Establishing, training and using method of TBM cutter wear detection model

Abstract

The invention provides a method for establishing, training and using a TBM cutter wear detection model, which comprises the following steps: constructing a basic network structure through a convolutional neural network, adding jump connection, extracting feature maps of all convolutional layers in the basic network structure, using up-sampling and transverse connection, constructing a feature pyramid network, obtaining a feature fusion module, and using a DANN algorithm to obtain a feature fusion model; constructing a domain confrontation training module, constructing a wear detection module by using an Elman algorithm, obtaining an initial TBM cutter wear detection model, obtaining an optimized TBM cutter wear detection model by using a fast sparrow optimization algorithm, training by using historical TBM cutter image data, and completing TBM cutter wear detection by using the trained TBM cutter wear detection model. The problem that detection results are inconsistent due to personal differences in existing TBM tool wear detection is solved, and the method is suitable for TBM tool wear detection.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to a method for establishing, training, and using a TBM tool wear detection model. Background Art

[0002] A Tunnel Boring Machine (TBM) is a large-scale mechanical device used for tunneling, especially suitable for tunneling in hard rock formations. The design of the tunnel boring machine enables it to perform efficient and precise tunnel excavation underground without large-scale excavation operations on the ground. The tools of the tunnel boring machine are indispensable tools in the construction of the tunnel boring machine, mainly used for rock breaking to facilitate tunnel excavation. The tools of the tunnel boring machine are prone to wear during the tunnel boring process, and the detection of the wear degree is crucial for ensuring construction safety and efficiency.

[0003] The existing TBM tool wear detection technology relies on the experience and subjective judgment of inspectors, which is prone to inconsistencies in detection results due to individual differences, with poor accuracy and affecting the construction progress. Summary of the Invention

[0004] The technical problem to be solved by the present invention: The present invention provides a method for establishing, training, and using a TBM tool wear detection model to solve the problem of inconsistent detection results caused by individual differences in the existing TBM tool wear detection.

[0005] The technical solution adopted by the present invention to solve the above technical problem: A method for establishing a TBM tool wear detection model includes the following steps:

[0006] S01. Use a convolutional neural network to construct a basic network structure to obtain a graph feature extraction module, and the basic network structure includes a convolutional layer and a pooling layer connected alternately in sequence;

[0007] S02. Add skip connections, extract the feature maps of all convolutional layers in the basic network structure, and use upsampling and horizontal connection to construct a feature pyramid network to obtain a feature fusion module;

[0008] S03. Use the DANN algorithm to construct a domain adversarial training module, and the domain adversarial training module includes a gradient reversal layer and a domain classifier connected in sequence;

[0009] S04. Use the Elman algorithm to construct a wear detection module;

[0010] S05. Connect both the domain adversarial training module and the wear detection module to the feature fusion module to obtain an initial TBM tool wear detection model including a graph feature extraction module, a feature fusion module, a domain adversarial training module, and a wear detection module;

[0011] S06. Combine the loss functions of the graph feature extraction module, the feature fusion module, the wear detection module, and the domain adversarial training module to obtain a comprehensive loss function;

[0012] S07. Use the fast sparrow search algorithm to optimize the parameters of the initial TBM cutter wear detection model with the goal of minimizing the value of the comprehensive loss function, and obtain the optimized TBM cutter wear detection model.

[0013] Further, in S07, it specifically includes the following steps:

[0014] S071. Set the parameters of the swarm intelligence optimization algorithm, where the parameters include population parameters, maximum number of iterations, search space, and fitness function, and the fitness function is the minimum value function of the comprehensive loss function;

[0015] S072. Initialize the population;

[0016] S073. Use the fitness function to obtain the fitness values of the initial individuals;

[0017] S074. Sort the initial individuals to obtain the initial discoverers, initial joiners, and initial predators;

[0018] S075. Update the initial FSSA population to obtain an updated population; the updated population includes updated discoverers, updated joiners, and updated predators;

[0019] S076. Use the dynamic opposition-based learning algorithm to perform dynamic opposition-based learning on the updated population to generate a dynamically opposed population;

[0020] S077. Calculate the fitness values of all individuals in the updated population and the dynamically opposed population according to the fitness function, and take the individual with the minimum fitness value as the optimal individual;

[0021] S078. If the current iteration number reaches the maximum iteration number or the fitness value of the optimal individual is lower than the fitness value threshold, output the optimal individual;

[0022] S079. Decode the optimal individual to obtain the optimal parameters, and adjust the parameters of the initial TBM cutter wear detection model according to the optimal parameters to obtain the optimized TBM cutter wear detection model.

[0023] The present invention also provides a training method for a TBM cutter wear detection model, which is applied to the optimized TBM cutter wear detection model obtained by using the above-mentioned method for establishing a TBM cutter wear detection model, and includes the following steps:

[0024] S11. Use the preprocessed historical TBM cutter image data as the training set;

[0025] S12. Input the image data in the training set and use the graph feature extraction module to extract the feature map;

[0026] S13. Use the feature fusion module to fuse the extracted feature maps to obtain the fused features;

[0027] S14. Use the wear detection module to predict the fused features to obtain the wear prediction label;

[0028] S15. Use the domain adversarial training module to perform gradient reversal on the fused features, obtain the fused features after gradient reversal, and perform domain classification to obtain the domain classification label;

[0029] S16. According to the domain classification label and the wear prediction label, use the comprehensive loss function to obtain the loss value;

[0030] S17. Traverse all the image data in the training set, repeat S12 to S16 until the loss value is the lowest, and obtain the trained TBM cutter wear detection model.

[0031] The present invention also provides a method for using a TBM cutter wear detection model, which is applied to the trained TBM cutter wear detection model obtained by using the training method of the TBM cutter wear detection model described in claim 3. The method for using includes: collecting TBM cutter image data and using the trained TBM cutter wear detection model for detection to obtain the real-time TBM cutter wear detection result.

[0032] Further, the detection result includes slight wear degree, general wear degree, and severe wear degree.

[0033] Advantages of the present invention: The present invention provides a method for establishing, training, and using a TBM cutter wear detection model. By constructing a basic network structure using a convolutional neural network, a graph feature extraction module is obtained. Skip connections are added to extract the feature maps of all convolutional layers in the basic network structure. Through upsampling and lateral connections, a feature pyramid network is constructed to obtain a feature fusion module. The DANN algorithm is used to construct a domain adversarial training module, and the Elman algorithm is used to construct a wear detection module, thereby obtaining an initial TBM cutter wear detection model and a corresponding comprehensive loss function. Using the fast sparrow optimization algorithm, with the goal of minimizing the value of the comprehensive loss function, the parameters of the initial TBM cutter wear detection model are optimized to obtain an optimized TBM cutter wear detection model. Then, using historical TBM cutter image data to train the optimized TBM cutter wear detection model, a trained TBM cutter wear detection model is obtained. Using the trained TBM cutter wear detection model to complete the TBM cutter wear detection, it solves the problem that the existing TBM cutter wear detection results are inconsistent due to individual differences. Description of the Drawings

[0034] Figure 1 is a schematic flow chart of a method for establishing a TBM cutter wear detection model provided by the present invention;

[0035] Figure 2 is a schematic flow chart of a method for training a TBM cutter wear detection model provided by the present invention. Detailed Embodiments

[0036] In view of the problem that the existing TBM cutter wear detection results are inconsistent due to individual differences, the present invention provides a method for establishing, training, and using a TBM cutter wear detection model. Thereby, using the TBM cutter wear detection model to complete the cutter wear detection through the cutter image, it solves the problem that the existing TBM cutter wear detection results are inconsistent due to individual differences.

[0037] Specifically, the method for establishing a TBM cutter wear detection model is as Figure 1 shown and includes the following steps:

[0038] S01. Use a convolutional neural network to construct a basic network structure to obtain a graph feature extraction module. The basic network structure includes sequentially and alternately connected convolutional layers and pooling layers;

[0039] S02. Add skip connections to extract the feature maps of all convolutional layers in the basic network structure. Through upsampling and lateral connections, construct a feature pyramid network to obtain a feature fusion module;

[0040] S03. Use the DANN algorithm to construct a domain adversarial training module, which includes a gradient reversal layer and a domain classifier connected in sequence;

[0041] S04. Use the Elman algorithm to construct a wear detection module, which is used to predict the wear degree of the tool;

[0042] S05. Connect both the domain adversarial training module and the wear detection module to the feature fusion module to obtain an initial TBM tool wear detection model that includes a graph feature extraction module, a feature fusion module, a domain adversarial training module, and a wear detection module. Specifically, the gradient reversal layer is located between the feature fusion module and the domain classification layer, and its function is to reverse the gradient during backpropagation, so that the domain classifier can learn the features that distinguish the source domain and the target domain during training, while the graph feature extraction module and the feature fusion module are trained to extract and learn domain-invariant features, thus realizing adversarial training. The domain classification layer is used to predict whether the feature belongs to the source domain or the target domain according to the reversed features output by the gradient reversal layer. The domain adversarial training module is used to analyze the differences between the source domain data and the target domain data, so that the graph feature extraction module can accurately extract the invariant features of the target domain, and make the graph feature extraction module and the feature fusion module gradually learn to ignore the domain-related information, improving the prediction efficiency;

[0043] S06. Combine the loss functions of the graph feature extraction module, the feature fusion module, the wear detection module, and the domain adversarial training module to obtain a comprehensive loss function. Specifically, the comprehensive loss function is usually the mean square error, and the loss value is obtained by calculating the mean square error value through each domain classification label and the true classification label.

[0044] S07. Use the fast sparrow optimization algorithm to optimize the parameters of the initial TBM tool wear detection model with the goal of minimizing the value of the comprehensive loss function, and obtain the optimized TBM tool wear detection model. In this way, by optimizing the parameters of the initial TBM tool wear detection model, the sensitivity of the optimized TBM tool wear detection model to the initial value is reduced.

[0045] In S07, it specifically includes the following steps:

[0046] S071. Set the parameters of the swarm intelligence optimization algorithm, and the parameters include population parameters, maximum iteration times, search space, and fitness function. The fitness function is the minimum value function of the comprehensive loss function;

[0047] S072. Initialize the population to obtain the initial population, and the population includes several initial individuals;

[0048] S073. Use the fitness function to obtain the fitness values of the initial individuals;

[0049] S074. Sort the initial individuals to obtain the initial discoverers, initial joiners, and initial predators;

[0050] S075. Update the initial population to obtain an updated population; the updated population includes updated discoverers, updated joiners, and updated predators;

[0051] S076. Use the dynamic reverse learning algorithm to perform dynamic reverse learning on the updated population to generate a dynamically reversed population;

[0052] S077. Calculate the fitness values of all individuals in the updated population and the dynamically reversed population according to the fitness function, and take the individual with the minimum fitness value as the optimal individual;

[0053] S078. If the current iteration number reaches the maximum iteration number or the fitness value of the optimal individual is lower than the fitness value threshold, output the optimal individual;

[0054] S079. Decode the optimal individual to obtain the optimal parameters, and adjust the parameters of the initial TBM tool wear detection model according to the optimal parameters to obtain an optimized TBM tool wear detection model.

[0055] The training method of the TBM tool wear detection model, as Figure 2 shown, includes the following steps:

[0056] S11. Use the preprocessed historical TBM tool image data as the training set;

[0057] S12. Input the image data in the training set and use the graph feature extraction module to extract the feature map;

[0058] S13. Use the feature fusion module to fuse the extracted feature maps to obtain the fused features;

[0059] S14. Use the wear detection module to predict the fused features to obtain the wear prediction labels;

[0060] S15. Use the domain adversarial training module to perform gradient reversal on the fused features to obtain the gradient-reversed fused features, and perform domain classification to obtain the domain classification labels;

[0061] S16. According to the domain classification labels and the wear prediction labels, use the comprehensive loss function to obtain the loss value; specifically, the loss value can be obtained by calculating the mean square error value between each domain classification label and the true classification label;

[0062] S17. Traverse all the image data in the training set, repeat S12 to S16 until the loss value is the lowest, and obtain the trained TBM tool wear detection model.

[0063] The usage method of the TBM tool wear detection model includes: collecting TBM tool image data, using the trained TBM tool wear detection model for detection, and obtaining the TBM tool wear detection result, where the detection result includes slight wear degree, general wear degree, and severe wear degree.

Claims

1. The method for establishing a TBM tool wear detection model is characterized by: The following steps are involved: S01. Use a convolutional neural network to construct a basic network structure to obtain a graph feature extraction module, wherein the basic network structure includes convolutional layers and pooling layers that are alternately connected in sequence; S02, adding skip connections, extracting feature maps of all convolutional layers in the basic network structure, using upsampling and lateral connections, constructing a feature pyramid network, and obtaining a feature fusion module; S03. Use the DANN algorithm to construct a domain adversarial training module, where the domain adversarial training module includes a gradient reversal layer and a domain classifier connected in sequence; S04, using Elman algorithm, construct a wear detection module; S05, connecting the domain adversarial training module and the wear detection module to the feature fusion module to obtain an initial TBM tool wear detection model including a graph feature extraction module, a feature fusion module, a domain adversarial training module and a wear detection module; S06, combining the loss function of the graph feature extraction module, the loss function of the feature fusion module, the loss function of the wear detection module, and the loss function of the domain adversarial training module to obtain a comprehensive loss function; S07. Using the fast sparrow optimization algorithm, with the goal of minimizing the value of the comprehensive loss function, the parameters of the initial TBM tool wear detection model are optimized to obtain the optimized TBM tool wear detection model.

2. The method for establishing a TBM tool wear detection model according to claim 1, characterized in that: S07 specifically includes the following steps: S071. Setting parameters of the swarm intelligence optimization algorithm, the parameters including population parameters, maximum number of iterations, search space and fitness function, the fitness function being a minimum function of the comprehensive loss function; S072, initializing the population; S073. Using the fitness function, obtain the initial individual fitness value; S074. Sort the initial individuals to obtain the initial discoverers, initial joiners and initial predators; S075. updating the initial population to obtain an updated population; the updated population includes an updated discoverer, an updated joiner, and an updated predator; S076. Using a dynamic reverse learning algorithm, dynamically reverse learning is performed on the updated population to generate a dynamic reverse population; S077. According to the fitness function, the fitness values ​​of all individuals in the updated population and the dynamically reversed population are calculated, and the individual with the smallest fitness value is taken as the optimal individual; S078. If the current number of iterations reaches the maximum number of iterations or the fitness value of the optimal individual is lower than the fitness value threshold, the optimal individual is output; S079. Decode the optimal individual to obtain the optimal parameters, adjust the parameters of the initial TBM tool wear detection model according to the optimal parameters, and obtain the optimized TBM tool wear detection model.

3. A training method for a TBM tool wear detection model, characterized in that: The optimized TBM tool wear detection model obtained by the method for establishing the TBM tool wear detection model according to claim 1 comprises the following steps: S11, using the preprocessed historical TBM tool image data as a training set; S12, input the image data in the training set, and use the image feature extraction module to extract the feature map; S13, using a feature fusion module to fuse the extracted feature maps to obtain fusion features; S14, using the wear detection module to predict the fused features and obtain a wear prediction label; S15. Use the domain adversarial training module to perform gradient inversion on the fused features to obtain the fused features after gradient inversion, and perform domain classification to obtain domain classification labels. S16, using a comprehensive loss function according to the domain classification label and the wear prediction label to obtain a loss value; S17, traverse all the image data in the training set, repeat S12 to S16 until the loss value is the lowest, and obtain the trained TBM tool wear detection model.

4. The method for using the TBM tool wear detection model is characterized by: The trained TBM tool wear detection model is applied to the training method of the TBM tool wear detection model as described in claim 3, and the use method includes: collecting TBM tool image data, using the trained TBM tool wear detection model to perform detection, and obtaining TBM tool wear detection results.

5. The method for using the TBM tool wear detection model according to claim 4, characterized in that: The test results include slight wear, moderate wear and severe wear.

Images