Shield machine tail gap soft measurement method based on deep learning

Abstract

A shield machine tail gap soft measurement method based on deep learning is provided, which effectively improves the accuracy and robustness of shield machine tail gap measurement. [Solution] The method includes: building a deep learning image segmentation model based on an improved U-Net network; inputting the preprocessed data to be measured into the trained image segmentation model to automatically calculate the projection positions of parallel laser lines on the inner wall and segments of the shield machine shell; and utilizing the spatial characteristics between the shield machine shell and the segments, and based on the dotted line coplanarity principle and pixel ratio relationship, adopting a multi-stage conversion method to convert the laser projection positions obtained by the model calculation into actual shield machine tail gap distances.

Description

[Technical Field]

[0001] The present invention relates to the technical field of shield machine tail gap measurement of shield machines, specifically to a shield machine tail gap soft measurement method based on deep learning. [Background technology]

[0002] Today, China's tunnel and underground construction technology and equipment are constantly improving and becoming more sophisticated. In particular, with the ongoing development of urban rail transit and regional transportation projects, China's tunnel projects are constantly expanding in scale and becoming more difficult. The shield tunneling method is one of the underground excavation methods. It uses a shield machine unit to ensure the stability of the excavation face while safely carrying out tunnel excavation and paving work. Compared to other tunnel construction methods, the shield tunneling method is highly automated, economical, safe, and efficient, and is less susceptible to weather conditions, gradually becoming the mainstream method for underground construction.

[0003] The shield machine tail gap refers to the gap between the inner wall of the shield machine shell and the outer diameter of the segment during shield construction. When joining each segment ring, a certain shield machine tail gap must be left to perform grease sealing, and a grouting system must be used to fill the gap, prevent segment deformation, and improve anti-seepage performance.

[0004] During tunnel construction, the shield machine's propulsion direction and the curvature of the tunnel design affect the length of each cylinder group in the shield machine's propulsion system, resulting in a constantly changing shield machine tail gap. A small shield machine tail gap increases the force exerted by the shield machine tail brushes on the segments, increasing friction between them and the excavation resistance of the propulsion system. This not only reduces excavation speed, but also accelerates wear on the shield machine tail brushes, causing extensive damage. In severe cases, this can lead to extrusion and collision between the segments and the shield machine shell, resulting in segment rupture, allowing water, sand, and mud to flow into the tunnel through the rupture, resulting in serious safety hazards such as ground collapse. Furthermore, the shield machine tail gap can reflect the relative position of the segments and the shield machine tail, providing data support and decision-making basis for segment selection, shield posture adjustment, and other factors, thereby preventing safety accidents. Therefore, fast and efficient shield gap measurement is essential for smooth tunnel construction.

[0005] However, in China's current shield construction process, the main measurement methods for shield machine tail gaps fall into three categories: manual measurement, contact measurement, and non-contact measurement. Manual measurement is the simplest and is widely adopted in construction projects. However, this method suffers from various issues, including low efficiency during construction, low measurement accuracy, and safety hazards. At the same time, the measurement point is not consistent with the manual measurement method, resulting in poor continuity of the shield machine tail gap data obtained. Contact measurement often involves the use of mechanical and electronic devices such as sensors to obtain data. However, the harsh environment of the shield machine tail can easily be affected by mud, dust, and water stains, requiring timely cleaning and maintenance. Furthermore, contact measurement devices wear out over time, necessitating periodic calibration and zero adjustment, or even replacement, which wastes materials and costs. In non-contact measurement, visual detection methods are often used, but segment end face images are easily affected by dust, moisture, mud, and other equipment on site, making it difficult to measure shield machine tail gaps using traditional visual methods. Summary of the Invention [Problem to be solved by the invention]

[0006] The present invention aims to provide a shield machine tail gap soft measurement method based on deep learning, which effectively improves the accuracy and robustness of shield machine tail gap measurement through feature extraction and modeling analysis of high-noise images, provides reliable technical support for the safety and efficiency of shield construction, and solves the problems in the above background technology. [Means for solving the problem]

[0007] To achieve the above objectives, the present invention provides the following technical solutions:

[0008] The shield machine tail gap soft measurement method based on deep learning is Constructing a deep learning image segmentation model based on an improved U-Net network, realized using a convolutional neural network; The image segmentation model adopts a composite loss function including weighted cross-entropy loss and Dice loss. The shield machine tail gap measurement system uses a parallel laser as an auxiliary device, and uses the laser image of the shield machine tail end face collected by an industrial camera as visual measurement data; inputting the preprocessed visual measurement data into the image segmentation model to train it and generate a corresponding model weight file for image segmentation; The pre-processed data to be measured is input into a trained image segmentation model, and the image segmentation model automatically calculates the projection position of the parallel laser line on the inner wall and segment of the shield machine shell; and utilizing the spatial characteristics between the shield machine shell and the segment, and based on the dotted line coplanarity principle and pixel ratio relationship, converting the laser projection position calculated by the image division model into an actual shield machine tail gap distance by adopting a multi-stage conversion method.

[0009] Preferably, the network structure of the image segmentation model is composed of four parts: an encoding module, a dilation and convolution module, a channel attention module, and a decoding module.

[0010] Preferably, the entire encoding and decoding module includes an input layer, four downsampling layers, four upsampling layers, and an output layer. Feature extraction is performed by convolution, downsampling is performed using a max pooling layer, upsampling and dimension recovery are performed using transposed convolution, and feature information from higher layers and feature information from lower layers are fused using a skip link structure. In the network, the outputs of the input layer and the fourth upsampling layer, the first downsampling layer and the third upsampling layer, the second downsampling layer and the second upsampling layer, and the third downsampling layer and the first upsampling layer are fused according to channel dimensions, and the fused feature results are used as the input for the next layer.

[0011] Preferably, the dilation and convolution module adopts a hybrid dilation and convolution design structure, that is, a dilation scheme in which the dilation coefficient is [1,2,5] and the convolution kernel is 3x3.

[0012] Preferably, the channel attention module is (Squeeze-and-Excitation) We employ an attention mechanism to redistribute the weights on the feature map X processed by the neural network, and generate a new feature map Outputs TIFF0007823853000002.tif7150.

[0013] W represents the width of the input feature, H represents the height of the input feature map, and C represents the number of channels in the input feature map. The SE module includes two operation stages: the push operation and the excitation operation.

[0014] In the push-out operation stage, global pooling is used to compress the output features of the convolutional layer into one feature vector, i.e., a 1 × 1 × C feature vector, to obtain feature information with a global receptive field that matches the number of input and output channels. The compression operation is completed by a global average pooling operation to calculate the average value of the feature map of each channel, and the calculation formula is as follows: TIFF0007823853000003.tif1362, X c represents the feature map of the c-th channel of the feature tensor X, and z c represents the average value of the global receptive field of the c channel obtained after the channel is extruded. Feature vector Z=[z1,z2,...z c ] is the feature vector solved in the extrusion operation stage.

[0015] In the excitation operation stage, two fully connected layers are used to obtain the required weight information w through learning, effectively capturing the correlation between channels, and processing the feature vector Z through the two fully connected layer parameter matrices to obtain the desired feature weight information for different channels. The calculation formula is as follows: TIFF0007823853000004.tif972In the formula, δ represents the ReLU activation function, w1 and w2 represent the weight matrices of the two fully connected layers, respectively, and σ represents the Sigmoid function.

[0016] The generated weight vector s is used to assign weights to the feature maps to create a new feature map. The SE module takes TIFF0007823853000005.tif11150 and its size remains unchanged. It simply multiplies the generated feature vector and the original feature map in order according to the channel, that is, it reweights the original features. The SE module does not change the size of the feature map. The calculation formula is as follows: TIFF0007823853000006.tif6150 During the ceremony, TIFF0007823853000007.tif6150 Represents the feature map of the c feature channel in TIFF0007823853000008.tif6150.

[0017] Preferably, we use a composite loss function L of weighted cross-entropy loss and Dice loss for image segmentation. mix is defined as follows: TIFF0007823853000009.tif954 where M represents the label image and G represents the predicted image. wce and L dice represent the weighted cross-entropy loss and Dice loss, respectively, and are calculated as follows: TIFF0007823853000010.tif1779 In the formula, N represents the number of training samples in each batch, and x n,c represents the probability that the predicted image obtained by the network model is in category c, and y n,c represents the probability that the label image corresponding to the predicted image is in category c, and w c represents different weights corresponding to different category predictions in the loss function, and in our case, C=2 represents dividing the image into two categories: background and laser stripe. i and TIFF0007823853000012.tif95 represents the label value and predicted value of pixel i, respectively.

[0018] Preferably, the shield machine tail gap measurement system of the parallel laser assisted device uses two rigidly connected semiconductor laser devices to project two parallel lasers toward the inner wall of the shield machine tail and the end face of the segment, and uses the laser gap images of the inner wall of the shield machine tail and the end face of the segment taken by an industrial camera as visual measurement data.

[0019] Preferably, the multi-step transformation method includes two scanning steps, a least squares step and a scaling step.

[0020] Preferably, the two scan steps mark and classify the laser projection positions obtained by model calculation. The image is scanned twice. In the first scan, label information is assigned to each pixel position, and during the scan, multiple different labels may be assigned to the same connected region. In the second scan, pixels with different label information while belonging to the same connected region are merged. That is, pixels marked with labels having an equivalent relationship are grouped, classified into the same connected region, and the same label information is re-assigned.

[0021] When performing labeling analysis of the connected region using a two-scan algorithm, a threshold operation is added. In the threshold operation, when the number of pixels in the connected region is less than 0.05% of the total number of pixels in the image, that region is discarded, and its calculation formula is as follows. TIFF0007823853000013.tif3649 In the formula, L i represents the label value of the label of the i-th connected region, B i represents the number of pixel values of the i-th connected region, and B0 represents the number of pixel values of the background.

[0022] Preferably, the least squares method step is used for fitting the laser straight line. Here, there are n points (x i , y i )(0 ≤ i < n) for the laser, and the equation of the finally fitted straight line is TIFF0007823853000014.tif830 That is, the slope of the straight line is k, the intercept of the straight line and the y-axis is b, and for each x [[ID=2" "]] i the fitting value is TIFF0007823853000015.tif922 The error function is shown by the following formula. TIFF0007823853000016.tif1359 When the value of the error function is minimized, the values corresponding to k and b become the parameters of the fitting equation. Taking the partial derivative functions with respect to k and b for the error function, the following formula is obtained. TIFF0007823853000017.tif1398 TIFF0007823853000018.tif1385By setting the partial derivatives to 0, we can rearrange to obtain the following system of equations in two variables: TIFF0007823853000019.tif2643 Substituting TIFF0007823853000020.tif1372 into the above formula and rearranging, we get the following: TIFF0007823853000021.tif2326In the formula, the values ​​corresponding to k and b are the parameter values ​​of the fitting line that needs to be solved. The ratio K of the actual distance to the pixel value of the camera image in the scaling stage is obtained by combining the following equations: TIFF0007823853000022.tif923 TIFF0007823853000023.tif917 where L x represents the pixel distance between parallel lasers on the segment, l represents the actual distance between parallel lasers on the segment, and x a ,x b represent the x-axis coordinates of the fitted vertical lines of the laser stripe projected on the left and right sides of the segment, respectively. [Effects of the Invention]

[0023] The present invention has the following advantageous effects compared to the prior art. The present invention introduces a dilation and convolution module to further enhance the ability to complement local and global features, and an SE channel attention module to modify the acquired features and adaptively enhance highly relevant feature information, resulting in excellent interference resistance and laser stripe detection capabilities. At the same time, a two-scan method is used to analyze the connected regions, and deep learning is used to label and classify the acquired laser stripe images. The extracted laser stripes are linearly fitted using the least squares method to indirectly determine the pixel distance of the shield machine tail gap based on the principle of dotted line coplanarity. The actual measured value of the parallel laser spacing and the pixel value are used to convert the shield machine tail gap to the actual distance, providing reliable technical support for the safety and efficiency of shield construction. [Brief explanation of the drawings]

[0024] [Figure 1] 1 is a flowchart of a shield machine tail gap soft measurement method based on deep learning of the present invention. [Figure 2] FIG. 1 is a schematic diagram of a laser stripe detection network model based on the improved U-Net of the present invention. [Figure 3] 1 is a schematic diagram of an inflatable folding structure of the present invention. [Figure 4] FIG. 2 is a schematic diagram of the SE channel attention module of the present invention. [Figure 5] FIG. 1 is a schematic diagram of the average loss of the model training of the present invention. [Figure 6] FIG. 1 is a schematic diagram of the similarity between the model training prediction results of the present invention and the actual labels. [Figure 7] FIG. 1 is a schematic diagram comparing the qualitative analysis results of the present invention and the conventional U-Net network laser stripe detection. [Figure 8] 1 is a schematic diagram of a fitting line for calculating pixel values ​​and actual values ​​of the shield machine tail gap of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described and clarified below with reference to the drawings in the embodiments of the present invention. The described embodiments are only some of the embodiments of the present invention, and are not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without any creative efforts fall within the protection scope of the present invention.

[0026] In practical applications, in conventional vision-based non-contact shield machine tail gap detection methods, the segment end face images are easily interfered with by dust, moisture, mud, and other materials in the construction environment. Therefore, it is difficult to accurately measure the shield machine tail gap using conventional vision methods. To solve this problem, this embodiment provides the following technical solution: The processes in the embodiments may be realized by, for example, a processor included in the processing device executing a program stored in a storage device. Alternatively, the processes in the embodiments may be performed by a hardware circuit included in the processing device.

[0027] FIG. 1 is a flowchart of the shield machine tail gap soft measurement method based on deep learning in this embodiment. Referring to FIG. 1, the shield machine tail gap soft measurement method based on deep learning includes the following steps: Step S0: Train the model. Step S01: Build a shield machine tail gap measurement system equipped with a parallel laser auxiliary device. Use two rigidly connected semiconductor lasers to irradiate the inner wall and segment end face of the shield machine tail with two parallel lasers. Use the laser gap images of the inner wall and segment end face of the shield machine tail taken with an industrial camera as visual measurement data. Step S02: Preprocess the visual measurement data. In this example, the Labelme annotation tool is used to annotate the data and mark the laser stripes in the shield machine tail gap image to separate them from the background. After obtaining the json file, convert it to a PNG image. In this example, 423 images are selected and divided into a training set, validation set, and test set. Step S03: Input the image data into a deep learning image segmentation model based on an improved U-Net network.

[0028] FIG. 2 is a schematic diagram of the laser stripe detection network model based on the improved U-Net of this embodiment. As shown in Figure 2, the network structure of the image segmentation model consists of four parts connected in series: an encoding module, a dilation and convolution module, a channel attention module, and a decoding module.

[0029] The entire encoding and decoding module includes an input layer, four downsampling layers, four upsampling layers, and an output layer. It performs feature extraction using convolution operations, downsampling using a max pooling layer, upsampling and dimension recovery using transposed convolution, and uses a skip link structure to fuse feature information from higher layers with feature information from lower layers. In the network, the outputs of the input layer and the fourth upsampling layer, the first downsampling layer and the third upsampling layer, the second downsampling layer and the second upsampling layer, and the third downsampling layer and the first upsampling layer are fused according to the channel dimensions, and the fused feature results are used as the input for the next layer.

[0030] FIG. 3 is a schematic diagram of the expansion and folding structure of this embodiment. Referring to FIG. 3, the dilation convolution module adopts a HDC (Hybird Dilated Convolution) design structure, that is, a dilation scheme in which the dilation coefficient is [1,2,5] and the convolution kernel is 3×3.

[0031] FIG. 4 is a schematic diagram of the SE channel attention module of this embodiment. Referring to Figure 4, the channel attention module adopts the Squeeze-and-Excitation (SE) attention mechanism. As shown in Figure 4, SE redistributes the weights for the feature map X processed by the neural network, and generates a new feature map The output is TIFF0007823853000024.tif6150. W represents the width of the input feature map, H represents the height of the input feature map, and C represents the number of channels in the input feature map. The SE module can combine two operation stages: extrusion operation and excitation operation.

[0032] The first step is the push-out operation, which uses global pooling to compress the output features of the convolutional layer into one feature vector, i.e., a 1x1xC feature vector, to obtain feature information with a global receptive field that matches the number of input and output channels. The compression operation is usually completed by a global average pooling operation, i.e., calculating the average value of the feature map of each channel. The formula is as follows: TIFF0007823853000025.tif1362, X c represents the feature map of the c-th channel of the feature tensor X, and z c represents the average value of the global receptive field of the c channel obtained after the channel is extruded. Feature vector Z=[z1,z2...z c ] is the feature vector solved in the extrusion operation stage.

[0033] The second stage is the excitation operation, which uses two fully connected layers to obtain the required weight information w through training, effectively capturing the correlation between channels, and then processes the feature vector Z through the two fully connected layer parameter matrices to obtain the desired feature weight information s for different channels. The formula is as follows: TIFF0007823853000026.tif972Here, δ represents the ReLU activation function, w1 and w2 represent the weight matrices of the two fully connected layers, respectively, and σ represents the Sigmoid function.

[0034] The final step is to generate a weight vector s and assign weights to the feature maps to create a new feature map The SE module does not change the size of the feature map. It simply multiplies the generated feature vector and the original feature map in order according to the channel. In other words, it reweights the original features. The SE module does not change the size of the feature map. The formula is as follows: TIFF0007823853000028.tif6150 where, TIFF0007823853000029.tif6150 This represents the feature map of the c feature channel of TIFF0007823853000030.tif16153. Step S04: Repeat step S03 to train the image segmentation model, and generate a corresponding model weight file to be used in the image segmentation model.

[0035] FIG. 5 is a schematic diagram of the average loss of the model training in this example. FIG. 6 is a schematic diagram of the similarity between the model training prediction results and the actual labels in this example. 5-6 show the model training situation in this embodiment. During the training process, the loss of the image segmentation model is calculated using a specially designed composite loss function L, which includes weighted cross-entropy loss and Dice loss. mix was adopted and the definition is as follows: TIFF0007823853000031.tif954 where M represents the label image and G represents the predicted image. wce and L dice represent the weighted cross-entropy loss and Dice loss, respectively, and are calculated as follows: TIFF0007823853000032.tif1779 TIFF0007823853000033.tif2355 where N represents the number of training samples in each batch, and x n,c represents the probability that the predicted image obtained by the network model is in category c, and y n,c represents the probability that the label image corresponding to the predicted image is in category c, and w c represents different weights corresponding to different category predictions in the loss function, and in our case, C=2 divides the image into two categories, background and laser stripe, and y i and TIFF0007823853000034.tif95 represents the label value and predicted value of pixel i, respectively.

[0036] To more accurately demonstrate the performance of the laser stripe detection algorithm provided in this invention, we performed detailed calculations of the laser stripe detection effectiveness through quantitative and qualitative analysis, comparing it with the evaluation indicators of the conventional U-Net algorithm based on manually labeled images. The quantitative analysis is shown in Table 1, and the qualitative analysis is shown in Figure 7. Figure 7 is a schematic diagram comparing the qualitative analysis results of this invention's laser stripe detection with those of the conventional U-Net network.

[0037] Table 1. Quantitative analytical evaluation of the two methods (percentage, %) TIFF0007823853000035.tif41145

[0038] Both the conventional U-Net algorithm and the method provided by the present invention have been shown to effectively avoid interference noise and accurately extract laser stripes. However, when large-scale noise pollution occurs near the laser stripes, the conventional U-Net algorithm may miss some pixels when removing noise interference. The method of the present invention has a high tolerance for noise interference and can demonstrate a certain fitting ability even when the laser stripes are hidden by noise, thereby identifying the laser stripe position and reducing the impact of noise on the accuracy of laser stripe extraction.

[0039] Step S1: Start: Using the method of step S01, obtain a laser measurement image of the shield machine tail gap. Step S2: Image preprocessing: The method of step S02 is used to preprocess the acquired shield machine tail gap laser measurement image. Step S3: Laser stripe extraction: In step S04, the weight file trained by the model is read, image data is input, and step S03 is repeated to output the projection positions of the parallel laser lines on the inner wall and segments of the shield machine shell calculated by the model. Step S4: Linear fitting: The extracted laser stripes are fitted using a multi-step transformation method including two scanning steps, a least-squares step, and a scaling step.

[0040] In this embodiment, the two scanning steps mark and classify the laser projection positions obtained by model calculation. The image is scanned twice. In the first scan, label information is assigned to each pixel position, and during the scanning process, multiple different labels may be assigned to the same connected region. In the second scan, pixels with different label information while belonging to the same connected region are merged. That is, pixels marked with labels having an equivalent relationship are grouped, classified into the same connected region, and the same label information is reassigned.

[0041] When performing labeling analysis of the connected region using a two-scan algorithm, a threshold operation is added. In the threshold operation, when the number of pixels in the connected region is less than 0.05% of the total number of pixels in the image, that region is discarded. The calculation formula is as follows. TIFF0007823853000036.tif3649Here, L i represents the label value of the i-th connected region, and B i represents the number of pixel values in the i-th connected region, and B0 represents the number of pixel values of the background.

[0042] In this embodiment, the least squares method step is used for fitting the laser straight line. Assuming that there are n points (x i , y i )(0 ≤ i < n) on the laser, the equation of the finally fitted straight line is TIFF0OO7823853000037.tif517, that is, the slope of the straight line is k, and the intercept of the straight line and the y-axis is b. For each x i , its fitting value is TIFF0007823853000038.tif922, and the error function is shown by the following formula. TIFF0007823853000039.tif1359When the value of the error function is the minimum, the values corresponding to k and b become the parameters of the fitting equation. Taking the partial derivative functions of k and b with respect to the error function, the following formula is obtained. TIFF0007823853000040.tif1398 TIFF0007823853000041.tif1385By setting the partial derivatives to 0, we can rearrange to obtain the following system of equations in two variables: TIFF0007823853000042.tif2643 Substituting TIFF0007823853000043.tif1372 into the above formula and rearranging, we get the following: TIFF0007823853000044.tif2326Here, the values ​​corresponding to k and b are the parameter values ​​of the fitting line that needs to be solved. The ratio K of the actual distance to the pixel value of the camera image in the scaling stage is obtained by combining the following equations: TIFF0007823853000045.tif923 TIFF0007823853000046.tif917 where L x represents the pixel distance between parallel lasers on the segment, l represents the actual distance between parallel lasers on the segment, and x a ,x b represent the axis coordinates of the fitted vertical lines of the laser stripe projected on the left and right sides of the segment, respectively.

[0043] Step S5: Calculate the pixel value and actual value of the shield machine tail gap: Figure 8 is a schematic diagram of the fitting line used to calculate the pixel value and actual value of the shield machine tail gap in this embodiment. Referring to Figure 8, in step S3, the laser stripe captured by the camera is extracted, annotated, and fitted with a line to obtain the line equations L1 and L2 of the laser projection on the same side segment and the inner wall of the shield machine shell. The equations of the two lines are as follows: TIFF0007823853000047.tif917 TIFF0007823853000048.tif923By solving the two linear equations simultaneously, the coordinate P2 (x2, y2) of the intersection of the two lines can be obtained. The intersection is a point on the inner radial line of the shield machine tail corresponding to the plane on which the end face of the shield machine tail is located. The distance between this point and the coordinate P1 (x1, y1) of the laser end point from the shield machine tail to the segment end face is calculated, and the pixel distance D of the shield machine tail gap is obtained. pix Get. D pix =y2-y1 Then, the pixel distance is converted to the actual distance D based on the ratio K between the actual length and the size of the pixel value. The actual distance of the shield machine tail gap is: TIFF0007823853000049.tif923

[0044] In summary, an extended convolution module is introduced to further enhance the ability to complement local features and global features, and an SE channel attention module is introduced to modify the acquired features and adaptively enhance highly relevant feature information, resulting in excellent anti-interference and laser stripe detection capabilities. At the same time, a two-scan method is used to analyze the connected area, and deep learning is used to label and classify the acquired laser stripe images. The extracted laser stripes are linearly fitted using the least squares method to indirectly obtain the pixel distance of the shield machine tail gap based on the principle of dotted line coplanarity, and the actual value and pixel value of the parallel laser spacing are used to convert the shield machine tail gap into an actual distance, providing reliable technical support to ensure the safety and efficiency of shield construction.

[0045] It should be noted that, as used herein, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another and do not necessarily require or imply an actual relationship or ordering between those entities or operations. Furthermore, the use of "comprises," "including," or other similar terms is intended to be non-exclusive, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed or inherent in such process, method, article, or apparatus.

[0046] While embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A shield machine tail gap measurement method based on deep learning, Constructing a deep learning image segmentation model based on an improved U-Net network, implemented using a convolutional neural network; The image segmentation model employs a composite loss function including a weighted cross-entropy loss and a Dice loss; The shield machine tail gap measurement system uses a parallel laser as an auxiliary device, and uses the laser image of the shield machine tail end face collected by an industrial camera as visual measurement data; inputting the preprocessed visual measurement data into the image segmentation model to train it and generate a corresponding model weight file for image segmentation; The pre-processed data to be measured is input into a trained image segmentation model, and the image segmentation model automatically calculates the projection position of the parallel laser line on the inner wall and segment of the shield machine shell; Utilizing the spatial characteristics between the shield machine shell and the segment, and based on the pixel ratio relationship, converting the laser projection position calculated by the image division model into the actual shield machine tail gap distance by adopting a multi-stage conversion method; Including, The network structure of the image segmentation model consists of four parts: an encoding module, a dilated convolution module, a channel attention module, and a decoding module. The entire encoding and decoding module includes an input layer, four downsampling layers, four upsampling layers, and an output layer. It performs feature extraction through convolution operations, downsampling using a max pooling layer, upsampling and dimension recovery using transposed convolution, and fuses feature information from higher layers with feature information from lower layers using a skip link structure. In the network, the outputs of the input layer and the fourth upsampling layer, the first downsampling layer and the third upsampling layer, the second downsampling layer and the second upsampling layer, and the third downsampling layer and the first upsampling layer are fused according to the channel dimensions, and the fused feature results are used as the input for the next layer. The dilation and convolution module adopts a mixed dilation and convolution design structure, that is, the dilation coefficient is [1, 2, 5], and the convolution kernel is 3 × 3. The channel attention module employs a squeeze-and-excitation (SE) attention mechanism to redistribute the weights on the feature map X processed by the neural network, thereby generating a new feature map X. and W represents the width of the input feature, H represents the height of the input feature map, and C represents the number of channels in the input feature map. Here, the SE module includes two operation stages: a squeeze operation and an excitation operation. In the push-out operation stage, global pooling is used to compress the output features of the convolutional layer into a 1x1xC feature vector, obtaining feature information with a global receptive field where the number of input and output channels is the same. The compression operation is completed through the global average pooling operation, and the average value of the feature map of each channel is solved, and the calculation formula is and where X c represents the feature map of the c-th channel of the feature tensor X, z c represents the mean value of the global receptive field of the c-th channel obtained after the channel is extruded, The feature vector Z = [z 1 , z 2 . . . z c ] is the feature vector solved in the extrusion operation stage, In the excitation operation stage, two fully connected layers are used to obtain the required weight information w through learning, and the correlation between channels is grasped. The feature vector Z is processed by the two fully connected layer parameter matrices to obtain the desired feature weight information s for different channels. The calculation formula is and where δ represents the ReLU activation function, w 1 and w 2 represent the weight matrices of the two fully connected layers, respectively, and σ represents the Sigmoid function. The weight vector s is used to assign weights to the feature maps to create a new feature map. The size of the feature vector is not changed, and the generated feature vector and the original feature map are multiplied in order according to the channel, that is, the original features are reweighted. The SE module does not change the size of the feature map, and the calculation formula is and During the ceremony, teeth represents the feature map of the c feature channel in The composite loss function L mix of the weighted cross-entropy loss and Dice loss in the image segmentation model is defined as follows: where M represents the label image, G represents the predicted image, L wce and L dice represent the weighted cross-entropy loss and Dice loss, respectively, which are calculated as follows: where N represents the number of training samples in each batch, x n,c represents the probability that the predicted image obtained by the network model is in category c, y n,c represents the possibility that the labeled image corresponding to the predicted image is in category c, w c represents different weights in the loss function corresponding to different category predictions, C=2 represents dividing the image into two categories, background and laser stripe, and y i and represent the label value and predicted value of pixel i, respectively, The shield machine tail gap measurement system of the parallel laser auxiliary device uses two rigidly connected semiconductor laser devices to project two parallel lasers toward the inner wall of the shield machine tail and the end face of the segment, and uses the laser gap image of the inner wall of the shield machine tail and the end face of the segment taken by an industrial camera as visual measurement data. A deep learning-based shield machine tail gap measurement method characterized in that the multi-stage transformation method includes two scanning stages, a least squares stage, and a scaling stage.

2. The two scanning stages mark and classify the laser projection positions obtained by model calculation, scan the image twice, in the first scan, assign one label information to each pixel position, during the scanning process, multiple different labels may be assigned to the same connected region, in the second scan, merge pixels that belong to the same connected region but have different label information, group pixels marked with labels that have an equivalent relationship, classify them into the same connected region, and reassign the same label information, When using the two-scan algorithm to perform connected region labeling analysis, a threshold operation is added. In the threshold operation, if the number of pixels in a connected region is less than 0.05% of the total number of pixels in the image, the region is discarded. The calculation formula is: and The deep learning-based shield machine tail gap measurement method of claim 1, characterized in that, in the formula, L i represents the label value of the i-th connected region, B i represents the number of pixel values ​​of the i-th connected region, and B 0 represents the number of pixel values ​​of the background.

3. The least squares step is used to fit a laser line, and there are n points (x i , y i ) (0≦i<n) on the laser, and the equation of the final fitted line is That is, the slope of the line is k, the intercept of the line with the y-axis is b, and for each x i , the fitted value is and the error function is given by the following equation: When the value of the error function is minimum, the values ​​corresponding to k and b become the parameters of the fitting equation, Taking the partial derivatives of the error function with respect to k and b gives us Setting the partial derivatives to zero, we can rearrange to obtain the following system of equations in two variables: Substituting into the above formula and rearranging, we get: where the values ​​corresponding to k and b are the parameter values ​​of the fitting line that needs to be solved, The ratio K of the actual distance to the pixel value of the camera image in the scaling stage is obtained by combining the following equations:

3. The deep learning-based shield machine tail gap measurement method of claim 2, wherein Lx represents the pixel distance between the parallel lasers on the segment, l represents the actual distance between the parallel lasers on the segment, and xa and xb represent the x-axis coordinates of the fitted straight lines of the laser stripes projected on the left and right sides of the segment, respectively.

Images