A Method for On-site Monitoring of Multi-source Information in Tunnels Based on IoT Wireless Transmission

Abstract

This invention discloses a method for on-site monitoring of multi-dimensional tunnel information based on IoT wireless transmission, relating to the technical field of tunnel monitoring and tunnel safety assessment. Specifically, it addresses the problem that traditional monitoring methods are insufficient to meet the requirements for comprehensive, real-time, and accurate tunnel monitoring. The invention employs a combination of IoT monitoring and on-site monitoring, using a deep learning model to identify tunnel lining cracks, water leakage, lining voids, and road cracks. Linear regression is used to predict the safety level vector of the tunnel's structural health, thereby determining the overall structural safety level of the tunnel. Based on the data analysis results, the monitoring and sleep modes of the sensors are controlled, achieving intelligent tunnel monitoring. Relevant monitoring and prediction information is transmitted to relevant personnel to facilitate the development of monitoring and maintenance plans.

Description

Technical Field

[0001] This invention belongs to the technical field of tunnel monitoring and tunnel safety assessment, specifically relating to a method for on-site monitoring of multi-source tunnel information based on wireless transmission via the Internet of Things. Background Technology

[0002] With more and more highway tunnels going into operation, the maintenance costs are increasing day by day. Moreover, traditional monitoring methods are difficult to meet the requirements for comprehensive, real-time and accurate monitoring of tunnels, which leads to the inability to detect risks in tunnel operation in a timely manner, posing a huge hidden danger to the safety of tunnel operation.

[0003] Therefore, establishing a tunnel monitoring system based on IoT wireless transmission can reduce the manpower costs of tunnel monitoring, and is more convenient and faster than traditional monitoring methods, meeting the needs of modern society for intelligence. Furthermore, monitoring methods that integrate diverse tunnel information can provide a more comprehensive overall assessment of the tunnel's structural safety. Based on the collected data, the tunnel's safety level can be determined. When the safety level is too high, it indicates a risk of property damage. In this case, information can be sent to relevant personnel to facilitate timely and effective protective measures, thereby preventing potential tunnel disasters. This is of great significance for the maintenance and management of highway tunnels during their operational phase. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a method for on-site monitoring of multi-source tunnel information based on IoT wireless transmission, specifically comprising: S1. Collecting image data of monitoring items and uploading it to a local server; S2. Based on the image data of the monitoring items, training a vehicle identification model and a ground-penetrating radar identification model, evaluating the predictive performance of the models, and deploying the vehicle identification model and the ground-penetrating radar identification model to an edge server; S3. Using the vehicle identification model and the ground-penetrating radar identification model, identifying and classifying the image data uploaded from the local server to the edge server, and obtaining the safety level vector of each monitoring item through the classification results; S4. Based on the safety level vector of each monitoring item, establishing a linear regression model and training the model, and obtaining the overall tunnel safety level assessment vector based on the trained linear regression model; S5. Obtaining the overall tunnel safety level assessment vector based on the overall tunnel safety level assessment vector, and subsequently obtaining the overall tunnel structural safety level; S6. Determining the corresponding operations based on the tunnel safety level assessment vector and the tunnel structural safety level. The method of this invention controls the monitoring and sleep of sensors based on the results of data analysis, realizing intelligent monitoring of tunnels. When the safety risk coefficient is high, relevant information is transmitted to the staff so that they can formulate monitoring and maintenance plans.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for on-site monitoring of multi-source tunnel information based on IoT wireless transmission includes:

[0007] S1. Collect monitoring project image data and upload it to the local server;

[0008] S2. Based on the image data of the monitoring project, train the vehicle identification model and the ground penetrating radar identification model, evaluate the prediction effect of the model, and deploy the vehicle identification model and the ground penetrating radar identification model to the edge server;

[0009] S3. Using the vehicle-mounted data acquisition model and the ground-penetrating radar model, the image data uploaded from the local server to the edge server is identified and classified, and the safety level vector of each monitoring item is obtained through the classification results;

[0010] S4. Based on the safety level vector of each monitoring item, establish a linear regression model and train the model. Based on the trained linear regression model, obtain the overall tunnel safety level assessment vector.

[0011] S5. Based on the overall tunnel safety level assessment vector, the overall tunnel safety level assessment vector is obtained, and the overall structural safety level of the tunnel is obtained;

[0012] Preferably, the training methods for the vehicle identification model and the ground-penetrating radar identification model in step S2 are the same, except that the uploaded datasets used in the training process of the vehicle identification model and the ground-penetrating radar identification model correspond to the datasets obtained by the vehicle and the datasets obtained by the ground-penetrating radar, respectively.

[0013] Preferably, the process of training the data acquisition vehicle recognition model and the ground-penetrating radar recognition model in step S2 includes:

[0014] Step 1. Preprocess the images in the dataset uploaded from the local server;

[0015] Step 2. Take 70% of the preprocessed dataset as the training set, 20% as the validation set, and 10% as the test set;

[0016] Step 3. Build a vehicle identification model and a ground-penetrating radar identification model using the training set, and train them.

[0017] Preferably, the process of training the data acquisition vehicle recognition model and the ground-penetrating radar recognition model in step 3 includes:

[0018] 1) Calculate the ratio of the prior box size to the original image size in the training set in a single multi-boundary box detector; the formula is:

[0019]

[0020] in, This is the ratio of the size of the prior bounding box corresponding to the k-th feature map to the size of the original image. , These are the maximum and minimum values ​​for that ratio, generally. , m represents the number of feature maps;

[0021] 2) Calculate the width of the prior box in a single-shot multi-boundary box detector. and high The calculation formula is:

[0022]

[0023]

[0024] in, ;

[0025] 3) Calculate the mapping relationship between the prior bounding boxes on the feature map and the image coordinates of the original images in the training set. The calculation formula is as follows:

[0026]

[0027]

[0028]

[0029]

[0030] in, , The width and height of the feature layer; , The width and height of the original image; , , , ) is the feature map of the k-th layer centered at (a, b) and of size ( , The prior bounding box is mapped to the coordinates of the original image;

[0031] 4) The location loss function obtained based on the distance difference between the predicted bounding box and the labeled bounding box is as follows:

[0032]

[0033]

[0034] in, For the positive sample set, Indicates the first Does the first prior box match the second...? If there are label boxes with category P, This indicates a successful match; otherwise, it is 0. Represents the prediction window center coordinates and width and height Compared to the real frame The offset, where N is the number of candidate region boxes that match the true target box;

[0035] 5). Based on whether the predicted category matches the labeled category, the confidence loss function is obtained as follows:

[0036]

[0037]

[0038] in, Indicates the first Softmax confidence regression of positive samples with respect to class P; The set of negative samples; For the first Confidence level of a negative sample Let be the confidence level of the i-th positive sample;

[0039] 6) Based on the confidence loss function and location loss function described above, the total loss function is obtained as follows:

[0040]

[0041] in, For confidence loss, For location loss, N is the number of candidate region boxes that match the ground truth bounding boxes; l is the predicted bounding box. For annotation boxes; This is the weighting parameter, which is generally set to 1;

[0042] 7) Based on the total loss function, the corresponding data acquisition vehicle identification model and ground-penetrating radar identification model are obtained.

[0043] Preferably, the process of evaluating the model's prediction performance includes:

[0044] Step 1. Use the trained vehicle recognition model and ground-penetrating radar recognition model to detect images in the corresponding test set to obtain TP, FP, and FN. Then, based on TP, FP, and FN, calculate the accuracy and recall of the computational model to predict relevant targets.

[0045] Where TP is the number of correctly classified and correctly positioned bounding boxes; FP is the number of incorrectly classified or incorrectly positioned bounding boxes; FN is the unpredicted bounding boxes; Recall: the model's ability to predict all relevant targets; Precision: the model's ability to predict relevant targets, its formula is:

[0046]

[0047] Step 2. Based on the numerical values ​​of precision and recall, evaluate the predictive performance of the model using mAP. The evaluation process includes:

[0048] first,

[0049]

[0050] in, It is based on and The result is that, even when using As the y-axis The curve formed by the x-axis;

[0051] Secondly

[0052]

[0053] Where AP is the average precision and C is the number of categories in the object detection task. It is a natural number;

[0054] When the mAP value is greater than 75, the corresponding model has good predictive performance; when the mAP value is greater than 90, the corresponding model has excellent predictive performance.

[0055] Preferably, the safety level vector for each monitoring item in step S3 includes:

[0056] The crack safety level vector is:

[0057]

[0058] in, Let n be the vector of crack safety levels, where n is the number of segments where monitoring points are arranged throughout the tunnel.

[0059] The safety level vector for lining leakage is:

[0060]

[0061] in, Let n be the safety level vector for water leakage in the lining, where n is the number of segments where monitoring points are arranged throughout the tunnel.

[0062] The safety level vector for lining voids is:

[0063]

[0064] in, Let n be the safety level vector for the lining voids, where n is the number of monitoring points arranged throughout the tunnel.

[0065] The safety level vector of steel reinforcement stress is:

[0066]

[0067] in, Let n be the safety level vector for the reinforcing steel, and n be the number of monitoring points in the entire tunnel; the safety level vector for the concrete strain is:

[0068]

[0069] in, Let n be the concrete strain safety level vector, and n be the number of segments where the monitoring points are arranged throughout the tunnel.

[0070] The safety level vector for cross-sectional convergence is:

[0071]

[0072] in, Let n be the cross-sectional convergence safety level vector, where n is the number of segments where monitoring points are arranged throughout the tunnel.

[0073] Preferably, the process of obtaining the overall tunnel safety level assessment vector in step S4 includes:

[0074] Step 1. Establish a linear regression model based on the safety level vector of each monitoring item;

[0075] Step 2. Train the linear regression model;

[0076] Step 3. Obtain the tunnel safety level assessment vector based on the trained linear regression model;

[0077] Preferably, the linear regression model established in step 1 is as follows:

[0078]

[0079] in, It is the first The first set of tunnel monitoring data Predicted values ​​for the condition level assessment of the tunnel lining structure at the monitoring points. This refers to the weight of each monitoring item in the overall tunnel risk assessment. It is the first The first set of tunnel monitoring data The safety level of cracks at the monitoring points of the section These are bias parameters;

[0080] The process of training the linear regression model in step 2 is as follows:

[0081] (1) Initialize the parameters of the linear regression model , ;

[0082] (2) Use the loss function to continuously adjust the initialized parameters. , Training ends when the loss function converges.

[0083] The loss function used is the squared loss function, i.e.

[0084]

[0085] Where m is the number of tunnel data sets in the entire training sample, and n is the number of segments where monitoring points are deployed throughout the tunnel. It is the first The first set of tunnel monitoring data The target value for the assessment of the condition level of the tunnel lining structure at the monitoring points;

[0086] (3) Optimized parameters of the loss function , Optimized , for:

[0087]

[0088]

[0089] in, , The weights of the six monitoring items in the overall tunnel risk assessment; , That is the number of samples in each mini-batch. Indicates the learning rate; The values ​​of the six monitoring indicators for the i-th tunnel segment;

[0090] (4) Optimize the parameters , Substituting these values ​​into the linear regression model expression, we obtain the predicted value for the condition level assessment of the tunnel lining structure.

[0091] Step 3 yields the overall tunnel safety level assessment vector. The formula is:

[0092]

[0093] in, This is the predicted value for the structural condition assessment of the nth tunnel lining.

[0094] Preferably, in step S5, the formula for calculating the overall structural safety level of the tunnel is:

[0095]

[0096] A tunnel multi-source information field monitoring system based on Internet of Things wireless transmission includes:

[0097] The data acquisition module is used to collect image data from the monitoring project and upload it to the local server;

[0098] The model component module is used to train a data acquisition vehicle recognition model and a ground-penetrating radar recognition model based on the image data of the monitoring project, evaluate the prediction effect of the model, and deploy the data acquisition vehicle recognition model and the ground-penetrating radar recognition model to the edge server.

[0099] Vector acquisition module one is used to identify and classify image data uploaded from the local server to the edge server using the data acquisition vehicle identification model and the ground penetrating radar identification model, and obtain the safety level vector of each monitoring item through the classification results;

[0100] The second vector acquisition module is used to establish and train a linear regression model based on the safety level vectors of each monitoring item, and obtain the overall tunnel safety level assessment vector based on the trained linear regression model.

[0101] The safety level acquisition module is used to obtain the overall tunnel safety level assessment vector based on the overall tunnel safety level assessment vector, and then obtain the overall structural safety level of the tunnel.

[0102] The control module is used to determine the corresponding operation based on the tunnel safety level assessment vector and the tunnel structural safety level.

[0103] The beneficial effects of this invention are: This invention discloses a method for on-site monitoring of multi-source tunnel information based on wireless transmission via the Internet of Things. Compared with the prior art, the improvement of this invention lies in:

[0104] This invention provides a tunnel multi-source information on-site monitoring system and method based on IoT wireless transmission. Compared with traditional monitoring methods, it combines IoT monitoring with on-site monitoring, reducing the risk of obtaining erroneous information due to equipment malfunction. It uses convolutional neural networks to identify tunnel lining cracks, water leakage, lining voids, cross-sectional convergence, and road cracks, reducing labor costs. Deploying the network model to an edge server improves data transmission and computation speed. Linear regression is used to predict the safety level vector of tunnel structural health, thereby determining the overall structural safety level of the tunnel. Based on the data analysis results, the monitoring and sleep modes of sensors are controlled, achieving intelligent tunnel monitoring. When the safety risk coefficient is high, relevant information is transmitted to staff so that they can take timely measures to avoid personal and property losses. Attached Figure Description

[0105] Figure 1 This is a flowchart of the monitoring method of the present invention;

[0106] Figure 2 This is a schematic diagram of the monitoring method of the present invention;

[0107] Figure 3 This is a structural diagram of the single-shot multi-boundary box detector of the present invention; Detailed Implementation

[0108] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0109] Example 1:

[0110] See attached document Figure 1-3 As shown, the method for on-site monitoring of multi-source tunnel information based on IoT wireless transmission is characterized by including:

[0111] S1. Collect monitoring project image data and upload it to the local server;

[0112] The first step is to use six key indicators as monitoring items: steel reinforcement stress, concrete strain, cross-sectional convergence, lining voids, lining surface cracks, and water leakage.

[0113] Step 2: Selection and deployment of IoT monitoring equipment;

[0114] Among them, concrete strain is monitored using embedded and vibrating wire concrete strain gauges, steel stress is monitored using embedded and vibrating wire steel stress gauges, and cross-sectional convergence is monitored using laser phase laser rangefinders. The sensors used should be connected to high-voltage power lines around them to provide power to the sensors.

[0115] In each monitoring section, 10 steel stress gauges and 10 concrete strain gauges should be installed. The specific number and monitoring points should be arranged according to the actual needs of the site. The monitoring points should be able to cover all sensitive parts of the section so as to achieve comprehensive monitoring of the entire arch.

[0116] Among them, the laser rangefinders deployed at each monitoring section need to cover the transverse and longitudinal directions of the section. The specific layout should be designed according to the tunnel structure's net height to determine the number and location of transverse convergence monitoring lines.

[0117] A tunnel integrated data acquisition vehicle equipped with multiple devices was used to collect information from the tunnel site. The equipment on the tunnel integrated data acquisition vehicle included a line scan camera, an optical lens, a supplementary light, and a power supply. The tunnel images captured were mainly used to identify cracks on the road surface, cracks on the tunnel lining surface, and water leakage.

[0118] Among them, the line array camera and optical lens are used to capture images of the tunnel surface and tunnel lining. Because the lighting conditions in the tunnel are poor, supplementary lights are needed to provide a light source during shooting to ensure the clarity and uniform brightness of the images. The power supply is used to provide power to the camera and supplementary lights. Considering cost factors, the time interval for using the tunnel integrated data acquisition vehicle to collect tunnel site information should be set to once a year. Staff can also conduct fixed-point collection at an appropriate frequency according to the information sent in step S6.

[0119] Ground-penetrating radar (GPR) is used to collect information on the tunnel lining structure. A longitudinal survey line is laid out at each of the left and right arch waists, left and right inverted arches, and the arch crown. Inspectors use a GPR antenna close to the tunnel lining surface and perform continuous, uniform inspections. The tunnel lining structure information collected by GPR is primarily used to determine if there are voids in the tunnel lining structure. Considering cost factors, the time interval for collecting tunnel lining structure information using GPR should be set to once a year. Alternatively, staff can perform fixed-point collection at an appropriate frequency based on the information sent in step S6.

[0120] The third step involves the monitoring equipment wirelessly transmitting the collected data on steel reinforcement stress, concrete strain, and cross-sectional convergence to a local server. ZigBee wireless transmission is suitable because it is inexpensive, consumes little power, and has a long transmission distance. Considering that the amount of data collected by the steel reinforcement stress gauge, concrete strain gauge, and laser rangefinder is not large, ZigBee wireless transmission is appropriate. The local server should be located in a central control room around the tunnel for temporary storage of monitoring information.

[0121] Tunnel safety accidents often result from quantitative changes in internal factors leading to qualitative changes, or from sudden phenomena caused by external factors. Therefore, in daily use, it is not necessary for all monitoring equipment to be in operation at all times. When encountering external influences such as rain, snow, and earthquakes, the monitoring interval can be shortened, or even set to real-time monitoring.

[0122] The fourth step is for on-site personnel to store the collected image and radar information into a local server. The data collected by on-site personnel using ground-penetrating radar and tunnel integrated data acquisition vehicle is too large, and wireless transmission would consume too much network resources. After completing the data collection, on-site personnel can go directly to the central control room to store the data into the local server.

[0123] Step 5: Upload the newly received monitoring data from the local server to the edge server;

[0124] S2. Using deep learning methods and combining image data from the monitoring project, train a data collection vehicle identification model and a ground-penetrating radar identification model, evaluate the prediction performance of the model, and deploy the data collection vehicle identification model and the ground-penetrating radar identification model to an edge server;

[0125] Specifically, the training methods for the data collection vehicle recognition model and the ground-penetrating radar recognition model are the same. The difference is that during the training process, the data collection vehicle recognition model uses data acquired by the data collection vehicle, while the ground-penetrating radar recognition model uses data acquired by the ground-penetrating radar. All other steps are the same.

[0126] Because the tunnel integrated data acquisition vehicle takes too many images, manual processing is too inefficient. Therefore, we are considering using deep learning methods to train an identification model for the acquisition vehicle and a ground-penetrating radar identification model to automatically identify cracks on the road surface, cracks on the lining surface, and water seepage. Although the ground-penetrating radar collects electromagnetic wave signals, the data used to train the model is radar images. The data collected by concrete strain gauges, steel stress gauges, and laser rangefinders are already very intuitive and do not require further preprocessing.

[0127] The steps for training the data collection vehicle recognition model and the ground-penetrating radar recognition model using deep learning methods and combined with image data of the monitoring items are as follows:

[0128] Step 1. Preprocess the images in the dataset uploaded by the local server; here, if it is necessary to train the vehicle recognition model, preprocess the images in the dataset uploaded by the vehicle; if it is necessary to train the ground penetrating radar recognition model, preprocess the images in the dataset uploaded by the ground penetrating radar.

[0129] The data used to train the model is an existing dataset, not data collected directly on-site.

[0130] The preprocessing steps are as follows:

[0131] Step 1. Manual annotation: Manually annotate the image dataset used to train the model; the information to be annotated includes: whether there are cracks in the lining and whether there is water leakage in the lining.

[0132] Step 2. Image normalization; using the following formula, transform the labeled image pixels from 0 to 255 to 0 to 1;

[0133]

[0134] in, This represents the input image pixel values. To output image pixel values, , These represent the maximum and minimum values ​​of the input pixels, respectively.

[0135] Step 3. Image Enhancement: Rotating and mirroring the image obtained in Step 2 can expand the sample size, and adding noise to the image can improve the generalization of the model.

[0136] Step 2. Take 70% of the preprocessed image dataset as the training set, the remaining 20% ​​as the validation set, and 10% as the test set;

[0137] Step 3. Build a vehicle identification model and a ground-penetrating radar identification model using the training set, and train them;

[0138] 1) Using the Single Shot MultiBox Detector (SSD) model structure with progressively smaller convolutional layers, feature maps are extracted from the training set images. The ratio of the prior bounding box size corresponding to the feature map to the prior bounding box size of the training set images is then calculated. The formula is as follows:

[0139]

[0140] in, This is the ratio of the size of the prior bounding box corresponding to the k-th feature map to the size of the original image. , These are the maximum and minimum values ​​for that ratio, generally. , , where m represents any natural number;

[0141] 2). Calculate the width of the prior box in a single-shot multi-boundary box detector. and high The calculation formula is:

[0142]

[0143]

[0144] in, ;

[0145] 3) Calculate the mapping relationship between the prior bounding boxes on the feature map and the image coordinates of the original images in the training set. The calculation formula is as follows:

[0146]

[0147]

[0148]

[0149]

[0150] in, , For the width and height of the feature layer, , The width and height of the original image, ( , , , ) is the feature map of the k-th layer centered at (a, b) and of size ( , The prior bounding box is mapped to the coordinates of the original image;

[0151] 4). The location loss function can be obtained based on the distance difference between the predicted bounding box and the labeled bounding box, where the labeled bounding box is manually labeled, and the predicted bounding box is generated based on the image's pixel count and other content; the formula for the location loss function is as follows:

[0152]

[0153]

[0154] in, For the positive sample set, Indicates the first Does the first prior box match the second...? If there are label boxes with category P, This indicates a successful match; otherwise, it is 0. Represents the prediction window center coordinates and width and height Compared to the real frame The offset, where N is the number of candidate region boxes that match the true target box;

[0155] 5). Based on whether the predicted category matches the labeled category, the confidence loss function is obtained as follows:

[0156]

[0157]

[0158] in, Indicates the first Softmax confidence regression of positive samples with respect to class P; The set of negative samples; For the first Confidence level of a negative sample Let be the confidence level of the i-th positive sample;

[0159] 6). Based on the confidence loss function and location loss function described above, the total loss function is obtained as follows:

[0160]

[0161] in, For confidence loss, For location loss, N is the number of candidate region boxes that match the ground truth bounding boxes; l is the predicted bounding box. For annotation boxes; This is the weighting parameter, which is generally set to 1;

[0162] 7) Based on the total loss function, obtain the corresponding vehicle recognition model and ground-penetrating radar recognition model. Use this model to detect images in the test set to obtain TP, FP, and FN. Then, calculate the accuracy and recall of the model in predicting relevant targets based on TP, FP, and FN. TP is the number of correctly classified and correctly positioned bounding boxes; FP is the number of incorrectly classified or incorrectly positioned bounding boxes; FN is the unpredicted bounding boxes; Recall represents the model's ability to predict all relevant targets, i.e., the proportion of the model's predictions that correctly match the real targets; Precision represents the model's ability to predict relevant targets, i.e., the proportion of all predictions that match the real targets. The formula is:

[0163]

[0164] 8) Evaluate the model's predictive performance; that is, use mAP (Mean Average Precision) as the metric for evaluating model performance; when the mAP value is greater than 75, the model can be considered to have good predictive performance, and when the mAP value is greater than 90, the model can be considered to have excellent predictive performance; the formula for calculating mAP is:

[0165]

[0166] Where AP is the average precision; C is the number of categories in the target detection task;

[0167]

[0168] in, It is based on and The result is that, even when using As the y-axis The curve formed by the x-axis;

[0169] S3. Using the vehicle-mounted data acquisition model and the ground-penetrating radar model, the image data uploaded from the local server to the edge server is identified and classified, and the safety level vector of each monitoring item is obtained through the classification results;

[0170] After classification and recognition by the data collection vehicle recognition model, all images with cracks can be found in the tunnel images. Based on the sequence number of the crack images in the tunnel images, the specific location of the cracks in the tunnel can be inferred. Similarly, all images with lining leakage can be found in the tunnel images. Based on the sequence number of the lining leakage images in the tunnel images, the specific location of the lining leakage phenomenon in the tunnel can be inferred. At this point, the safety level assessment of lining cracks and highway cracks is summarized together.

[0171] Table 1 provides a reference for assessing the safety level of crack phenomena. As shown in the table, the assessment also requires consideration of crack length and width. Relevant personnel are needed to conduct the assessment, and should evaluate the identified and screened images, referring to Table 1 for the appropriate level assessment.

[0172] Relevant personnel conducted a safety level assessment of the lining cracks and highway cracks. The resulting crack safety level vector is as follows:

[0173]

[0174] Where n is the number of segments where monitoring points are arranged throughout the tunnel. This represents the crack safety level vector.

[0175] After the recognition and classification by the data acquisition vehicle's recognition model, all images with lining leakage can be found in the tunnel images. Based on the sequence number of the lining leakage image in the tunnel images, the specific location of the lining leakage phenomenon in the tunnel can be inferred.

[0176] Table 1 can be used as a reference for safety level assessment of lining leakage. As can be seen from the table, safety level assessment also needs to be combined with the surrounding environment and relevant personnel need to conduct the assessment. The personnel should first combine the surrounding weather to determine whether the lining leakage is seasonal, then determine whether there is mud flow on the sidewall of the leakage area, whether there is water gushing at the bottom of the tunnel, and finally determine whether the leakage in the arch directly hits the contact wire.

[0177] The safety level vector for lining leakage obtained by staff after conducting a safety level assessment of the lining leakage phenomenon is as follows:

[0178]

[0179] in, Let n be the safety level vector for water leakage in the lining, where n is the number of segments where monitoring points are arranged throughout the tunnel.

[0180] After identification by the ground-penetrating radar identification model, all radar images containing lining voids can be identified from the radar images. Based on the sequence number of the radar image containing the lining void in the entire radar image, the specific location of the lining void in the tunnel lining can be inferred.

[0181] For the safety level assessment of the lining void phenomenon, you can refer to Table 1. As can be seen from the table, when conducting the safety level assessment, it is also necessary to determine whether cracks and water leakage have appeared around the void. After obtaining the specific location of the lining void, you can query the crack location and lining leakage location obtained in step S3 to know whether cracks and water leakage have appeared around the lining void.

[0182] The safety level vector of the lining voids obtained after the safety level assessment of the lining void phenomenon is:

[0183]

[0184] in, Let n be the safety level vector for the lining voids, where n is the number of monitoring points arranged throughout the tunnel.

[0185] Referring to Table 1, the safety level vector of steel reinforcement stress is:

[0186]

[0187] in, Let n be the safety level vector for the reinforcing steel, and n be the number of segments where monitoring points are arranged throughout the tunnel.

[0188] Referring to Table 1, the safety level vector for concrete strain is:

[0189]

[0190] in, Let n be the concrete strain safety level vector, and n be the number of segments where the monitoring points are arranged throughout the tunnel.

[0191] Referring to Table 1, the safety level vector for cross-sectional convergence is:

[0192]

[0193] in, Let n be the cross-sectional convergence safety level vector, where n is the number of segments where monitoring points are arranged throughout the tunnel.

[0194] Table 1: Reference Table for Tunnel Safety Level Assessment

[0195]

[0196] Tunnel disasters are generally caused by quantitative changes leading to qualitative changes, so real-time monitoring of equipment is not necessary. In actual monitoring, when the safety level of the monitored steel reinforcement stress, concrete strain, or cross-sectional convergence is greater than level 2, it indicates that the tunnel structure has begun to deteriorate and monitoring of that section of the tunnel needs to be strengthened.

[0197] The specific methods for strengthening tunnel monitoring are as follows:

[0198] When the safety level of the reinforcing steel stress is level 3, that is, when the reinforcing steel stress... satisfy At that time, the steel stress gauge, concrete strain gauge and laser rangefinder at that location were set to collect data once every 1 hour;

[0199] When the safety level of the reinforcing steel stress is level 4, that is, when the reinforcing steel stress... satisfy At that time, the steel stress gauge, concrete strain gauge and laser rangefinder at that location were set to collect data every 30 minutes.

[0200] When the safety level of the reinforcing steel stress is level 5, that is, when the reinforcing steel stress... satisfy At that time, the steel stress gauge, concrete strain gauge and laser rangefinder at that location were set to collect data once every 1 minute;

[0201] S4. Based on the safety level vector of each monitoring item, establish a linear regression model and train the model. Based on the trained linear regression model, obtain the overall tunnel safety level assessment vector.

[0202] First, based on the safety level vectors of each monitoring item, a linear regression model is established. The expression of this model is as follows:

[0203]

[0204] in, It is the first The first set of tunnel monitoring data Predicted values ​​for the condition level assessment of the tunnel lining structure at the monitoring points. This refers to the weight of each monitoring item in the overall tunnel risk assessment. It is the first The first set of tunnel monitoring data The safety level of cracks at the monitoring points of the section These are bias parameters;

[0205] The process of training a linear regression model is as follows:

[0206] (1) Initialize the parameters of the linear regression model , ;

[0207] Specifically, a method of randomly initializing model parameters is used to adjust the weights in the linear regression model. Random sampling is performed from a normal distribution with a mean of 0 and a standard deviation of 0.01, and the bias parameter is... Initialize to 0;

[0208] (2) Define a loss function and use the loss function to continuously adjust the initialized parameters. , Training ends when the loss function converges.

[0209] The loss function used is the squared loss function, i.e.

[0210]

[0211] Where m is the number of tunnel data sets in the entire training sample, and n is the number of segments where monitoring points are deployed throughout the tunnel. It is the first The first set of tunnel monitoring data The target value for the assessment of the condition level of the tunnel lining structure at the monitoring points;

[0212] (3) Parameters of the loss function adjusted using mini-batch stochastic gradient descent , Optimized , for:

[0213]

[0214]

[0215] in, , The weights of the six monitoring items in the overall tunnel risk assessment; , That is the number of samples in each mini-batch. Indicates the learning rate; The values ​​of the six monitoring indicators for the i-th tunnel segment;

[0216] (4) Optimize the parameters , Substituting these values ​​into the linear regression model expression, we obtain the predicted value for the condition level assessment of the tunnel lining structure.

[0217] (5) Based on the predicted value of the tunnel lining structure condition level assessment, the overall tunnel safety level assessment vector is obtained; wherein the overall tunnel safety level assessment vector is obtained from the tunnel safety level assessment vectors of multiple segments, and the tunnel safety level assessment vectors of multiple segments are obtained from the predicted value of the tunnel lining structure condition level assessment.

[0218] Specifically, the input parameters required for each overall safety level assessment of the tunnel include the crack safety level vector. lining leakage safety level vector Safety level vector of lining voids Safety level vector of steel reinforcement stress Safety level vector of concrete strain Safety level vector of cross-section convergence Based on the above parameters, the overall tunnel safety level assessment vector is obtained. , The formula is:

[0219]

[0220] in, This is the predicted value for the structural condition assessment of the nth tunnel lining.

[0221] S5. Based on the overall tunnel safety level assessment vector, the overall tunnel safety level assessment vector is obtained, and then the overall structural safety level of the tunnel is obtained;

[0222] The formula for calculating the overall structural safety level of the tunnel is as follows:

[0223]

[0224] Based on the tunnel safety level assessment vector and the overall structural safety level of the tunnel, determine the corresponding operation;

[0225] When there is a greater than 2 in the predicted tunnel safety level assessment vector, the edge server sends the information of the corresponding tunnel monitoring section to the staff via SMS.

[0226] The information sent includes the tunnel structure safety level, steel reinforcement stress safety level and monitoring value, concrete strain safety level and monitoring value, cross-sectional convergence safety level and detection value, lining leakage safety level and lining leakage image, tunnel crack safety level and crack image, and lining void safety level.

[0227] When encountering extreme weather or geological disasters, the monitoring instruments will increase the monitoring frequency accordingly and send relevant information to the staff.

[0228] When the required tunnel structural safety level value is between 2 and 4, staff should arrange for personnel to go to the tunnel site for inspection and maintenance; when the tunnel structural safety level value is between 4 and 5, it indicates that the tunnel poses a risk of causing personal and property damage, and staff should immediately arrange for the tunnel to stop operation and implement an emergency repair plan.

[0229] Staff can also access the edge server to view the data; after the data monitoring data is uploaded to the edge server, staff can view the monitoring information and analysis information on the webpage.

[0230] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for on-site monitoring of multi-source tunnel information based on IoT wireless transmission, characterized in that: include: S1. Collect monitoring project image data and upload it to the local server; S2. Based on the image data of the monitoring project, train the vehicle identification model and the ground penetrating radar identification model, evaluate the prediction effect of the model, and deploy the vehicle identification model and the ground penetrating radar identification model to the edge server; S3. Using the vehicle-mounted data acquisition model and the ground-penetrating radar model, the image data uploaded from the local server to the edge server is identified and classified, and the safety level vector of each monitoring item is obtained through the classification results; S4. Based on the safety level vector of each monitoring item, establish a linear regression model and train the model. Based on the trained linear regression model, obtain the overall tunnel safety level assessment vector. S5. Based on the overall tunnel safety level assessment vector, the overall tunnel safety level assessment vector is obtained, and the overall structural safety level of the tunnel is obtained; The safety level vectors for each monitoring item in step S3 include: The crack safety level vector is: ； in, Let n be the vector of crack safety levels, where n is the number of segments where monitoring points are arranged throughout the tunnel. The safety level vector for lining leakage is: ； in, Vector for the safety level of lining leakage; The safety level vector for lining voids is: ； in, Vector for the safety level of lining voids; The safety level vector of steel reinforcement stress is: ； in, The safety level vector for reinforcing steel is: The safety level vector for concrete strain is: ； in, This represents the concrete strain safety level vector. The safety level vector for cross-sectional convergence is: ； in, This is the cross-sectional convergence safety level vector.

2. The method for on-site monitoring of multi-source tunnel information based on IoT wireless transmission according to claim 1, characterized in that: In step S2, the training methods for the data collection vehicle identification model and the ground penetrating radar identification model are the same, except that the uploaded datasets used in the training process of the data collection vehicle identification model and the ground penetrating radar identification model correspond to the datasets obtained by the data collection vehicle and the datasets obtained by the ground penetrating radar, respectively.

3. The method for on-site monitoring of multi-source tunnel information based on IoT wireless transmission according to claim 2, characterized in that: The process of training the data acquisition vehicle recognition model and the ground-penetrating radar recognition model in step S2 includes: Step 1. Preprocess the images in the dataset uploaded from the local server; Step 2. Take 70% of the preprocessed dataset as the training set, 20% as the validation set, and 10% as the test set; Step 3. Build a vehicle identification model and a ground-penetrating radar identification model using the training set, and train them.

4. The method for on-site monitoring of multi-source tunnel information based on IoT wireless transmission according to claim 3, characterized in that: Step 3, the process of training the data acquisition vehicle recognition model and the ground-penetrating radar recognition model, includes: 1) Calculate the ratio of the prior box size to the original image size in the training set in a single multi-boundary box detector; the formula is: ； in, This is the ratio of the size of the prior bounding box corresponding to the k-th feature map to the size of the original image. , These are the maximum and minimum values ​​for this ratio, respectively. , m represents the number of feature maps; 2) Calculate the width of the prior box in a single-shot multi-boundary box detector. and high The calculation formula is: ； ； in, ; 3) Calculate the mapping relationship between the prior bounding boxes on the feature map and the image coordinates of the original images in the training set. The calculation formula is as follows: ； ； ； ； in, , The width and height of the feature layer; , The width and height of the original image; , , , ) is the feature map of the k-th layer centered at (a, b) and of size ( , The prior bounding box is mapped to the coordinates of the original image; 4) The location loss function obtained based on the distance difference between the predicted bounding box and the labeled bounding box is as follows: ； ； in, For the positive sample set, Indicates the first Does the first prior box match the second...? If there are label boxes with category P, This indicates a successful match; otherwise, it is 0. Represents the prediction window center coordinates and width and height Compared to the real frame The offset, where N is the number of candidate region boxes that match the true target box; 5) Based on whether the predicted category matches the labeled category, the confidence loss function is obtained as follows: ； ； in, Indicates the first Softmax confidence regression of positive samples with respect to class P; The set of negative samples; For the first Confidence level of a negative sample Let be the confidence level of the i-th positive sample; 6) Based on the confidence loss function and location loss function described above, the total loss function is obtained as follows: ； in, For confidence loss, For location loss, N is the number of candidate region boxes that match the ground truth bounding boxes; l is the predicted bounding box. For annotation boxes; The weight parameter is set to 1; 7) Based on the total loss function, the corresponding data acquisition vehicle identification model and ground-penetrating radar identification model are obtained.

5. The method for on-site monitoring of multi-source tunnel information based on IoT wireless transmission according to claim 4, characterized in that: The process of evaluating the model's predictive performance includes: Step 1. Use the trained vehicle recognition model and ground-penetrating radar recognition model to detect images in the corresponding test set to obtain TP, FP, and FN. Then, based on TP, FP, and FN, calculate the accuracy and recall of the computational model to predict relevant targets. Where TP is the number of correctly classified and correctly positioned bounding boxes; FP is the number of incorrectly classified or incorrectly positioned bounding boxes; FN is the unpredicted bounding boxes; Recall: the model's ability to predict all relevant targets; Precision: the model's ability to predict relevant targets, its formula is: ； Step 2. Based on the numerical values ​​of precision and recall, evaluate the predictive performance of the model using mAP. The evaluation process includes: first, ； in, It is based on and The result is that, even when using As the y-axis, The curve formed by the x-axis; Secondly ； Where AP is the average precision and C is the number of categories in the object detection task. It is a natural number; When the mAP value is greater than 75, the corresponding model has good predictive performance; when the mAP value is greater than 90, the corresponding model has excellent predictive performance.

6. The method for on-site monitoring of multi-source tunnel information based on IoT wireless transmission according to claim 1, characterized in that: The process of obtaining the overall tunnel safety level assessment vector in step S4 includes: Step 1. Establish a linear regression model based on the safety level vector of each monitoring item; Step 2. Train the linear regression model; Step 3. Obtain the tunnel safety level assessment vector based on the trained linear regression model.

7. The method for on-site monitoring of multi-source tunnel information based on IoT wireless transmission according to claim 6, characterized in that: The linear regression model established in step 1 is as follows: ； in, It is the first The first set of tunnel monitoring data Predicted values ​​for the condition level assessment of the tunnel lining structure at the monitoring points. This refers to the weight of each monitoring item in the overall tunnel risk assessment. It is the first The first set of tunnel monitoring data The safety level of cracks at the monitoring points of the section These are bias parameters; The process of training the linear regression model in step 2 is as follows: (1) Initialize the parameters of the linear regression model , ; (2) Use the loss function to continuously adjust the initialized parameters. , Training ends when the loss function converges. The loss function used is the squared loss function, i.e. ； Where m is the number of tunnel data sets in the entire training sample, and n is the number of segments where monitoring points are deployed throughout the tunnel. It is the first The first set of tunnel monitoring data The target value for the assessment of the condition level of the tunnel lining structure at the monitoring points; (3) Optimized parameters of the loss function , Optimized , for: ； ； in, , The weights of the six monitoring items in the overall tunnel risk assessment; , That is the number of samples in each mini-batch. Indicates the learning rate; The values ​​of the six monitoring indicators for the i-th tunnel segment; (4) Optimize the parameters , Substituting these values ​​into the linear regression model expression, we obtain the predicted value for the condition level assessment of the tunnel lining structure. Step 3 yields the overall tunnel safety level assessment vector. The formula is: ； in, This is the predicted value for the structural condition assessment of the nth tunnel lining.

8. The method for on-site monitoring of multi-source tunnel information based on IoT wireless transmission according to claim 7, characterized in that: In step S5, the formula for calculating the overall structural safety level of the tunnel is as follows: 。

Images