Fine identification method for flexure structure based on transformer algorithm

Abstract

This invention provides a method for fine-grained identification of flexural structures based on the Transformer algorithm, comprising: extracting a set of flexural-sensitive attributes from 3D seismic data, performing attribute fusion and sample construction to obtain fused attribute samples and corresponding label datasets; combining a U-shaped network and the Transformer algorithm to obtain a hybrid network, and performing serialization and semantic embedding on the fused attribute samples to obtain serialized embedding features; inputting the serialized embedding features into the hybrid network, performing hybrid encoding and global context modeling, as well as upsampling reconstruction and skip fusion, to obtain pixel-level flexural segmentation results; performing training configuration and hyperparameter setting to obtain hyperparameter configuration, using the hyperparameters for network training and convergence to obtain an identification model, and outputting flexural structure identification results. This invention improves the expression of complex geometry and scale changes in flexural structures, achieving fine-grained identification and improved accuracy of small-scale flexural structures.

Description

Technical Field

[0001] This invention relates to the field of seismic facies identification, and in particular to a method for fine identification of flexural structures based on the Transformer algorithm. Background Technology

[0002] With the further development of mechanized coal mining and the expansion of mining areas into deeper regions, geological structures such as flexures have brought significant technical challenges to coal mine exploration and mining. The existence of these geological structures leads to uneven distribution of underground coal seams, directly affecting gas migration and coal mining efficiency. Therefore, accurate interpretation of underground geological structures is particularly urgent. Flexures are often associated with accidents such as roof falls, water inrushes, and gas outbursts, which not only increase the difficulty of coal seam mining and result in low overall efficiency of fully mechanized mining, but also pose hidden safety hazards.

[0003] Traditional anomaly identification methods often rely on expert experience to characterize anomalies. For example, fault identification involves analyzing discontinuities in the reflected wave phase axes along the seismic profile. While traditional anomaly location characterization based on expert experience can identify some relatively clear and large anomalies, it lacks objectivity and reliability for detailed anomaly descriptions. Advances in computer technology have enabled the application of numerous advanced techniques in seismic signal processing, solving many substantial problems. For instance, seismic attribute and edge detection techniques are used for fault identification. Methods for cave identification include seismic attribute analysis and seismic inversion methods.

[0004] Currently, research on deflection identification methods is limited. Based on domestic and international research, two deflection identification techniques are summarized: 1) conventional deflection identification methods, including seismic profile identification and auxiliary detection methods; 2) deflection identification methods based on seismic attributes, including some layer-by-layer or volume attributes that are sensitive to deflection, such as coherence attributes, variance attributes, and curvature attributes. With the continuous development and deepening of geophysical exploration, complex and deep-seated unconventional oil and gas reservoirs are gradually coming into view. This makes traditional structural interpretation methods no longer adequate for the current exploration challenges. Conventional algorithms are no longer sufficient for processing large amounts of seismic data and performing refined target identification. Although deep learning has shown advantages in faults and collapse columns, the locality of convolution limits global dependency modeling; when structures are non-locally correlated or have large scale spans, it is prone to "not seeing far ahead," and the intelligent identification of small-scale deflection still needs improvement. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a fine recognition method for flexural structures based on the Transformer algorithm, which improves the expression of complex geometry and scale changes in flexural structures, and achieves fine recognition and accuracy improvement of small-scale flexural structures.

[0006] To achieve the above objectives, the present invention provides the following solution: a method for fine recognition of flexural structures based on the Transformer algorithm, comprising: A set of flexural sensitive attributes is extracted from 3D seismic data. Based on the set of flexural sensitive attributes, attribute fusion and sample construction are performed to obtain fused attribute samples and corresponding label datasets. By combining the U-shaped network and the Transformer algorithm, a hybrid network is obtained, and the fused attribute samples are serialized and semantically embedded to obtain serialized embedded features. The serialized embedded features are input into the hybrid network for hybrid encoding, global context modeling, upsampling reconstruction, and skip fusion to obtain pixel-level flexural segmentation results; Based on the pixel-level flexure segmentation results, training configuration and hyperparameter settings are performed to obtain the hyperparameter configuration. The hyperparameters are then used for network training and convergence to obtain the recognition model and output the flexure structure recognition results.

[0007] Optionally, a set of flexural sensitivity attributes is extracted from the 3D seismic data, and attribute fusion and sample construction are performed based on the set of flexural sensitivity attributes to obtain fused attribute samples and corresponding label datasets, including: The root mean square amplitude, fracture orientation, fracture density, fracture simulation, fracture intensity, location of maximum value within time window, instantaneous frequency, instantaneous phase, smart ants, and coherent enhancement are extracted from 3D seismic data to obtain relevant sensitive attributes. Based on the delineated flexural boundary, the relevant sensitive attributes are compared and analyzed with the delineated area to screen out fracture density, smart ants, and coherent enhancement, thus obtaining a set of flexural sensitive attributes. Based on the set of flexure-sensitive attributes, crack density, smart ants and coherence enhancement are fused in three channels to obtain fused attribute slices. Combined with the flexure-defined image, the fused attribute slices are manually cropped to generate fused feature sample pairs and binarized flexure label pairs. The fused sample pairs are augmented and expanded, and divided into training, validation and test sets in a ratio of 7:1.5:1.5 to obtain fused attribute samples and corresponding label datasets.

[0008] Optionally, a hybrid network is obtained by combining a U-shaped network and the Transformer algorithm, and the fused attribute samples are serialized and semantically embedded to obtain serialized embedding features, including: The U-shaped network is used as the basic skeleton to process local details and multi-scale features, and the Transformer is used as a global vision supplement to capture long-distance dependencies, resulting in a hybrid network that includes encoding, global context, decoding reconstruction, and skip fusion. Based on the fused attribute samples, a high-dimensional feature map is extracted using a convolutional backbone. The high-dimensional feature map is then divided into two-dimensional image blocks of fixed size. The two-dimensional image blocks are flattened into vectors, and each two-dimensional image block is embedded into a high-dimensional space of the model dimension through linear mapping. Then, a positional encoding is added to each two-dimensional image block to inject spatial positional information, resulting in a serialized embedded feature.

[0009] Optionally, the serialized embedded features are input into the hybrid network for hybrid encoding and global context modeling, as well as upsampling reconstruction and skip fusion, to obtain pixel-level flexural segmentation results, including: The serialized embedded features are input into the transformer encoder of the hybrid network and stacked in multiple layers. Each encoder layer performs global association capture and feature expression enhancement on the serialized embedded features in sequence to iteratively generate progressively enhanced global context representations, while retaining the large multi-scale feature maps of the convolutional encoders and outputting high-dimensional representations including global dependency information. Based on the high-dimensional representation, a cascaded upsampler is introduced to upsample step by step to gradually restore the spatial resolution. Then, the high-dimensional representation is fused with the features of the corresponding layer at the encoding end through skip connections to output a pixel-level flexural segmentation result with the same resolution as the input. Each stage of the cascaded upsampler includes two upsampling, convolution, and activation operations.

[0010] Optionally, based on the pixel-level flexure segmentation results, training configuration and hyperparameter settings are performed to obtain the hyperparameter configuration. The hyperparameters are then used for network training and convergence to obtain the recognition model, which outputs the flexure structure recognition results, including: Select the loss function, design an adaptive moment estimation optimization algorithm with a learning rate of 0.01, set the batch size and maximum period, complete the training configuration, perform iterative training of the hybrid network, and perform early convergence observation after the 60th iteration to output the training configuration and hyperparameter configuration, obtain the recognition model, and output the flexural structure recognition result. The average crossover ratio and pixel accuracy are set as indicators. Based on the indicators, the U-shaped network and the hybrid network are compared to obtain the comparison results, and the quantitative evaluation of the model is completed. Then, the flexural structure recognition results are visualized and applied in actual engineering.

[0011] This invention discloses the following technical effects by providing a fine-grained identification method for flexural structures based on the Transformer algorithm: 1. Global-local collaboration: The long-range dependency modeling of the converter complements the detail and boundary recovery of the U-shaped network, which can see the whole picture and grasp the details, and is particularly suitable for flexural structures with large shape changes and diverse scales.

[0012] 2. Attribute-driven input fusion: The three attributes of crack density, smart ants and coherence enhancement are fused into three channels with minimal information redundancy, focusing on key clues that are sensitive to flexural distortion, thereby improving the signal-to-noise ratio and learning efficiency.

[0013] 3. Reproducible and efficient engineering: Clear training configuration, data augmentation and convergence strategies, stable after 60 iterations, training time of about 30 minutes, meeting the engineering time efficiency requirements.

[0014] 4. Significant quantization improvement: Compared with the U-shaped network, it significantly improves the average crossover ratio and pixel accuracy, verifying the advantages of the hybrid architecture in recognizing small-scale deflection.

[0015] 5. Visually friendly: The pixel-level output probabilities and masks are conducive to linkage with geological interpretation and risk assessment, supporting well location optimization, tunnel design and safety early warning.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 2 Earthquake crack density slice diagram provided for embodiments of the present invention; Figure 3 This is a slice image of an earthquake-resistant intelligent ant provided in an embodiment of the present invention; Figure 4 This is a slice diagram of coherence enhancement properties provided in an embodiment of the present invention; Figure 5 This is a fusion attribute slice diagram provided in an embodiment of the present invention; Figure 6 This is a fusion attribute slice cropping image provided in an embodiment of the present invention; Figure 7 This invention provides a fusion attribute slice cropping binarized label image for embodiments of the invention. Figure 8 The model training prediction graph provided in the embodiments of the present invention; Figure 9 The graph shows the model training loss function provided in the embodiments of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, this invention provides a method for fine recognition of flexural structures based on the Transformer algorithm, including: Step 1: Extract the set of flexural sensitive attributes from the 3D seismic data, and perform attribute fusion and sample construction based on the set of flexural sensitive attributes to obtain fused attribute samples and corresponding label datasets.

[0022] Step 1 specifically includes: 1.1 Extract root mean square amplitude, fracture azimuth, fracture density, fracture simulation, fracture intensity, location of maximum value within time window, instantaneous frequency, instantaneous phase, smart ants, and coherent enhancement from 3D seismic data to obtain relevant sensitive attributes. Based on the delineated flexural boundary, compare and analyze the relevant sensitive attributes with the delineated area to screen out fracture density, smart ants, and coherent enhancement, thus obtaining a set of flexural sensitive attributes.

[0023] 1.2 such as Figure 5 , Figure 6 , Figure 7 As shown, based on the set of flexure-sensitive attributes, crack density, intelligent ants, and coherence enhancement are fused in three channels to obtain fused attribute slices. Combined with the flexure-defined image, the fused attribute slices are manually cropped to generate fused feature sample pairs and binarized flexure label pairs. Wherein: 1) Seismic crack density slice like Figure 2 As shown, seismic fracture density slicing technology is a geological analysis method based on seismic data, used to identify and quantify the distribution and density of underground fractures. By processing and analyzing seismic reflection data, it generates slice images reflecting the distribution and density of underground fractures, providing an effective means for identifying and quantifying underground fractures. Flexure is often accompanied by fractures and faults in strata. By analyzing fracture density slices, areas of high fracture density can be identified, revealing densely populated areas of fractures related to fault activity, thereby aiding in fault identification.

[0024] 2) Earthquake-induced intelligent ant body slices like Figure 3As shown, intelligent ant slicing technology is a geological analysis method based on image processing and intelligent algorithms. It utilizes an ant colony algorithm to simulate the path-searching behavior of ants in seismic data, thereby extracting continuous structural information from seismic reflection features. By simulating the "crawling" trajectory of ants in seismic profiles, it automatically identifies potential faults, fractures, and other structural boundaries, generating slice images that reflect the connectivity and geometric features of geological structures, providing an intelligent means for structural interpretation. The connectivity paths extracted by intelligent ants often correspond to underground fault zones, fault planes, and fracture channels. By analyzing intelligent ant slices, discontinuous areas within seismic reflections can be effectively highlighted, revealing possible fault activity locations, thus aiding in the extraction of fault boundaries and the reconstruction of structural morphology.

[0025] The paths extracted by the intelligent ants are densely distributed in a band near the flexural boundary, forming a continuous direction of the structural boundary. The sample flexures are roughly distributed inside or near the dense path bands. This extraction effect is highly consistent with the structural stress concentration characteristics during the flexural formation process.

[0026] 3) Coherence Enhancement Attribute like Figure 4 As shown, coherence enhancement is a geological analysis method based on the coherence of seismic signals. By calculating the similarity of seismic waveforms, it enhances the continuity and clarity of structures such as reflection interfaces and geological faults in seismic data. This attribute can highlight the geometric features of subsurface structures, help identify fractures, faults, and other discontinuities, and provide more intuitive and accurate information support for seismic interpretation. Coherence enhancement attributes are usually closely related to faults and fractures, and their variations reflect the integrity and degree of fragmentation of the internal structure of geological bodies. By analyzing coherence enhancement slices, stratigraphic interfaces and fracture zones can be clearly distinguished, revealing the spatial distribution of fault development and assisting in fault location and reservoir evaluation.

[0027] The coherent enhancement property manifests as a distinct continuous fracture zone in the flexural region, reflecting the dense distribution of faults in the tectonic deformation area. The sample flexural region is located near the boundary of the high coherent enhancement value region, showing a distribution pattern closely related to the tectonic stress field and fracture development.

[0028] 1.3 The fused sample pairs are augmented and expanded, and divided into training, validation and test sets in a ratio of 7:1.5:1.5 to obtain fused attribute samples and corresponding label datasets.

[0029] For example, the research dataset originally contained 128 samples, which were expanded to 2560 samples after data augmentation. These samples were then divided into training, validation, and test sets in a ratio of 7:1.5:1.5. The training set contained 1792 samples and was used to update model parameters. The validation and test sets each contained 384 samples. The former was used to monitor model performance and adjust hyperparameters during training, while the latter was used to evaluate the model's generalization ability.

[0030] Step 2: Combine the U-shaped network and the Transformer algorithm to obtain a hybrid network, and perform serialization and semantic embedding on the fused attribute samples to obtain serialized embedding features.

[0031] Step 2 specifically includes: 2.1 The U-shaped network is used as the basic skeleton to process local details and multi-scale features, and the Transformer is used as a global vision supplement to capture long-distance dependencies, resulting in a hybrid network that includes encoding, global context, decoding reconstruction, and jump fusion.

[0032] 1) U-shaped network: Based on the concept of fully convolution, an "encoder-decoder + skip connections + feature concatenation" structure is adopted. Upsampling is performed at the decoding end, and shallow positional information is fused with deep semantic information through concatenation, improving detail recovery and localization accuracy. This fusion, based on convolution after concatenation, has a high computational cost, but it is beneficial for multi-scale detail and semantic fusion, improving resolution and segmentation accuracy.

[0033] 2) Transformer: It consists of stacked encoders and decoders, with multi-head self-attention, feedforward networks, residual connections, and layer normalization at its core. The attention mechanism works by calculating the similarity weights between the "query" and all "keys," then weighting and aggregating the "values" after "soft maximum" normalization to achieve adaptive global information aggregation.

[0034] Multi-head strategy: The query, key, and value are split into multiple "heads" in the dimension, the dependencies of different subspaces are learned in parallel, and finally the data are spliced ​​and merged to extract a more comprehensive representation.

[0035] Feedforward network: Two layers of linear transformation and activation are used for nonlinear feature mapping and channel transformation; residuals and normalization are used for stable training and gradient propagation.

[0036] 3) Hybrid networks: Using the convolutional features of a U-shaped network as input, the algorithm first performs "image blockization and linear embedding + positional encoding" and feeds it into the transformer encoder to model global dependencies. At the decoding end, cascaded upsampling (CUP) and skip connections are used to fuse the global context with shallow local details, progressively restoring the original resolution. This hybrid paradigm addresses both global and local aspects of image segmentation, overcoming the weaknesses of pure CNNs in handling long-range dependencies and the insufficient detail restoration of pure transformers.

[0037] 2.2 Based on the fused attribute samples, a high-dimensional feature map is extracted using a convolutional backbone. The high-dimensional feature map is then divided into two-dimensional image blocks of fixed size. The two-dimensional image blocks are flattened into vectors, and each two-dimensional image block is embedded into a high-dimensional space of the model dimension through linear mapping. Then, a positional encoding is added to each two-dimensional image block to inject spatial positional information, thereby obtaining serialized embedded features.

[0038] Step 3: Input the serialized embedded features into the hybrid network to perform hybrid encoding and global context modeling, as well as upsampling reconstruction and skip fusion, to obtain pixel-level flexural segmentation results; Step 3 specifically includes: The serialized embedded features are input into the transformer encoder of the hybrid network and stacked in multiple layers. Each encoder layer performs global association capture and feature expression enhancement on the serialized embedded features in sequence to iteratively generate progressively enhanced global context representations, while retaining the large multi-scale feature maps of the convolutional encoders and outputting high-dimensional representations including global dependency information. Based on the high-dimensional representation, a cascaded upsampler is introduced to upsample step by step to gradually restore the spatial resolution. Then, the high-dimensional representation is fused with the features of the corresponding layer at the encoding end through skip connections to output a pixel-level flexural segmentation result with the same resolution as the input. Each stage of the cascaded upsampler includes two upsampling, convolution, and activation operations.

[0039] Step 4, as follows Figure 8 , Figure 9 As shown, based on the pixel-level flexure segmentation results, training configuration and hyperparameter settings are performed to obtain the hyperparameter configuration. The hyperparameters are then used for network training and convergence to obtain the recognition model and output the flexure structure recognition results.

[0040] Step 4 specifically includes: 4.1 Select the loss function, design an adaptive moment estimation optimization algorithm with a learning rate of 0.01, set the batch size and maximum period, complete the training configuration, perform iterative training of the hybrid network, and perform early convergence observation after the 60th iteration to output the training configuration and hyperparameter configuration, obtain the recognition model, and output the flexural structure recognition result.

[0041] For example, the TransUNet model was trained using a preprocessed training dataset, with an epoch of 100 and a batch size of 8. As training progressed, the loss function gradually decreased, converging after 60 iterations. Training stopped at 100 iterations, with a total training time of 30 minutes. Test results showed that the loss function stabilized after 60 iterations, and the change was less than 0.01 after more than 60 iterations, indicating that excessive iterations had minimal impact on the experimental results. To balance computational efficiency and training stability, 60 iterations were ultimately selected. Furthermore, multiple experiments verified that a training time of 30 minutes and stable recognition results were achieved with an epoch of 100 and a batch size of 8; therefore, these were chosen as the final hyperparameter settings.

[0042] 4.2 The average crossover ratio and pixel accuracy are set as indicators. Based on the indicators, the U-shaped network and the hybrid network are compared to obtain the comparison results, and the quantitative evaluation of the model is completed. Then, the flexural structure recognition results are visualized and applied in actual engineering.

[0043] The comparison results are as follows: U-shaped network: average crossover ratio 0.571560, pixel accuracy 0.720236; Hybrid network: Average crossover ratio 0.735125, pixel accuracy 0.851403.

[0044] Conclusion: The hybrid network significantly outperforms the U-shaped network in terms of flexural segmentation accuracy and overall pixel-level accuracy.

[0045] Therefore, this invention provides a fine recognition method for flexural structures based on the Transformer algorithm, which improves the expression of complex geometry and scale changes in flexural structures, and achieves fine recognition and accuracy improvement of small-scale flexural structures.

[0046] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0047] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for fine recognition of flexural structures based on the Transformer algorithm, characterized in that, include: A set of flexural sensitive attributes is extracted from 3D seismic data. Based on the set of flexural sensitive attributes, attribute fusion and sample construction are performed to obtain fused attribute samples and corresponding label datasets. By combining the U-shaped network and the Transformer algorithm, a hybrid network is obtained, and the fused attribute samples are serialized and semantically embedded to obtain serialized embedded features. The serialized embedded features are input into the hybrid network for hybrid encoding, global context modeling, upsampling reconstruction, and skip fusion to obtain pixel-level flexural segmentation results; Based on the pixel-level flexure segmentation results, training configuration and hyperparameter settings are performed to obtain the hyperparameter configuration. The hyperparameters are then used for network training and convergence to obtain the recognition model and output the flexure structure recognition results.

2. The method for fine recognition of flexural structures based on the Transformer algorithm according to claim 1, characterized in that, A set of flexural sensitivity attributes is extracted from 3D seismic data. Based on this set, attribute fusion and sample construction are performed to obtain fused attribute samples and corresponding label datasets, including: The root mean square amplitude, fracture orientation, fracture density, fracture simulation, fracture intensity, location of maximum value within time window, instantaneous frequency, instantaneous phase, smart ants, and coherent enhancement are extracted from 3D seismic data to obtain relevant sensitive attributes. Based on the delineated flexural boundary, the relevant sensitive attributes are compared and analyzed with the delineated area to screen out fracture density, smart ants, and coherent enhancement, thus obtaining a set of flexural sensitive attributes. Based on the set of flexure-sensitive attributes, crack density, smart ants and coherence enhancement are fused in three channels to obtain fused attribute slices. Combined with the flexure-defined image, the fused attribute slices are manually cropped to generate fused feature sample pairs and binarized flexure label pairs. The fused sample pairs are augmented and expanded, and divided into training, validation and test sets in a ratio of 7:1.5:1.5 to obtain fused attribute samples and corresponding label datasets.

3. The method for fine recognition of flexural structures based on the Transformer algorithm according to claim 2, characterized in that, By combining a U-shaped network and the Transformer algorithm, a hybrid network is obtained. The fused attribute samples are then serialized and semantically embedded to obtain serialized embedding features, including: The U-shaped network is used as the basic skeleton to process local details and multi-scale features, and the Transformer is used as a global vision supplement to capture long-distance dependencies, resulting in a hybrid network that includes encoding, global context, decoding reconstruction, and skip fusion. Based on the fused attribute samples, a high-dimensional feature map is extracted using a convolutional backbone. The high-dimensional feature map is then divided into two-dimensional image blocks of fixed size. The two-dimensional image blocks are flattened into vectors, and each two-dimensional image block is embedded into a high-dimensional space of the model dimension through linear mapping. Then, a positional encoding is added to each two-dimensional image block to inject spatial positional information, resulting in a serialized embedded feature.

4. The method for fine recognition of flexural structures based on the Transformer algorithm according to claim 3, characterized in that, The serialized embedded features are input into the hybrid network for hybrid encoding, global context modeling, upsampling reconstruction, and skip fusion to obtain pixel-level flexural segmentation results, including: The serialized embedded features are input into the transformer encoder of the hybrid network and stacked in multiple layers. Each encoder layer performs global association capture and feature expression enhancement on the serialized embedded features in sequence to iteratively generate progressively enhanced global context representations, while retaining the large multi-scale feature maps of the convolutional encoders and outputting high-dimensional representations including global dependency information. Based on the high-dimensional representation, a cascaded upsampler is introduced to upsample step by step to gradually restore the spatial resolution. Then, the high-dimensional representation is fused with the features of the corresponding layer at the encoding end through skip connections to output a pixel-level flexural segmentation result with the same resolution as the input. Each stage of the cascaded upsampler includes two upsampling, convolution, and activation operations.

5. The fine-grained identification method for flexural structures based on the Transformer algorithm according to claim 4, characterized in that, Based on the pixel-level flexure segmentation results, training configuration and hyperparameter settings are performed to obtain the hyperparameter configuration. The hyperparameters are then used for network training and convergence to obtain the recognition model, which outputs the flexure structure recognition results, including: Select the loss function, design an adaptive moment estimation optimization algorithm with a learning rate of 0.01, set the batch size and maximum period, complete the training configuration, perform iterative training of the hybrid network, and perform early convergence observation after the 60th iteration to output the training configuration and hyperparameter configuration, obtain the recognition model, and output the flexural structure recognition result. The average crossover ratio and pixel accuracy are set as indicators. Based on the indicators, the U-shaped network and the hybrid network are compared to obtain the comparison results, and the quantitative evaluation of the model is completed. Then, the flexural structure recognition results are visualized and applied in actual engineering.

Images