Underground disease data enhancement and rapid identification method based on semi-supervised learning

Abstract

The invention relates to an underground disease data enhancement and rapid identification method based on semi-supervised learning, and the method comprises the following steps: S1, constructing a preliminary underground disease data set; s2, obtaining a preliminary training model; s3, enhancing the real detection data set by using a CycleGAN image conversion model, and generating a disease enhanced image data set; s4, a CycleGAN enhanced data set with a pseudo tag is obtained; s5, combining the CycleGAN enhanced data set with the pseudo tag in the step S4 and the preliminary underground disease data set in the step S1 to obtain a final enhanced data set; and S6, inputting the final enhanced data set in the step S5 into the lightweight YOLOv8 model for training to obtain a disease identification model meeting actual engineering requirements. According to the method, the accuracy of pseudo labels in semi-supervised learning is remarkably improved, the problem of data imbalance is solved, and a lightweight model meeting actual requirements is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of ground penetrating radar detection applications, and in particular to a method for enhancing and quickly identifying underground disease data based on semi-supervised learning. Background Art

[0002] With the acceleration of urbanization, underground disease problems (such as cavities, voids, and pipeline aging) pose a major threat to the safety and stability of infrastructure. Ground penetrating radar (GPR), as an efficient non-destructive detection technology, is widely used in the detection of underground diseases. However, traditional GPR data analysis methods rely on manual annotation and empirical judgment, with small amounts of data and uneven quality, especially in the category of underground diseases. This data imbalance problem not only limits the training effect of existing machine learning models, but also makes it difficult for the recognition accuracy and generalization ability of the model to meet the requirements in actual engineering applications. Therefore, how to effectively expand the underground disease dataset and improve the performance of the recognition model has become a hot topic and difficulty in current research.

[0003] In recent years, semi-supervised learning and data enhancement techniques have shown great potential in the field of target detection. By combining numerical simulation and generative adversarial networks (GANs), a large amount of high-quality simulated data can be generated based on limited real data, thus making up for the lack of data. At the same time, lightweight target detection models (such as YOLOv8) have gradually become the preferred choice in engineering practice due to their efficient computing performance and good detection accuracy. However, existing methods still have problems such as insufficient matching between simulated data and real data, limited data enhancement effect, and insufficient optimization of model structure. Therefore, it is urgent to propose a new method that is more suitable for underground disease identification. Summary of the invention

[0004] In view of the shortcomings of the background technology, the technical problem to be solved by the present invention is to provide a method for enhancing and quickly identifying underground disease data based on semi-supervised learning. The method solves the problems of lack and imbalance of underground disease data on roads, while optimizing the recognition model structure and improving the accuracy and efficiency of the actual application of the model.

[0005] The present invention is completed by adopting the following technical solution: a method for underground disease data enhancement and rapid identification based on semi-supervised learning, the steps are as follows: S1. Generate simulated data of underground disease radar images through numerical simulation methods, and fuse the simulated data with the real detection data set to construct a preliminary underground disease data set; S2. Input the preliminary underground disease dataset constructed in S1 into the lightweight YOLOv8 model for training to obtain a preliminary training model; S3. Use the CycleGAN image conversion model to enhance the real detection dataset and generate a disease-enhanced image dataset; S4. Based on the preliminary training model in S2, the disease enhanced image dataset generated in S3 is annotated with pseudo labels to obtain a CycleGAN enhanced dataset with pseudo labels; S5. Combine the CycleGAN enhanced dataset with pseudo labels in S4 and the preliminary underground disease dataset in S1 to obtain the final enhanced dataset; S6. Input the final enhanced data set in S5 into the lightweight YOLOv8 model for training to obtain a disease recognition model that meets the actual needs of the project.

[0006] Furthermore, the numerical simulation method in S1 is based on the forward simulation data of the gprMax open source software. During the simulation process, the simulation is performed according to the formation conditions of the measured data, and the target parameters and radar parameters are randomly set to generate differential images of the disease under the same formation conditions.

[0007] Furthermore, in S2 and S6, the lightweight YOLOv8 model is a lightweight improvement of the YOLOv8 model, and introduces the CBAM attention mechanism and the Focal loss loss function.

[0008] Furthermore, the YOLOv8 model is lightweight and the MobileNetv3-small structure replaces the backbone structure of YOLOv8. The MobileNetV3-small structure uses an efficient feature extraction layer and a lighter network structure.

[0009] Furthermore, the CBAM attention mechanism combines channel attention and spatial attention modules.

[0010] Furthermore, in S3, different images containing target objects, underground diseases and structures are input into the CycleGAN image conversion model to perform image domain style conversion, generate target object images with similar content and different styles, complete the real detection data set enhancement, and generate a disease enhanced image data set.

[0011] Furthermore, the CycleGAN image conversion model adds adversarial loss function and cycle consistency loss function during the conversion process to convert the image styles in two different domains. The formula of the adversarial loss function is as follows: The formula of the cycle consistency loss function is as follows: By combining the two, the final loss function is as follows: Among them, E is a symbol in mathematics, representing the expected value; p data (x) and p data (y) represents the real data distribution sampled from domains X and Y; G and F Represents two generators, G represents the first generator, which is used to convert the x target domain to the y target domain; F represents the second generator, which converts the y target domain to the x target domain; D Denotes the discriminator, D X Determine whether the generated image G(y) comes from the target domain X, D Y Used to determine whether the generated image G(x) comes from the target domain Y; G Y (y) represents the discriminator of the target domain Y for the real image y, D X (x) represents the discrimination of the real image x by the discriminator of the target domain X; D Y (G(x)) represents the discriminator of the target domain Y to the image generator x, D X (G(y)) represents the discriminator of the target domain X to the image generator y; F(G(x)) represents the conversion of x to the target domain Y through the generator G, and then through the generator F The result of converting back to the source domain X; G(F(y)) represents the result of converting y to the target domain X through the generator F, and then converting it back to the source domain Y through the generator G; the weighting factor λ controls the degree of fitting during the image domain conversion process.

[0012] Furthermore, the final enhanced dataset in S5 includes simulated data, real data, and pseudo-labeled data.

[0013] Furthermore, the pseudo-label annotations in S4 are automatically generated by the preliminary training model in S2.

[0014] In the present invention, the radar underground disease data is enhanced by numerical simulation and image generation model, and the enhanced data is introduced in an orderly manner in the semi-supervised learning process to reduce the manual annotation cost and improve the model accuracy. At the same time, the lightweight structure, CBAM and Focal loss are introduced to optimize the YOLOv8 model to ensure high-precision recognition, reduce the number of model parameters, and improve the reasoning speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flowchart of a method for underground disease data enhancement and rapid identification based on semi-supervised learning; Figure 2 This is a numerical simulation diagram of a method for underground disease data enhancement and rapid identification based on semi-supervised learning; Figure 3 A comparison chart of the inference speed of different model structures for underground disease data enhancement and rapid identification based on semi-supervised learning; Figure 4 A comparison chart of the model accuracy of a semi-supervised learning-based underground disease data enhancement and rapid identification method with and without simulated data enhancement datasets; Figure 5 This is a comparison chart of the model accuracy of supervised learning and semi-supervised learning for a method of underground disease data enhancement and rapid identification based on semi-supervised learning. DETAILED DESCRIPTION

[0016] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0017] Reference Figure 1-5 As shown, the present invention provides a method for underground disease data enhancement and rapid identification based on semi-supervised learning, and the specific steps are as follows: S1. Generate simulated data of underground disease radar images through numerical simulation methods, and fuse the simulated data with the real detection data set to construct a preliminary underground disease data set to obtain a gain effect.

[0018] The above-mentioned real detection data set is derived from engineering measurements. However, in actual survey conditions, underground disease data is small in quantity and poor in quality, and is far less than the data of other common targets. This leads to the problem of imbalance in the data volume of diseases and structures in common training data sets. Therefore, the simulation data of underground diseases is generated by numerical simulation methods, and combined with the real detection data set, a preliminary underground disease data set is constructed. Among them, the simulation data mainly covers two types of diseases, voids and voids, to alleviate the problem of small data volume of such targets in real data. The real detection data set includes five targets: voids, voids, pipelines, loose soil, and structures.

[0019] The above numerical simulation method is based on the forward simulation data of the gprMax open source software. The forward simulation establishes a multi-layer road structure and randomly sets parameters such as the center frequency of the ground penetrating radar, the size and shape of the target object to make up for the lack of disease data in the real detection data set, so as to generate diverse radar disease image simulation data that matches the real data, and at the same time improve the accuracy of subsequent pseudo-label annotation. For details, refer to Figure 2As shown in the figure, the simulation is performed according to the formation conditions of the measured data, that is, the parameter settings of the simulated data are consistent with the formation conditions of the real data, and the formation conditions include the number of layers, layer thickness, dielectric constant and other correlation parameters that affect imaging. The target object parameters and radar parameters are randomly set, wherein the target object parameters (i.e., the relevant parameters of the disease) include the shape, size, depth, location or medium of the disease, etc. The center frequency of the randomly set ground penetrating radar is preferably randomly selected between 200MHz, 400MHz, 600MHz and 800MHz. Such a random setting effectively enriches the disease samples and effectively generates differential images of diseases under the same formation conditions, thereby effectively improving the generalization performance of the recognition model and narrowing the gap in the number of disease data sets and other target object samples.

[0020] S2. Use the preliminary underground disease dataset constructed in S1 as the preliminary training dataset and input it into the lightweight YOLOv8 model for training to obtain a preliminary training model, which provides a basis for data enhancement of the subsequent image generation model.

[0021] S3. Use the CycleGAN image conversion model to enhance the real detection dataset and generate a disease enhanced image dataset. Figure 1 The cycle generative adversarial network model in the paper is the CycleGAN image conversion model (abbreviated as CycleGAN, full name Cycle-Consistent Generative Adversarial Network).

[0022] S4. Based on the preliminary training model in S2, the disease enhanced image dataset generated in S3 is annotated with pseudo labels to obtain a CycleGAN enhanced dataset with pseudo labels. Among them, the pseudo labels are automatically generated by the preliminary training model in S2. These labels do not require manual proofreading. As the core part of the semi-supervised learning process, they are intended to provide additional training data for the model, further optimize the model performance, reduce the cost of manual labeling, and improve recognition efficiency.

[0023] S5. Combine the pseudo-labeled CycleGAN enhanced dataset in S4 and the preliminary underground disease dataset in S1 to obtain the final enhanced dataset. The final enhanced dataset includes simulated data, real data, and pseudo-labeled data. It not only balances the sample distribution in the dataset, but also significantly optimizes the generalization ability of the model while improving the model accuracy, and adapts to a wider range of engineering application scenarios.

[0024] S6. Input the final enhanced data set in S5 into the lightweight YOLOv8 model for training to obtain a disease recognition model that meets the actual needs of the project.

[0025] In summary, the present invention combines semi-supervised learning, enhances the accuracy of pseudo labels by simulating data, and uses the CycleGAN model to generate more valid data, thereby improving model performance and reducing labor costs. At the same time, the lightweight YOLOv8 model introduces the CBAM attention mechanism and the Focal loss function to further enhance the model capabilities, and finally obtains an efficient recognition model suitable for actual surveys.

[0026] In S2 and S5 of the above-mentioned underground disease data enhancement and rapid identification method based on semi-supervised learning, the lightweight YOLOv8 model is Figure 1 Model 6 in the figure is also called the Light-weight YOLOv8 model, which is an optimized YOLOv8 model. Specifically, the YOLOv8 model is lightweighted and improved, and the CBAM attention mechanism (Convolutional Block Attention Module) and Focal loss function are introduced to form an efficient lightweight YOLOv8 model (Light-weight YOLOv8 model), which further reduces the number of model parameters while ensuring the accuracy of model recognition, thereby improving the practicability of the model in outdoor surveys.

[0027] Specifically, the YOLOv8 model is light-weighted and improved by replacing the backbone structure of YOLOv8 with the MobileNetv3-small structure. The MobileNetV3-small structure uses an efficient feature extraction layer and a lighter network structure to reduce model parameters and improve the inference speed. MobileNetV3 can significantly reduce the number of model parameters, improve the inference speed, and effectively improve the computational efficiency and application performance in the actual survey process. Focal loss can effectively alleviate the imbalance problem of positive and negative samples in target detection tasks, especially when dealing with engineering application scenarios with significant data imbalance, it can optimize sample weights and improve the model's attention to difficult-to-identify samples. The CBAM attention mechanism combines channel attention and spatial attention modules. The CBAM attention mechanism can provide more effective feature extraction capabilities for convolutional neural networks, improve the model's attention to key features, and thus improve the accuracy of model recognition. The final model is the Light-weight YOLOv8 model, which can alleviate the problems of poor practicality and complex operation caused by too many model parameters.

[0028] In S3 of the above-mentioned method for data enhancement and rapid identification of underground diseases based on semi-supervised learning, different images containing target objects, underground diseases and structures are input into the CycleGAN image conversion model for image domain style conversion, and target object images with similar content and different styles are generated to complete the enhancement of the real detection data set and generate a disease-enhanced image data set, further enriching the data set and helping to improve the generalization ability of the recognition model. Among them, the size of each type of image containing the target object remains consistent. During the conversion process, the CycleGAN image conversion model adds an adversarial loss function and a cycle consistency loss function to convert the image styles in two different domains. The formula of the adversarial loss function is as follows: (1) (2) The formula of the cycle consistency loss function is as follows: (3) By combining the two, the final loss function is as follows: (4) Among them, E is a symbol in mathematics, representing the expected value; p data (x) and p data (y) represents the real data distribution sampled from domains X and Y; G and F Represents two generators, G represents the first generator, which is used to convert the x target domain to the y target domain; F represents the second generator, which converts the y target domain to the x target domain; D Denotes the discriminator, D X Determine whether the generated image G(y) comes from the target domain X, D Y Used to determine whether the generated image G(x) comes from the target domain Y; G Y (y) represents the discriminator of the target domain Y for the real image y, D X (x) represents the discrimination of the real image x by the discriminator of the target domain X; D Y (G(x)) represents the discriminator of the target domain Y to the image generator x, D X (G(y)) represents the discriminator of the target domain X to the image generator y; F(G(x)) in formula (3) represents the conversion of x to the target domain Y through the generator G, and then FThe result of converting back to the source domain X. G(F(y)) represents the result of converting y to the target domain X through the generator F, and then converting it back to the source domain Y through the generator G. The weighting factor λ in formula (4) controls the degree of fitting in the image domain conversion process. The larger the value of λ, the more similar the generated images of the two different style domains are. The preferred value is 10. The combination of formulas (1) to (3) is the final model loss.

[0029] Reference Figure 3 As shown in the figure, in order to further verify the superiority of the Light-weight YOLOv8 model over the ordinary model, the inference speed evaluation indicators of different combinations of improvement methods under the same data set are given (the evaluation indicators include inference speed and memory usage). Figure 3 Model 1 represents the YOLOv8 model; Model 2 represents the YOLOv8+Focalloss model; Model 3 represents the YOLOv8+CBAM model; Model 4 represents the Attention-enhanced YOLOv8 model (YOLOv8+CBAM+Focalloss); Model 5 represents the YOLOv8+MobileNetV3-small model; Model 6 refers to the Light-weight YOLOv8 model (YOLOv8+MobileNetV3-small+CBAM+Focal loss).

[0030] The Light-weight YOLOv8 model is based on a dataset integrated with simulated data, real data, and pseudo-labeled data, and the model accuracy obtained by training different models. Among them, the real data comes from a road survey data, and the target object labels of the real data are manually annotated by experts. At the same time, since the amount of data on the two types of diseases, voids and cavities, is too small in the real data, and these two types of diseases belong to the same type of disease, only the depth and disease characteristics are different, so they are merged into one. Finally, a total of four types of targets are obtained, namely voids and cavities, pipelines, looseness, and structures.

[0031] Table 1 shows the evaluation indicators of different combinations of improvement methods.

[0032] In the present invention, the evaluation index Precision mainly represents the ratio of correctly predicted positive samples to all predicted positive samples, reflecting the accuracy of the model in identifying underground target features; Recall measures the proportion of correctly predicted positive samples to all positive label samples, reflecting the comprehensiveness of the model in detecting underground target features; MAP50 is mainly the average accuracy when the IOU threshold is above 0.5, that is, the area under the precision-recall curve of single category target recognition; FPS measures the recognition speed of the model, indicating the number of frames processed per second; Memory reflects the memory usage of the model.

[0033] By comparing the data in Table 1, it can be found that YOLOv8+MobileNetV3-small+CBAM+Focal loss (i.e., the Light-weight YOLOv8 model proposed in this invention) shows the best balance between prediction accuracy and inference speed. Although its accuracy is slightly lower than that of the standard YOLOv8 model, the lightweight version provides acceptable detection accuracy. More importantly, the lightweight model has more than doubled the improvement in memory efficiency and FPS, reducing memory usage from 11.41 MB to 4.54 MB and increasing FPS from 110 frames / second to 251 frames / second. In actual detection tasks, this improvement is particularly suitable for fast detection on mobile devices.

[0034] Table 1 Evaluation indicators of different improvement method combinations

[0035] Reference Figure 4 As shown in the figure, the accuracy of pseudo-label data with and without simulated data enhancement is shown. DS1 is a data set without simulated data, and DS2 is a data set with simulated data. According to the bar chart, the recognition accuracy of the data set with simulated data is significantly improved in the three categories, which shows that it is feasible to improve the accuracy of pseudo-labels by adding simulated data, proving the effectiveness of the method proposed in the present invention.

[0036] refer to Figure 5 As shown in the figure, the optimization effect of adding the CycleGAN dataset on the model accuracy is demonstrated. The model evaluation indicators obtained by supervised learning and unsupervised learning are compared, and the evaluation indicators include precision, recall rate and average precision. Obviously, semi-supervised learning with the CycleGAN enhanced dataset shows higher model accuracy, proving the superiority of the method proposed in the present invention.

[0037] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A method for underground disease data enhancement and rapid identification based on semi-supervised learning, characterized by: steps as follows: S1. Generate simulated data of underground disease radar images through numerical simulation methods, and fuse the simulated data with the real detection data set to construct a preliminary underground disease data set; S2. Input the preliminary underground disease dataset constructed in S1 into the lightweight YOLOv8 model for training to obtain a preliminary training model; S3. Use the CycleGAN image conversion model to enhance the real detection dataset and generate a disease-enhanced image dataset; S4. Based on the preliminary training model in S2, the disease enhanced image dataset generated in S3 is annotated with pseudo labels to obtain a CycleGAN enhanced dataset with pseudo labels; S5. Combine the CycleGAN enhanced dataset with pseudo labels in S4 and the preliminary underground disease dataset in S1 to obtain the final enhanced dataset; S6. Input the final enhanced data set in S5 into the lightweight YOLOv8 model for training to obtain a disease recognition model that meets the actual needs of the project.

2. The method for underground disease data enhancement and rapid identification based on semi-supervised learning according to claim 1 is characterized by: The numerical simulation method in S1 is based on the forward simulation data of the gprMax open source software. During the simulation process, the simulation is carried out according to the formation conditions of the measured data, and the target parameters and radar parameters are randomly set to generate differential images of the disease under the same formation conditions.

3. The method for underground disease data enhancement and rapid identification based on semi-supervised learning according to claim 1 is characterized by: In S2 and S6, the lightweight YOLOv8 model is a lightweight improvement of the YOLOv8 model, and introduces the CBAM attention mechanism and Focal loss function.

4. The method for underground disease data enhancement and rapid identification based on semi-supervised learning according to claim 3 is characterized by: The YOLOv8 model is lightweight and the MobileNetv3-small structure replaces the backbone structure of YOLOv8. The MobileNetV3-small structure uses an efficient feature extraction layer and a lighter network structure.

5. The method for underground disease data enhancement and rapid identification based on semi-supervised learning according to claim 3 or 4, characterized in that: The CBAM attention mechanism combines channel attention and spatial attention modules.

6. The method for underground disease data enhancement and rapid identification based on semi-supervised learning according to claim 1 is characterized by: In S3, different images containing target objects, underground diseases and structures are input into the CycleGAN image conversion model to perform image domain style conversion, generate target object images with similar content but different styles, complete the real detection data set enhancement, and generate the disease enhanced image data set.

7. The method for underground disease data enhancement and rapid identification based on semi-supervised learning according to claim 6 is characterized by: During the conversion process, the CycleGAN image conversion model adds an adversarial loss function and a cycle consistency loss function to convert the image styles in two different domains. The formula of the adversarial loss function is as follows: The formula of the cycle consistency loss function is as follows: By combining the two, the final loss function is as follows: Among them, E is a symbol in mathematics, representing the expected value; p data (x) and p data (y) represents the real data distribution sampled from domains X and Y; G and F Represents two generators, G represents the first generator, which is used to convert the x target domain to the y target domain; F represents the second generator, which converts the y target domain to the x target domain; D Denotes the discriminator, D X Determine whether the generated image G(y) comes from the target domain X, D Y Used to determine whether the generated image G(x) comes from the target domain Y; G Y (y) represents the discriminator of the target domain Y for the real image y, D X (x) represents the discrimination of the real image x by the discriminator of the target domain X; D Y (G(x)) represents the discriminator of the target domain Y to the image generator x, D X (G(y)) represents the discriminator of the target domain X to the image generator y; F(G(x)) represents the conversion of x to the target domain Y through the generator G, and then through the generator F The result of converting back to the source domain X; G(F(y)) represents the result of converting y to the target domain X through the generator F, and then converting it back to the source domain Y through the generator G; the weighting factor λ controls the degree of fitting during the image domain conversion process.

8. The method for underground disease data enhancement and rapid identification based on semi-supervised learning according to claim 1 is characterized by: The final augmented dataset in S5 includes simulated data, real data, and pseudo-labeled data.

9. The method for underground disease data enhancement and rapid identification based on semi-supervised learning according to claim 1 is characterized by: The pseudo-label annotations in S4 are automatically generated by the preliminary training model in S2.

Images