CN114972906A - Soil quality type identification method for excavation surface of soil pressure balance shield - Google Patents

Abstract

The invention relates to the field of earth pressure balance shield tunnel construction. An earth pressure balance shield excavation surface soil type identification method is proposed, which is characterized by comprising the following steps: step 1, establishing a dregs classification system; step 2, establishing a dregs identification database; step 3, dregs identification model 's build. The method of the invention can not only reflect the nature of the soil itself scientifically, but also is suitable for the actual needs of shield construction, and the constructed muck visual identification criterion provides a basis for the production of data set labels; The identification can quickly and accurately obtain the soil type information in front of the excavation, and has the characteristics of real-time and low cost.

Description

A method for identifying soil types of earth pressure balance shield excavation surface

technical field

The invention relates to the field of earth pressure balance shield tunnel construction.

Background technique

With the continuous acceleration of urbanization in China, the mileage and scale of subways have also expanded rapidly, which also brings great challenges to the safety of tunnel construction and the control of the impact on the surrounding environment. The shield method is widely used in the construction of subway tunnels because of its fast construction speed, little impact on the surrounding environment, and high degree of mechanization. However, the construction decision during shield tunneling depends on the understanding of the stratum ahead. When the actual soil quality is inconsistent with the known situation, it will bring a great threat to the construction quality and construction safety. At present, the geological information along the tunnel project mastered by the construction personnel only comes from the geological description provided by the geological exploration before construction. When the actual stratum does not match the stratum information provided in the geological survey report, tunnel quality problems may occur during the excavation process, and even lead to safety accidents in severe cases. Therefore, a system that uses muck images to identify the soil quality of the excavation surface in real time can ensure that the construction personnel can grasp the stratum changes in time, adjust the construction decisions, and ensure the safe and stable excavation of the tunnel.

At present, there are mainly the following solutions for the real-time identification of soil quality on the excavation surface of the shield machine:

The Chinese patent application with application number 202110183343.6 proposes a method for determining stratum characteristics based on real-time excavation parameters of shield tunneling. First, the stratum conditions are pre-classified according to the survey report before tunnel construction, and at the same time, the parameters collected in real time by the shield machine are transformed and analyzed. Process, draw the processed index into a two-dimensional plan, judge whether a new formation type is generated and update the number of formation types. Finally, the standardized parameters are input into the K-Means algorithm, and the formation type determined by the corresponding parameters is output.

The Chinese patent application with the application number of 201911421797.1 proposes a method for driving geological inversion based on the parameter data of the earth pressure balance shield machine. The historical operation data of the earth pressure balance shield machine is processed to prepare the data set, the random forest model is learned on the training set, and the model is verified on the test set. The geological condition information can be obtained by extracting the real-time excavation data of the earth pressure balance shield machine and inputting it into the random forest model.

The Chinese patent application with the application number of 202110152053.5 proposes a real-time soil type identification method based on vibration signals. Firstly, a sensor is installed on the excavation mechanism to collect the vibration signal generated by the excavation, and then the average shear modulus of the contacting soil is analyzed according to the vibration signal. Finally, the analysis result is compared with the data in the soil type database to determine the soil type.

The Chinese patent application with the application number of 202110558201.3 proposes a real-time prediction system and method for soil quality of a shield tunnel construction excavation surface. The prediction system includes a similar engineering data acquisition module, a soil quality information processing module, a construction data processing module, and a soil quality predictor building module. And soil prediction module. The soil quality predictor building module uses the convolutional neural network to construct the first soil quality predictor and the second soil quality predictor based on the grayscale map of the soil quality characteristics and the muck image obtained from the construction data, and combines the prediction results of the two predictors to obtain the real-time soil quality detection. result.

Disadvantages of the prior art

Shield construction parameters are related to various factors such as soil type, shield machine selection, tunnel burial depth, and slag improvement state. Therefore, the geological identification model established by using real-time tunneling parameters to invert geological types can only be constructed under similar construction conditions. Otherwise, the identification effect is difficult to guarantee; the method of using sensors to collect vibration signals to analyze soil parameters requires additional sensors, and the soil can only be classified based on the average shear modulus, a parameter index; using slag soil In the method of soil quality identification by image, the soil quality information corresponding to the image is obtained by analysis and calculation based on the geological survey results before tunnel construction, and the geological survey results themselves have the characteristics of incomplete and inaccurate information, and the distribution of soil layers between boreholes. There is a lot of uncertainty, and the inaccuracy of the label directly affects the recognition effect of the model.

SUMMARY OF THE INVENTION

Technical problem to be solved by the present invention

A method for classifying muck and a system for real-time identification of soil types based on muck monitoring video are designed to solve the problem that it is difficult for shield drivers to obtain information on the stratum ahead. The invention can realize real-time and accurate identification of the soil type of the stratum in front of the cutter head, and provide information for construction decisions such as subsequent adjustment of construction parameters and slag soil improvement plans by the shield driver.

object of the present invention

The core of the present invention is to develop a muck classification system that can combine the engineering properties and engineering experience of the soil itself, and real-time identify the type of soil ahead according to the muck image captured by the shield machine monitoring camera. Using this method, the classification results of the muck based on the construction characteristics of the shield can be obtained, and the muck can be located in real time based on the slag discharge video collected by the surveillance camera, and the geological type ahead can be judged in real time. Features: low cost, automatic, suitable for tunnels in various soil layers.

Technical scheme of the present invention

An earth pressure balance shield excavation surface soil type identification method includes classifying earth pressure balance shield dregs based on shield construction characteristics, and performing image recognition on the excavated dregs based on a deep learning algorithm.

This method determines the soil type of the excavation surface based on the identification of the excavated soil. Therefore, it is necessary to determine the classification of the soil type based on the classification of the soil type. Firstly, a slag classification system is established that considers both the engineering characteristics of the soil and the experience of shield construction. Then collect the muck data, select and process the data to obtain the earth pressure balance shield muck identification data set. Then build and train to get the muck target detection model. Finally, input the monitoring video of the muck during the construction of the earth pressure balance shield to realize the real-time identification of the muck, and obtain the soil type information of the excavation surface.

The steps of this method are as follows:

Step 1. Establish a muck classification system

In order to realize the identification of soil types during shield tunneling, it is the first step to understand the engineering properties of the shallow strata in the area where the shield tunnel is located. Static penetration test (CPT) is an in-situ test method that can comprehensively reflect the physical and mechanical properties of soil. By analyzing the statistical eigenvalues of penetration resistance of different strata, the engineering characteristics of soil mass in each stratum can be grasped, which lays a foundation for the classification of slag.

1.1 First, collect the geological survey report of the area where the shield construction is located, sort out the static penetration layer parameter table, count the penetration resistance value of each borehole through the stratum, and eliminate the abnormal value, that is, the obviously unreasonable data.

1.2 Then calculate the statistical characteristic values of the penetration resistance index of each layer, including the mean, standard deviation and coefficient of variation.

First, the soil layers are classified into categories according to the name of each layer, such as: clay (CL), silt (SI), and sand (SA).

According to the average penetration resistance, which can reflect the properties of soil engineering, it is divided into details, such as clay (CL) is divided into soft clay (SC), ordinary clay (NC), and hard clay (HC).

According to this method, the strata in a certain area can be combined and classified according to their engineering properties, and the classification of soil types suitable for the adjustment of EPB shield parameters can be obtained.

1.3 After classifying the soil types in a certain area, determine the classification of the muck on this basis.

First, consider whether there are special soils or confined water that have a great impact on the earth pressure balance shield construction. If other strata are combined into one category, such strata should be classified separately. In addition, considering that the excavation surface often passes through multi-layer strata during the construction of the shield method, the slag discharged at this time can be designated as a special type of "mixed soil". So far, the classification results of slag soil can be obtained by adjusting on the basis of the classification of soil quality, taking into account the engineering properties of the soil itself and construction experience.

1.4 In the process of establishing the data set later, each muck image needs to be marked with its muck type, so finally, based on the visual characteristics of the muck, a criterion for visually identifying the muck category is constructed, which is convenient for the production of subsequent labels. The muck identification criterion of this method uses four kinds of muck apparent characteristics to determine the muck category, namely muck shape (MS), muck section shape (CS), muck surface flatness (SC), and muck shape (MS), respectively. Color (MC).

Step 2. Establish a muck identification database

2.1 First, the monitoring videos of muck in the area where the method is applied are widely collected.

The general surveillance video resolution is 1920*1080, preferably no less than 1280*720.

The muck monitoring camera should be installed directly above or diagonally above the belt conveyor, aiming at the excavation port of the screw excavator to ensure that the complete slag process and muck form can be photographed. Install lighting equipment above the excavation opening and keep the lens surface clean, so that the captured images are bright, clear and easy to identify.

2.2 Preprocessing the data after collecting the video data.

First, cut out the valid video of the stable unearthed stage, and then extract video frames at a frequency of one frame every 30 frames (the specific frame frequency can be adjusted according to the advancing speed of the shield machine). Images were selected according to the principle of changing the shape of the muck and the image background as much as possible, and the images taken under the contamination of the lens, as well as the poor mosaic images and redundant images caused by video freezes were excluded.

2.3 Use LabelImg image annotation software to manually mark the muck image, frame the muck target in each image with a rectangular frame, and input the muck category of each target frame. The label information is in PASCAL VOC format, and the real target box position information and corresponding muck category information of each image are saved as XML files. In both training and validation sets, each image must be labeled, and the information recorded in it is called the ground truth (GT). This ground truth is provided to the training process in 3.3.4.

2.4 The finally obtained muck image file and corresponding label file is the muck data set, which is divided into training set, validation set and test set according to the ratio of 8:1:1.

Step 3. Construction of muck identification model

The construction process of the muck identification model is also an integral part of the code and the writing order. The present invention uses the convolutional neural network to establish the muck identification model, and its task belongs to the target detection task, so a target detection network needs to be built. The model is built on the deep learning framework Pytorch and implemented in python language.

The muck identification model includes: a custom data set module, a muck detection network building module, a construction training process module, a model testing and evaluation index module, and a muck identification result visualization module.

3.1 In custom dataset module

3.1.1 It is necessary to load the muck image data and the corresponding label information from the data set storage path, and save them in the list respectively.

3.1.2 Then perform certain processing and conversion on the image data and label data. Convert image data from PIL format to Tensor format of shape (C,H,W), where C is the number of image channels, H is the picture height, and W is the picture width, and normalize all pixel values by dividing by 255 to between 0-1.

3.1.3 Then, before the image is input into the network, a certain random image enhancement operation is performed, and the image is enhanced by horizontal flipping, grid masking, random cropping, color jittering and blurring with a probability of 0.5, so as to improve the image quality of the database. Diversity to enhance the robustness of the model.

3.1.4 Calculate the mean and standard deviation of the three-channel pixel values of the picture in the entire data set, standardize the picture to facilitate the learning of the model, and provide it to step 3.3.

3.2 In the muck detection network building module, the network is mainly divided into three parts: the backbone network, the neck network and the top network, which are provided to step 3.3.

The main function of the backbone network is to extract image features, and its basic structure is a combination of convolution layer, batch normalization layer and activation layer. The backbone network can be built by stacking this basic structure. Adding residual connections to the backbone network can increase the depth of the network and effectively improve its feature extraction capability.

The main function of the neck network is feature enhancement and fusion. For example, but not limitation, the spatial pyramid pooling (spatial pyramid pooling, SPP) and path aggregation network (path aggregation network, PANet) used in the present invention are commonly used modules of the neck network. . Among them, the SPP module can effectively expand the receptive field and separate the most important contextual information, and PANet can enhance the extracted image features by fusing the parameters of different levels of the backbone network. The neck network is of great help to the performance improvement of the entire object detection network.

The top network is the detector, and its main function is to make the final regression prediction for the location and category information of the muck. The higher-resolution output feature map contains more detailed features of the input image and is good at detecting small objects, while the lower-resolution output feature map contains coarser features of the input image and is good at detecting large objects.

As an example, the backbone network adopts Darknet53, the neck network adopts SPP and PANet, and the top network consists of object detectors with three output scales. The picture is first input into the backbone network, which consists of 53 layers of convolutional combinations, in which a certain number of residual connections are inserted. Residual connection is to solve the problem of network degradation when the network reaches a certain number of layers when training a deep neural network, that is, it not only fails to improve the expressive ability of the model, but also makes the effect of the model worse. Each convolution combination can be regarded as a function F, and the output predicted value y=F(x) can be obtained by inputting the observed value. The residual connection is divided into two lines, one line is the observation value x input represents F(x) in the convolution combination of the function F, and the other line directly transmits the observation value x, and the final predicted value is the output result of the two lines Add, that is, F(x)+x. If the entire residual connection is regarded as a function H, and the input observation value is x, the predicted value y=H(x)=F(x)+x. F(x)=H(x)-x is the residual, which is the difference between the predicted value y and the observed value x. In this way, the next layer after the residual connection not only contains the information of the previous layer after nonlinear change (convolution combination), but also contains the original information of the previous layer. This processing makes the information only possible to increase layer by layer, and the performance of the model Nor does it decrease as the depth of the network increases.

After the layer-by-layer feature extraction of the backbone network, the feature map already contains high-level semantic information that can identify the muck, but it inevitably loses many detailed features, which is not conducive to the detection of small targets. Therefore, the feature maps trained by the backbone network are input to the neck network for further feature fusion and enhancement.

The feature map at the top of the backbone network is first input into the SPP structure of the neck network. The SPP network performs three different scale pooling operations on the feature map and then stitches them together in the channel dimension, which can solve the problem of repeated feature extraction by the convolutional network. It greatly improves the speed of generating detection candidate frames and saves computing costs. Then, the top feature map of the backbone network and the feature maps of two different levels in the middle of the backbone network are input into the PANet network of the neck network for top-down and bottom-up bidirectional fusion, so that it has both deep semantic information. It also has shallow texture, color and other basic information to ensure the integrity and diversity of features and improve the final prediction effect.

The feature maps of three different levels in the backbone network ("feature map 1, feature map 2, feature map 3") are fused and enhanced by the neck network and then input into the final top network detector respectively, and each undergoes a simple convolution. Combination, regression to get the final prediction result "prediction output 1" "prediction output 2" "prediction output 3". The three feature maps of different sizes contain features of different scales, and also correspond to predicting targets of different sizes. The feature map of prediction output 1 is larger and contains the most detailed information, and is responsible for predicting small target objects; the feature map of prediction output 3 is smaller, which is easier to distinguish the overall information, and is responsible for predicting large target objects. The results of the above three prediction outputs are combined together to be the prediction results of the entire network for the picture.

3.3 In building the training process module

3.3.1 First, import the custom data set module and the muck detection network building module, instantiate the muck data set and the muck detection network, and build a data loader to customize the way of inputting pictures and label data into the network.

3.3.2 Design the loss function of the network. The loss function of the target detection network used in this method is divided into three parts: localization loss, target frame confidence loss and classification loss. ), as shown in formula (1).

Loss=Localization loss+Confidence loss+Classification loss

(1)

The confidence of the target frame represents whether the target frame contains the slag body and the size of the intersection of the target frame and the real frame when it is included.

3.3.3 Many hyperparameters need to be set during network training, including the initial value of the learning rate and how it changes with the increase of the number of training cycles, the optimizer and the parameters of the optimizer, the batch size of input images, the number of training cycles, and the loss function. weight, etc. After setting the loss function and hyperparameters, you can go to step 3.3.4 to start training.

3.3.4 Each batch of images is input to the network, and the prediction result will be obtained. The current loss value can be calculated by inputting the predicted value and the ground truth (GT) in the label of step 2.3 into the loss function, that is, the difference between the predicted value and the real value. the distance. Calculate the derivative of the loss value for all network parameters, and use the optimizer to optimize and update the network parameters, which is a round of training iterations. When all the pictures in the training set are input into the network for training in turn, it is a training loop. After a training cycle is completed, the images of the validation set are input into the network in turn to calculate the network accuracy. Based on this, the network training situation is observed and used as the basis for adjusting the hyperparameters for the next network training. When the loss decreases and converges to a certain stable value, the training can be ended, and the network parameters after training are saved, that is, a recognition model that can accurately locate the muck target and judge the muck type is obtained.

3.4 In the model test and evaluation index module, load the trained model and parameters, and enter the muck image to be tested to obtain the regressed muck location information and category information. Write the evaluation index calculation code to quantitatively evaluate the detection results of the test set, and evaluate the effect of the detection model.

3.5 In the visualization module of muck identification results, the positioning information and category information of muck body target frame obtained by network regression are drawn on the original image to visualize the identification results. Use the center point information x0, y0 and the length and width information h, w of the target frame returned by the network to draw the target frame in the original image, and write the returned slag category information and the corresponding confidence in each target frame in text form At the same time, the target boxes and category information of different muck types are drawn and written in different colors.

beneficial effect

1. The present invention proposes a method for classifying slag based on soil engineering properties and shield construction experience and a criterion for identifying slag vision. The classification result obtained by this method can not only reflect the nature of the soil itself but also be suitable for shield construction. According to the actual needs, the constructed muck visual identification criterion provides a basis for the production of dataset labels.

2. The excavated soil is the most direct data to reflect the type of soil in front, and the determinant of its apparent characteristics is the soil condition in front. Therefore, the deep learning algorithm can be used to identify the excavated soil images quickly and accurately. Cubic soil type information. Moreover, with the expansion of the muck image database, the application range of the muck identification model updated will become wider and wider, and there is no limitation that only a model trained from similar projects can be used to predict a certain engineering soil type.

3. In the process of data collection and practical application, the method proposed in the present invention only needs to use the original muck monitoring video data on site, and does not require additional sensors or other equipment. The reasoning speed of the algorithm itself is also higher than the frame rate of the surveillance video, so this method has the characteristics of real-time and low cost.

Description of drawings

Fig. 1 is the flow chart of the method of the present invention

Fig. 2 is a flow chart for establishing the slag classification system of the present invention

Fig. 3 is the establishment process flow of the muck identification database of the present invention

Fig. 4 is a schematic diagram of data acquisition of muck images in the present invention

Fig. 5 is the construction flow of the muck identification model of the present invention

FIG. 6 is a schematic diagram of a network structure according to an embodiment of the present invention

Detailed ways

The technical solutions provided by the present application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of the present application will become more apparent in conjunction with the following description.

Example

An earth pressure balance shield excavation surface soil type identification method includes classifying earth pressure balance shield dregs based on shield construction characteristics, and performing image recognition on the excavated dregs based on a deep learning algorithm.

The overall flow of the invention is shown in Figure 1.

This method determines the soil type of the excavation surface based on the identification of the excavated soil. Therefore, it is necessary to determine the classification of the soil type based on the classification of the soil type. Firstly, a slag classification system is established that considers both the engineering characteristics of the soil and the experience of shield construction. Then collect the muck data, select and process the data to obtain the earth pressure balance shield muck identification data set. Then build and train to get the muck target detection model. Finally, input the monitoring video of the muck during the construction of the earth pressure balance shield to realize the real-time identification of the muck, and obtain the soil type information of the excavation surface.

Step 1. Establish a muck classification system

The establishment process of the muck classification system is shown in Figure 2. In order to realize the identification of soil types during shield tunneling, it is the first step to understand the engineering properties of the shallow strata in the area where the shield tunnel is located. Static penetration test (CPT) is an in-situ test method that can comprehensively reflect the physical and mechanical properties of soil. By analyzing the statistical eigenvalues of penetration resistance of different strata, the engineering characteristics of soil mass in each stratum can be grasped, which lays a foundation for the classification of slag.

1.1 First, collect the geological survey report of the area where the shield construction is located, sort out the static penetration layer parameter table, count the penetration resistance value of each borehole through the stratum in excel, and eliminate the abnormal value, which is obviously unreasonable. data.

1.2 Then calculate the statistical characteristic values of the penetration resistance index of each layer, including the mean, standard deviation and coefficient of variation. First, the soil layers are classified into categories according to the name of each layer, such as: clay (CL), silt (SI), and sand (SA). According to the average penetration resistance, which can reflect the properties of soil engineering, it is divided into details, such as clay (CL) is divided into soft clay (SC), ordinary clay (NC), and hard clay (HC). According to this method, the strata in a certain area can be combined and classified according to their engineering properties, and the classification of soil types suitable for the adjustment of EPB shield parameters can be obtained.

1.3 After classifying the soil types in a certain area, determine the classification of the muck on this basis. First, consider whether there are special soils or confined water that have a great impact on the earth pressure balance shield construction. If other strata are combined into one category, such strata should be classified separately. In addition, considering that the excavation surface often passes through multi-layer strata during the construction of the shield method, the slag discharged at this time can be designated as a special type of "mixed soil". So far, the classification results of slag soil can be obtained by adjusting on the basis of the classification of soil quality, taking into account the engineering properties of the soil itself and construction experience.

1.4 In the process of establishing the data set later, each muck image needs to be marked with its muck type, so finally, based on the visual characteristics of the muck, a criterion for visually identifying the muck category is constructed, which is convenient for the production of subsequent labels. The muck identification criterion of this method uses four kinds of muck apparent characteristics to determine the muck category, namely muck shape (MS), muck section shape (CS), muck surface flatness (SC), and muck shape (MS), respectively. Color (MC).

Step 2. Establish a muck identification database

Figure 3 shows the process of establishing the muck identification database.

2.1 First, collect the monitoring video of the muck in the area where the method is applied. Generally, the resolution of the monitoring video is 1920*1080, preferably not less than 1280*720. Figure 4 shows the schematic diagram of muck image data collection. The muck monitoring camera should be installed directly above or diagonally above the belt conveyor, aiming at the excavation port of the screw excavator to ensure that the complete slag process and muck form can be photographed. . Install lighting equipment above the excavation opening and keep the lens surface clean, so that the captured images are bright, clear and easy to identify.

2.2 Preprocessing the data after collecting the video data. First, cut out the valid video of the stable unearthed stage, and then extract video frames at a frequency of one frame every 30 frames (the specific frame frequency can be adjusted according to the advancing speed of the shield machine). Images were selected according to the principle of changing the shape of the muck and the image background as much as possible, and the images taken under the contamination of the lens, as well as the poor mosaic images and redundant images caused by video freezes were excluded.

2.3 Use LabelImg image annotation software to manually mark the muck image, frame the muck target in each image with a rectangular frame, and input the muck category of each target frame. The label information is in PASCAL VOC format, and the real target box position information and corresponding muck category information of each image are saved as XML files. In both training and validation sets, each image must be labeled, and the information recorded in it is called the ground truth (GT). This ground truth is provided to the training process in 3.3.4.

2.4 The finally obtained muck image file and corresponding label file is the muck data set, which is divided into training set, validation set and test set according to the ratio of 8:1:1.

Step 3. Construction of muck identification model

The construction process of the muck identification model is shown in Figure 5, which is also a part of the code and the writing sequence. The present invention uses the convolutional neural network to establish the muck identification model, and its task belongs to the target detection task, so a target detection network needs to be built. The model is built on the deep learning framework Pytorch and implemented in python language.

3.1 In custom dataset module

3.1.1 It is necessary to load the muck image data and the corresponding label information from the data set storage path, and save them in the list respectively.

3.1.2 Then perform certain processing and conversion on the image data and label data. Convert image data from PIL format to Tensor format of shape (C, H, W), where C represents the number of image channels, H represents the image height, W represents the image width, and normalize all pixel values by dividing by 255 to between 0-1.

3.1.3 Then, before the image is input into the network, a certain random image enhancement operation is performed, and the image is enhanced by horizontal flipping, grid masking, random cropping, color jittering and blurring with a probability of 0.5, so as to improve the image quality of the database. Diversity to enhance the robustness of the model.

3.1.4 Calculate the mean and standard deviation of the three-channel pixel values of the picture in the entire data set, standardize the picture to facilitate the learning of the model, and provide it to step 3.3.

3.2 In the muck detection network building module, the network is mainly divided into three parts: the backbone network, the neck network and the top network, which are provided to step 3.3.

The main function of the backbone network is to extract image features, and its basic structure is a combination of convolution layer, batch normalization layer and activation layer. The backbone network can be built by stacking this basic structure. Adding residual connections to the backbone network can increase the depth of the network and effectively improve its feature extraction capability.

The main function of the neck network is feature enhancement and fusion. For example, but not limitation, the spatial pyramid pooling (spatial pyramid pooling, SPP) and path aggregation network (path aggregation network, PANet) used in the present invention are commonly used modules of the neck network. . Among them, the SPP module can effectively expand the receptive field and separate the most important contextual information, and PANet can enhance the extracted image features by fusing the parameters of different levels of the backbone network. The neck network is of great help to the performance improvement of the entire object detection network.

The top network is the detector, and its main function is to make the final regression prediction for the location and category information of the muck. The higher-resolution output feature map contains more detailed features of the input image and is good at detecting small objects, while the lower-resolution output feature map contains coarser features of the input image and is good at detecting large objects.

Figure 6 is a schematic diagram of the structure of the embodiment network. The backbone network adopts Darknet53, the neck network adopts SPP and PANet, and the top network consists of target detectors with three output scales. The picture is first input into the backbone network, which consists of 53 layers of convolutional combinations, in which a certain number of residual connections are inserted. Residual connection is to solve the problem of network degradation when the network reaches a certain number of layers when training a deep neural network, that is, it not only fails to improve the expressive ability of the model, but also makes the effect of the model worse. Each convolution combination can be regarded as a function F, and the output predicted value y=F(x) can be obtained by inputting the observed value. The residual connection is divided into two lines, one line is the observation value x input represents F(x) in the convolution combination of the function F, and the other line directly transmits the observation value x, and the final predicted value is the output result of the two lines Add, that is, F(x)+x. If the entire residual connection is regarded as a function H, and the input observation value is x, the predicted value y=H(x)=F(x)+x. F(x)=H(x)-x is the residual, which is the difference between the predicted value y and the observed value x. In this way, the next layer after the residual connection not only contains the information of the previous layer after nonlinear change (convolution combination), but also contains the original information of the previous layer. This processing makes the information only possible to increase layer by layer, and the performance of the model Nor does it decrease as the depth of the network increases.

After the layer-by-layer feature extraction of the backbone network, the feature map already contains high-level semantic information that can identify the muck, but it inevitably loses many detailed features, which is not conducive to the detection of small targets. Therefore, the feature maps trained by the backbone network are input to the neck network for further feature fusion and enhancement.

The feature map at the top of the backbone network is first input into the SPP structure of the neck network. The SPP network performs three different scale pooling operations on the feature map and then stitches them together in the channel dimension, which can solve the problem of repeated feature extraction by the convolutional network. It greatly improves the speed of generating detection candidate frames and saves computing costs. Then, the top feature map of the backbone network and the feature maps of two different levels in the middle of the backbone network are input into the PANet network of the neck network for top-down and bottom-up bidirectional fusion, so that it has both deep semantic information. It also has shallow texture, color and other basic information to ensure the integrity and diversity of features and improve the final prediction effect.

The feature maps of three different levels in the backbone network ("feature map 1, feature map 2, feature map 3") are fused and enhanced by the neck network and then input into the final top network detector respectively, and each undergoes a simple convolution. Combination, regression to get the final prediction result "prediction output 1" "prediction output 2" "prediction output 3". The three feature maps of different sizes contain features of different scales, and also correspond to predicting targets of different sizes. The feature map of prediction output 1 is larger and contains the most detailed information, and is responsible for predicting small target objects; the feature map of prediction output 3 is smaller, which is easier to distinguish the overall information, and is responsible for predicting large target objects. The results of the above three prediction outputs are combined together to be the prediction results of the entire network for the picture.

3.3 In the module that builds the training process

3.3.1 First, import the custom data set module and the muck detection network building module, instantiate the muck data set and the muck detection network, and build a data loader to customize the way of inputting pictures and label data into the network.

3.3.2 Design the loss function of the network. The loss function of the target detection network used in this method is mainly divided into three parts: localization loss, target frame confidence loss and classification loss. loss), as shown in formula (1).

Loss=Localization loss+Confidence loss+Classificaltion loss (1)

The confidence of the target frame represents whether the target frame contains the slag body and the size of the intersection of the target frame and the real frame when it is included.

3.3.3 Many hyperparameters need to be set during network training, including the initial value of the learning rate and how it changes with the increase of the number of training cycles, the optimizer and the parameters of the optimizer, the batch size of input images, the number of training cycles, and the loss function. weight, etc. After setting the loss function and hyperparameters, you can go to step 3.3.4 to start training.

3.3.4 Each batch of images is input to the network, and the prediction result will be obtained. The current loss value can be calculated by inputting the predicted value and the ground truth (GT) in the label of step 2.3 into the loss function, that is, the difference between the predicted value and the real value. the distance. Calculate the derivative of the loss value for all network parameters, and use the optimizer to optimize and update the network parameters, which is a round of training iterations. When all the pictures in the training set are input into the network for training in turn, it is a training loop. After a training cycle is completed, the images of the validation set are input into the network in turn to calculate the network accuracy. Based on this, the network training situation is observed and used as the basis for adjusting the hyperparameters for the next network training. When the loss decreases and converges to a certain stable value, the training can be ended, and the network parameters after training are saved, that is, a recognition model that can accurately locate the muck target and judge the muck type is obtained.

3.4 In the model test and evaluation index module, load the trained model and parameters, and enter the muck image to be tested to obtain the regressed muck location information and category information. Write the evaluation index calculation code to quantitatively evaluate the detection results of the test set, and evaluate the effect of the detection model.

3.5 In the visualization module of muck identification results, the positioning information and category information of muck body target frame obtained by network regression are drawn on the original image to visualize the identification results. Use the center point information x ₀ , y ₀ and the length and width information h and w of the target frame returned by the network to draw the target frame in the original image, and write the returned slag category information and the corresponding confidence in the form of text in each The upper left corner of the target box, and the target boxes and category information of different muck types are drawn and written in different colors.

The muck classification system and the muck visual identification criterion obtained by the method proposed by the present invention will be updated with the continuous expansion of the collection range of the geological prospecting report and the continuous expansion of the muck database, and the muck classification system and the muck classification system with a correspondingly wider application range can be obtained. Visual identification criteria, and updating the training muck identification model at the same time can obtain a muck identification model with a wider application range. Therefore, the method proposed in the present invention is theoretically applicable to the identification of the soil type in front of the earth pressure balance shield excavation in all areas.

Innovative point of the present invention

1. The present invention provides a method for classifying earth pressure balance shield slag based on soil engineering properties and shield construction experience. The classification results of the muck types obtained by this classification method can not only be identified in the image through the deep learning algorithm, but also meet the actual engineering needs during shield construction, laying a classification basis for the subsequent identification of muck images.

2. The present invention provides a method for visual identification of dregs. According to the four important appearance characteristics of muck, namely muck shape, muck cross-section shape, muck surface flatness and muck color, the identification criteria of different muck types are formulated to provide identification basis for the subsequent production of dataset labels.

3. The present invention provides a method for identifying the soil quality type of an earth pressure balance shield excavation surface, that is, based on deep learning, the image recognition of the slag is carried out to obtain the information of the soil quality type ahead. The method can use monitoring video data to perform model training and real-time recognition during application, and has the characteristics of accuracy, speed and low cost.

The above description is only a description of the preferred embodiments of the present application, and is not intended to limit the scope of the present application. Any changes or modifications made by any person of ordinary skill in the field according to the technical contents disclosed above shall be regarded as equivalent effective embodiments, and all belong to the protection scope of the technical solutions of the present application.

Attachment: Terminology Explanation:

Stratigraphy: The general term for all layered rocks, which can be consolidated rocks or unconsolidated sediments, namely soil, which is a layer or a group of uniform features and properties that are distinct from the upper and lower layers. Distinctive rock (soil) layers.

Soil quality: refers to the structure and properties of soil. It also refers to the quality and structure of the soil.

Geological Exploration: The use of geological exploration methods such as surveying and mapping, geophysical exploration, geochemical prospecting, drilling, pit exploration, sampling testing, and geological remote sensing to investigate the geological conditions of rocks, stratigraphic structures, minerals, groundwater, and landforms in a certain area research work.

Static penetration test: The static penetration test is to press the conical probe into the soil at a constant speed with static pressure to measure the penetration resistance (including the resistance of the cone head and the frictional resistance of the side wall or the frictional resistance ratio), and The soil layers are divided according to the resistance they receive to determine the engineering properties of the soil.

Statistical eigenvalues: refers to the representative quantitative characteristics that can accurately describe the distribution of statistical data obtained after sorting out the original data of statistical surveys.

Tunnel: An engineering building buried in the ground, which is a form of human utilization of underground space.

Subway: fast, large-capacity, electric-driven rail transit built in the city, built in tunnels.

Shield method: A tunnel construction method, which uses a shield machine to excavate the ground and assemble the tunnel segments.

Shield machine: a kind of construction machinery, which consists of shell, cutter head, pushing equipment, assembly equipment and other supporting equipment. The shell is a cylinder, which plays a protective role, and other equipment is inside the shell.

Earth pressure balance shield: a type of shield machine, the soil that is cut by the rotation of the front cutter head when the shield is propelled is filled with the soil chamber, and its passive earth pressure is basically balanced with the earth pressure and water pressure on the excavation surface, which makes the excavation The face and the shield face are in equilibrium.

Earth pressure balance shield muck: spoil excavated when the earth pressure balance shield is pushed forward.

Cutter: The equipment used by the shield machine to cut the stratum, it is located at the front end of the shield machine, and the soil is cut off by rotating and extruding.

Face: The face where the excavation tunnel is continuously advancing.

Spiral excavator: a machine that transports the soil cut by the cutter head to the belt conveyor through the rotation of the helical member. The rotation speed of the helical member determines the speed of excavation.

Belt Conveyor: A belt device that transports the soil from the screw excavator to the muck truck.

Shield construction parameters: various parameters that need to be set during the construction of the shield machine, such as cutter head speed, etc. Whether the parameter setting is reasonable determines the safety and quality of shield construction.

Artificial Intelligence: A new technology science that studies and develops theories, methods, techniques and applied systems for simulating, extending and expanding human intelligence.

Machine learning: a multi-field interdisciplinary subject involving probability theory, statistics, approximation theory, algorithmic complexity theory and other subjects. It is the core of artificial intelligence and the fundamental way to make computers intelligent.

Deep learning: It is a new research direction in the field of machine learning. It learns the inherent laws and representation levels of sample data, and the obtained information is very helpful for the interpretation of data such as text, images and sounds.

Target detection: Target detection is an important application of deep learning, which is to identify objects in pictures and mark the location of objects.

Sample: The basic unit of data used by machine learning methods, which can be a one-dimensional matrix or a high-dimensional matrix.

Label: The real information of the data.

LabelImg: A tool for labeling datasets for object detection tasks.

Dataset: The collection of training, validation, and test sets is called a dataset, including data and labels.

Training set: is the data sample fitted by the model and used to debug the neural network.

Validation set: It is a sample set set aside separately during the model training process. It can be used to adjust the hyperparameters of the model and to perform a preliminary evaluation of the model's ability to check the training effect.

Test set: used to evaluate the generalization ability of the final model. However, it cannot be used as the basis for algorithm-related selections such as parameter tuning and feature selection. It is used to test the actual learning ability of the network.

Convolutional Neural Network: It is a type of feedforward neural network that includes convolutional computation and has a deep structure, and is one of the representative algorithms of deep learning.

Image channel: In RGB color mode, it refers to three channels of red, green and blue.

Image enhancement: Add some information or transform data to the original image by certain means, selectively highlight interesting features in the image or suppress (mask) some unwanted features in the image, so that the image matches the visual response characteristics.

Grid mask: Generate a grid with the same resolution as the original image, the gray area value is 1, the black area value is 0, and the enhanced image is multiplied with the original image, which realizes the deletion of information in specific areas, essentially It can be understood as a regularization method.

Color dithering: refers to generating a new image by randomly adjusting the saturation, brightness, and contrast of the original image.

Robustness: refers to the characteristic that the system can still maintain certain performance even when it is disturbed or uncertain.

Convolutional layer: The convolutional layer consists of several convolutional units, and the parameters of each convolutional unit are optimized by the back-propagation algorithm. The purpose of the convolution operation is to extract different features of the input. The first convolution layer may only extract some low-level features such as edges, lines, and corners. More layers of the network can iteratively extract more complex features from the low-level features. feature.

Batch Normalization Layer: Batch Normalization (BN) is a technique used to improve the performance and stability of artificial neural networks. It can normalize the input of any layer in the neural network, fixing the mean and variance of the input signal of each layer.

Activation layer: Activate the input data, that is, nonlinear transformation, map the features to high-dimensional nonlinear intervals for interpretation, and solve problems that cannot be solved by linear models.

Receptive field: The feature vector of a certain position in the characteristic map of a certain layer is calculated from the input of a fixed area in a previous layer, then this area is the receptive field of this position. Commonly used is the receptive field of the input image corresponding to the feature map of a certain layer.

Feature map: The layer obtained after the image is extracted from the convolutional layer.

Loss function: used to evaluate the degree to which the predicted value of the model is different from the true value.

Hyperparameters: Hyperparameters are parameters whose values are set before starting the learning process, not parameter data obtained through training. Usually, hyperparameters need to be optimized, and a set of optimal hyperparameters is selected for the model to improve the performance and effect of learning.

Learning rate: is the step size for each update of the model parameters.

Batch size: The number of samples input to the network for each training session.

Claims (5)

1. A method for identifying soil property types of an excavation surface of a soil pressure balance shield is characterized by comprising the following steps:

step 1, establishing a muck classification system

1.1 collecting geological survey reports of the area where shield construction is located, sorting a static sounding layering parameter table in the geological survey reports, counting penetration resistance values of all drill holes passing through a stratum, and eliminating abnormal values, namely obvious and unreasonable data;

1.2 calculating the statistical characteristic values of the injection resistance indexes of each stratum, including a mean value, a standard deviation and a variation coefficient; firstly, carrying out large-class classification according to soil layer names of all layers; dividing in detail according to the average value of the penetration resistance which can reflect the engineering property of the soil body;

according to the method, strata in a certain area are combined and classified according to the engineering properties of the strata, and soil property type division suitable for adjusting the soil pressure balance shield parameters is obtained;

1.3 after the soil property types of a certain area are divided, determining the classification of the muck on the basis;

firstly, considering whether special soil or confined water and other conditions which have great influence on the construction of the earth pressure balance shield exist, if the stratum with special properties exists, and the stratum and other strata are combined into one when the stratum is classified according to the penetration resistance value, the stratum should be divided separately; in addition, considering the condition that the excavation surface passes through a plurality of layers of stratums when shield construction is carried out, the discharged muck can be determined as a special type of 'mixed soil'; therefore, the classification result of the muck classification considering the engineering property of the soil body and the construction experience can be obtained by adjusting on the basis of the classification of the soil property;

1.4 constructing a criterion for identifying the type of the muck with the eyesight based on the visual characteristics of the muck, so as to facilitate the manufacture of subsequent labels;

step 2, establishing a muck identification database

2.1, widely collecting muck monitoring videos of an area to which the method is applied;

2.2 preprocessing the data after collecting the video data;

intercepting effective videos at a stable unearthing stage, and extracting video frames at a frequency of one frame per 30 frames, wherein the specific frame extraction frequency can be adjusted according to the propelling speed of the shield tunneling machine; selecting images according to the principle that the shape of the muck and the image background are changed as much as possible, and removing images shot under the condition that a lens is stained and mosaic poor images and redundant images caused by video blockage;

2.3 manually marking the muck image by adopting LabelImg image marking software, framing out a muck body target in each image by using a rectangular frame, and inputting the muck type of each target frame;

the label information is in a PASCAL VOC format, and the real target frame position information and the corresponding muck type information of each picture are stored as XML files;

in the training set and the verification set, each image must be provided with a label, and the information recorded in the image is called a true value which is provided for the training process of 3.3.4;

2.4, the obtained muck image file and the corresponding label file are the muck data set, and the muck image file and the corresponding label file are calculated according to the following steps of 8: 1: 1, dividing the ratio into a training set, a verification set and a test set;

step 3, construction of muck recognition model

The muck identification model comprises: the system comprises a user-defined data set module, a muck detection network building module, a construction training process module, a model testing and evaluation index module and a muck identification result visualization module;

3.1 in custom data set Module

3.1.1 loading the muck image data and corresponding label information from the data set storage path, and respectively storing the data and the corresponding label information in a list;

3.1.2 then processing and converting the image data and the label data; converting the image data from a PIL format into a Tensor format with the shape of (C, H, W), wherein C represents the number of image channels, H represents the height of the picture, W represents the width of the picture, and dividing all pixel values by 255 to be normalized to be between 0 and 1;

3.1.3, performing certain random image enhancement operation on the image before the image is input into the network, and performing five kinds of data enhancement of horizontal turning, grid mask, random cutting, color dithering and blurring on the image with the probability of 0.5, so as to improve the diversity of database pictures and enhance the robustness of a model;

3.1.4, calculating the mean value and the standard deviation of three-channel pixel values of the picture in the whole data set, standardizing the picture so as to facilitate the learning of the model, and providing the standard deviation and the mean value for the three-channel pixel values of the picture in the whole data set to the step 3.3;

3.2 in the muck detection network building module, the network is divided into three parts: a backbone network, a neck network and a top network, which are provided for step 3.3;

the main function of the backbone network is to extract image features, the basic structure of the backbone network is the combination of a convolutional layer, a batch normalization layer and an activation layer, and the basic structure is stacked to complete the construction of the backbone network; the depth of the network can be increased by adding residual connection in the backbone network, and the capability of extracting the features of the network is effectively improved;

the main function of the neck network is feature enhancement and fusion;

the top network is a detector, and the main function of the detector is to perform final regression prediction on the position and the category information of the residue soil; the output characteristic diagram with higher resolution contains more detailed characteristics of the input image and is good for detecting small targets, and the output characteristic diagram with lower resolution contains coarser characteristics of the input image and is good for detecting large targets;

3.3 in building the training Process Module

3.3.1, firstly, importing a custom data set module and a muck detection network building module, instantiating a muck data set and a muck detection network, and building a data loader to input a custom picture and tag data into a network;

3.3.2 designing the loss function of the network, wherein the loss function of the target detection network used by the method is divided into three parts: the muck positioning loss, the target frame confidence loss and the muck classification loss are shown in formula (1),

Loss＝Localization loss+Confidence loss+Classification loss (1)

the confidence coefficient of the target frame represents whether the target frame contains the slag soil body or not and the size of the intersection ratio of the target frame and the real frame when the target frame contains the slag soil body;

3.3.3 need set up the hyper-parameter when the network trains, include: the initial value of the learning rate and the change mode of the initial value along with the increase of the training cycle number, the parameters of an optimizer and the optimizer, the batch size of input pictures, the training cycle number, the weight of each item of a loss function and the like; after the loss function and the over-parameters are set, the step 3.3.4 can be entered for training;

3.3.4 each batch of images is input into the network to obtain a prediction result, and the current loss value, namely the distance between the prediction value and the true value can be calculated by inputting the prediction value and the true value in the label in the step 2.3 into a loss function; calculating the derivative of the loss value to all the network parameters, and optimizing and updating the network parameters by using an optimizer, namely a round of training iteration; when all pictures in the training set are input into the network for training in turn, a training cycle is formed; after finishing a training cycle, inputting the pictures of the verification set into the network in turn to calculate the network precision, observing the network training condition according to the network precision, and taking the observed network training condition as a basis for adjusting the hyper-parameters in the next network training; when the loss is reduced and converged to a certain stable value, the training can be finished, and the trained network parameters are stored, so that an identification model capable of accurately positioning the muck target and judging the muck type is obtained;

3.4 loading the trained model and parameters in a model testing and evaluating index module, and inputting a muck picture to be tested to obtain regressed muck positioning information and classification information; compiling an evaluation index calculation code, carrying out quantitative evaluation on the detection result of the test set, and evaluating the effect of the detection model;

3.5 in a muck recognition result visualization module, drawing the positioning information and the category information of the muck body target frame obtained by network regression on an original image to visualize the recognition result; and drawing the target frames in the original image by using the position information x0 and y0 of the central points of the target frames and the length and width information h and w obtained by network regression, writing the regressed muck type information and the corresponding confidence coefficient in the upper left corner of each target frame in a character form, and simultaneously drawing and writing the target frames and the type information of different muck types by adopting different colors.

2. The method for identifying the soil property type of the excavation face of the earth pressure balance shield as claimed in claim 1, wherein the muck identification criterion determines the type of muck by using the apparent characteristics of four types of muck, which are respectively the shape of muck, the shape of the cross section of muck, the surface flatness of muck and the color of muck.

3. The method for identifying the soil property type of the excavation face of the earth pressure balance shield as claimed in claim 1, wherein the resolution of the muck monitoring video is 1920 x 1080 and is not lower than 1280 x 720;

the muck monitoring camera is arranged right above or obliquely above the belt conveyor and is aligned with the muck outlet of the spiral muck remover, so that the complete muck removing process and muck form are ensured to be shot; and the lighting equipment is arranged above the soil outlet, and the surface of the lens is kept clean, so that the shot image is bright and clear and is easy to identify.

4. The method for identifying the soil property of the excavation surface of the earth pressure balance shield as claimed in claim 1, wherein the muck identification model is built on a deep learning frame Pythrch and is implemented by adopting python language.

5. The method for identifying the soil property type of the excavation surface of the earth pressure balance shield as claimed in claim 1, wherein in the step 3.2, a space pyramid pooling and path aggregation network adopted by the residue soil detection network building module is a neck network module; the SPP module can effectively expand the receptive field and separate the most important context information, and the PANet can enhance the extracted image characteristics by fusing parameters of different levels of a backbone network; the neck network is greatly helpful for improving the performance of the whole target detection network;

the backbone network adopts Darknet53, the neck network adopts SPP and PANet, and the top network consists of three target detectors with output scales; firstly, inputting pictures into a backbone network, wherein the backbone network is formed by 53 layers of convolution combination, and a certain number of residual error connections are inserted; the residual error connection is used for solving the problem that when the deep neural network is trained, the network is degraded after a certain number of layers of the network is reached, namely, the expression capability of the model cannot be improved, and the effect of the model is poor; each convolution combination can be regarded as a function F, and an output predicted value y is obtained by inputting an observation value; the residual error is connected and divided into two lines, one line is F (x) in the convolution combination of the observation value x input representation function F, the other line directly transmits the observation value x, and the final predicted value is the addition of the output results of the two lines, namely F (x) + x; if the whole residual join is regarded as a function H and the input observation value is x, the predicted value y ═ H (x) ═ f (x) + x; f (x) h (x) -x is the residual, i.e. the difference between the predicted value y and the observed value x; therefore, the next layer connected by the residual errors not only contains the information of the previous layer after nonlinear change (convolution combination), but also contains the original information of the previous layer, so that the information can only increase gradually layer by processing, and the performance of the model cannot be reduced due to the increase of the network depth;

after layer-by-layer feature extraction of a backbone network, the feature map already contains high-level semantic information capable of identifying the muck, but many detailed features are not lost, so that the detection of small targets is not facilitated; therefore, the feature map obtained by backbone network training is input into the neck network for further feature fusion and enhancement;

the feature graph at the top end of the backbone network is firstly input into an SPP structure of a neck network, the SPP network performs three kinds of pooling operations with different scales on the feature graph and then splices the feature graph with channel dimensions, so that the problem of repeated extraction of features by a convolutional network can be solved, the speed of generating a detection candidate frame is greatly improved, and the calculation cost is saved; secondly, inputting the feature map at the topmost end of the backbone network and the feature maps at two different levels in the middle of the backbone network into a PANet network of a neck network for bidirectional fusion from top to bottom and from bottom to top, so that the feature maps have not only deep semantic information but also basic information such as shallow texture and color, the integrity and diversity of features are ensured, and the final prediction effect is improved;

three feature graphs of different levels in a backbone network, namely a feature graph 1, a feature graph 2 and a feature graph 3, are respectively input into a final top network detector after fusion enhancement of a neck network, and are respectively subjected to simple convolution combination and regression to obtain final prediction results, namely a prediction output 1, a prediction output 2 and a prediction output 3; the three feature maps with different sizes respectively contain features with different scales, and the targets with different sizes are correspondingly predicted; the feature map of the prediction output 1 is large, contains most detail information and is responsible for predicting small target objects; the feature map of the prediction output 3 is smaller, so that the overall information is easier to distinguish, and the prediction output 3 is responsible for predicting a large target object; the three results output by prediction are merged together to obtain the prediction result of the whole network on the picture.

Images