain the prediction probability, and train the generator based on the prediction probability with the goal of minimizing the generator loss value, thus obtaining the updated generator.
[0216] The formula for calculating the generator loss value is as follows:
[0217] ;
[0218] in, Let be the generator loss function. To help the generator combat loss, For the trade-off function, Reconstruct the loss for the generator.
[0219] Furthermore, clustering module 802 is also used for:
[0220] Traverse the sensor data matrix, extract the spatial coordinates, radial stress values and axial strain values of each measuring point, and stitch them together to obtain the feature dataset;
[0221] Based on the feature dataset, the K-Means++ algorithm is used to initialize the cluster centers, resulting in initial cluster centers.
[0222] Based on the Euclidean distance from each measurement point in the feature dataset to each initial cluster center, each measurement point is assigned to each initial cluster center to obtain a list of category labels;
[0223] Based on the list of category labels, the cluster centers of the feature dataset are updated to obtain the updated cluster centers. Then, based on the Euclidean distance between the updated cluster centers and the initial cluster centers, a convergence check is performed to obtain the convergence flag.
[0224] Based on the convergence criteria, the rock mass state is classified according to the updated cluster centers to obtain classification labels. Based on the spatial coordinates, each measuring point is assigned to a classification label to obtain a rock mass state classification map.
[0225] Furthermore, the evaluation module 804 is also used for:
[0226] The radial stress values of each earth pressure cell measuring point are extracted from the sensor data after grouting, and the standard deviation of the radial stress values is calculated to obtain the pressure uniformity value.
[0227] The axial strain values of each fiber optic strain sensor measurement point are extracted from the grouting post-sensing data, and linear regression calculation is performed based on the axial strain values to obtain the strain stability rate value.
[0228] The pressure equalization value and strain stability rate value are compared with the preset stability threshold to obtain the evaluation conclusion.
[0229] The formula for calculating the priority score is as follows:
[0230] ;
[0231] in, For the first The first anchor bolt Priority score for each measurement point For the first The first anchor bolt Normalized fractions of radial stress values at each measuring point For the first The first anchor bolt Normalized fractions of axial strain values at each measuring point For the first The first anchor bolt The normalized score of the risk clustering results for each measurement point. , and These are the weighting coefficients.
[0232] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the intelligent grouting sensing anchor control method for layered soft rock as described above.
[0233] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.
[0234] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0235] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.
Claims
1. A method for controlling intelligent grouting sensing anchor bolts suitable for layered soft rock, characterized in that, The method includes: The radial stress and axial strain values at each measuring point on the anchor bolt are obtained to obtain the original sensing data matrix. Based on the generative adversarial filling network, the missing values of the original sensing data matrix are filled to obtain the sensing data matrix. Based on the sensor data matrix, cluster analysis is performed using an unsupervised clustering algorithm to obtain a rock mass state classification map; wherein, the rock mass state classification map includes high-risk weak areas, stable transition areas, and high-stress bearing areas; Based on the rock mass condition classification map and the sensor data matrix, the priority score of each measuring point is calculated, and based on the priority score, the grouting priority decision is made to obtain the grouting priority control command; wherein, the grouting priority control command is used to instruct the anchor to perform the grouting action and to feed back the sensor data after grouting. Based on the grouting post-injection sensing data, the grouting effect is evaluated, and an evaluation conclusion is obtained; wherein, the evaluation conclusion is used to indicate the next round of grouting.
2. The method according to claim 1, characterized in that, The process of imputing missing values in the original sensor data matrix using a generative adversarial imputation network to obtain a sensor data matrix includes: Traverse the original sensor data matrix and mark the mask matrix positions corresponding to elements with empty values or values exceeding a preset physical range as to be filled, to obtain the missing value mask matrix; The original sensor data matrix and the missing value mask matrix are input into the generator of the generative adversarial imputation network to generate missing values and obtain the adversarial generative data matrix. Based on the missing value mask matrix, the missing location features are extracted from the original sensor data matrix to obtain the missing features; The missing features are input into the random forest model to predict missing values, and the random forest imputation data matrix is obtained. Based on the missing value mask matrix, the original sensor data matrix, the adversarial generative data matrix, and the random forest imputation data matrix are weighted and summed to obtain the sensor data matrix.
3. The method according to claim 2, characterized in that, The generative adversarial filling network is trained using the following method: Based on historical complete sensor data, a training missing mask matrix is randomly generated, and based on the training missing mask matrix, the historical complete sensor data is replaced with real values to obtain a missing sample matrix. Based on the missing sample matrix and the training missing mask matrix, a cue matrix is constructed. Based on the cue matrix, the missing sample matrix, the training missing mask matrix, and the historical complete sensor data, the generator and discriminator are iteratively trained through iterative training steps until the loss value converges, thus obtaining the generative adversarial incomplete network. The iterative training steps include: S1: Input the missing sample matrix and the training missing mask matrix into the generator to generate imputation data and obtain the generated data matrix; S2: Concatenate the cue matrix and the historical complete sensor data into positive samples, and concatenate the generated data matrix and the cue matrix into negative samples. Train the discriminator based on the binary cross-entropy loss to obtain the updated discriminator. S3: Input the generated data matrix and the prompt matrix into the updated discriminator to predict the original observations, obtain the prediction probability, and train the generator based on the prediction probability with the goal of minimizing the generator loss value, to obtain the updated generator.
4. The method according to claim 3, characterized in that, The step of training the generator based on the predicted probability, with the objective of minimizing the generator loss value, to obtain the updated generator includes: The formula for calculating the generator loss value is as follows: ; in, Let be the generator loss function. To help the generator combat loss, For the trade-off function, Reconstruct the loss for the generator.
5. The method according to claim 1, characterized in that, The process of performing cluster analysis using an unsupervised clustering algorithm based on the sensor data matrix to obtain a rock mass state classification map includes: Traverse the sensor data matrix, extract the spatial coordinates of each measuring point, the radial stress value and the axial strain value, and stitch them together to obtain a feature dataset; Based on the aforementioned feature dataset, cluster centers are initialized using the K-Means++ algorithm to obtain initial cluster centers. Based on the Euclidean distance from each measurement point in the feature dataset to each initial cluster center, each measurement point is assigned to each initial cluster center to obtain a list of category labels; Based on the category label list, the cluster centers of the feature dataset are updated to obtain updated cluster centers. Then, based on the Euclidean distance between the updated cluster centers and the initial cluster centers, a convergence check is performed to obtain a convergence flag. Based on the convergence flag, the updated cluster centers are classified into rock mass states to obtain classification labels. Based on the spatial coordinates, each measuring point is assigned to the classification labels to obtain the rock mass state classification map.
6. The method according to claim 5, characterized in that, The evaluation of the grouting effect based on the post-grouting sensing data, and the resulting evaluation conclusions, include: The radial stress values of each earth pressure cell measuring point are extracted from the grouting post-sensing data, and the standard deviation of the radial stress values is calculated to obtain the pressure uniformity value. The axial strain values of each fiber optic strain sensor measurement point are extracted from the grouting post-sensing data, and linear regression calculation is performed based on the axial strain values to obtain the strain stability rate value. The pressure equalization value and the strain stability rate value are compared with preset stability thresholds to obtain the evaluation conclusion.
7. The method according to claim 1, characterized in that, The priority score for each measuring point is calculated based on the rock mass condition classification map and the sensor data matrix, including: The formula for calculating the priority score is: ; in, For the first The first anchor bolt Priority score for each measurement point For the first The first anchor bolt Normalized fractions of radial stress values at each measuring point For the first The first anchor bolt Normalized fractions of axial strain values at each measuring point For the first The first anchor bolt The normalized score of the risk clustering results for each measurement point. , and These are the weighting coefficients.
8. A smart grouting sensing anchor control system suitable for layered soft rock, characterized in that, The system includes: The filling module is used to obtain the radial stress value and axial strain value of each measuring point on the anchor rod to obtain the original sensing data matrix, and fill the missing value of the original sensing data matrix based on the generative adversarial filling network to obtain the sensing data matrix. The clustering module is used to perform clustering analysis based on the sensor data matrix using an unsupervised clustering algorithm to obtain a rock mass state classification map; wherein, the rock mass state classification map includes high-risk weak areas, stable transition areas, and high-stress bearing areas; The decision module is used to calculate the priority score of each measuring point based on the rock mass condition classification map and the sensing data matrix, and to make a grouting priority decision based on the priority score to obtain a grouting priority control command; wherein, the grouting priority control command is used to instruct the anchor to perform grouting action and to feed back the sensing data after grouting. The evaluation module is used to evaluate the grouting effect based on the post-grouting sensing data and obtain an evaluation conclusion; wherein the evaluation conclusion is used to indicate the next round of grouting.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.
