Fracture neighborhood corresponding stress thermal map prediction and prediction model training method based on deep learning
By using an end-to-end pixel-level regression network based on deep learning, the relative stress distribution heatmap can be predicted directly from crack images. This solves the efficiency and accuracy problems of stress distribution prediction in the crack neighborhood in existing technologies, and achieves fast and accurate localization of stress concentration areas, making it suitable for engineering assessment under complex conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies struggle to quickly and accurately predict relative stress distribution thermograms from crack images, and their robustness and adaptability are insufficient, especially under complex conditions, failing to meet the needs of rapid health assessment for engineering structures.
An end-to-end pixel-level regression network based on deep learning is adopted. The model is trained through crack neighborhood images and combined with an encoder, a bridging multi-scale context fusion module and a decoder to output a high-resolution relative stress distribution heatmap. A composite loss function is introduced to improve prediction accuracy and robustness.
It achieves rapid and accurate mapping from crack images to stress field distribution, reduces computational costs, improves the efficiency and accuracy of engineering assessment, is applicable to complex real-world scenarios, and has the ability to accurately locate areas of high stress concentration.
Smart Images
Figure CN122154194A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of crack stress field inversion technology, specifically a method for predicting relative stress thermograms in the crack neighborhood and training prediction models based on deep learning. Background Technology
[0002] Concrete structures (especially bridge structures) account for a large proportion of transportation and construction industries and have long service lives. With increasing service life, the performance of structural materials deteriorates, and problems such as cracks gradually appear and develop. The appearance of cracks not only affects the aesthetics and durability of the structure but also directly reflects the redistribution of stress within the structure, easily leading to stress concentration, accelerated fatigue damage, reduced load-bearing capacity, and other safety hazards, seriously threatening structural safety. If high-stress concentration areas and their distribution patterns can be quickly and accurately identified in the early stages of crack initiation or propagation, it can provide crucial information for preventative maintenance, repair, and reinforcement decisions, thereby effectively reducing life-cycle maintenance costs and significantly improving safety risk management capabilities.
[0003] Currently, technologies for monitoring and evaluating structural stress states are mainly divided into two categories: contact physical measurement and numerical simulation.
[0004] (1) Contact sensor measurement. This type of method obtains the strain or dynamic response at local locations by deploying sensors (such as resistance strain gauges, fiber optic grating sensors, accelerometers, etc.) on the surface or key parts of the structure, and then calculates the stress state. Although this technology is relatively mature, it has obvious limitations: First, the number of deployment points is limited, the information obtained is discrete, and it is difficult to achieve high spatial resolution full-field stress distribution coverage; second, the sensors are susceptible to interference from environmental temperature fluctuations, electromagnetic noise, etc., and the stability and durability of long-term monitoring are challenged, resulting in high maintenance costs; in addition, for areas where cracks have appeared, the sensor installation itself may introduce new stress concentrations or it is difficult to effectively deploy them in key areas such as crack tips.
[0005] (2) Finite element numerical simulation. This method calculates the stress distribution across the entire field by solving the mechanical control equations based on the accurate geometric model of the structure, the constitutive relations of the materials, the boundary conditions, and the load input. The finite element method can provide high-resolution stress field results, but its application has significant bottlenecks: First, establishing an accurate finite element model itself requires detailed geometric and material parameters, and the calculation process is complex and time-consuming, making it difficult to meet the needs of rapid screening or near real-time evaluation of large-scale infrastructure; Second, the calculation accuracy of the model depends heavily on the accurate setting of boundary conditions and material parameters, while these parameters of actual engineering structures are often uncertain, requiring complex parameter inversion and model calibration, which further increases the difficulty and cost of application.
[0006] In recent years, with the rapid development of computer vision and deep learning technologies, non-contact image monitoring methods have been widely used in bridge inspection and surface damage identification. Numerous studies have used deep learning to achieve automatic detection, localization, and quantification (e.g., width and length measurement) of surface defects such as cracks and spalling, or to visually measure the overall displacement and deflection of structures. However, these works mainly focus on extracting surface geometric information or macroscopic deformation information from images, falling under the category of "morphological perception." They cannot yet bridge the physical gap by directly inferring the invisible internal stress concentration distribution around cracks from image features, thus failing to achieve "mechanical state perception."
[0007] On the other hand, to accelerate the finite element analysis process, the academic community has proposed finite element surrogate model technology, which utilizes neural networks to learn the mapping relationship from finite element input parameters (such as geometric dimensions, loads, boundary conditions, etc.) to the output stress field. However, existing surrogate models mostly rely on manually defined abstract parameters as input, and their application still presupposes that a complete mechanical model and working conditions are known or assumed. For the most common situation in engineering practice, where only images of the crack surface can be obtained, there is a lack of a solution that can use the original crack neighborhood image as direct input and output the corresponding relative stress heat map end-to-end. In addition, the surface of real bridge structures has complex textures, stains, uneven lighting, and other interferences, and the robustness and adaptability of existing visual methods and surrogate models under such complex conditions still need to be improved.
[0008] Therefore, there is an urgent need in this field to develop an innovative technical solution that can take easily obtainable crack neighborhood images as input, quickly and stably output the corresponding relative stress distribution heat map with low computational cost, and accurately locate high stress concentration areas, so as to fill the gap in existing technology and meet the urgent needs of rapid health assessment and intelligent operation and maintenance of engineering structures. Summary of the Invention
[0009] In view of this, the purpose of this invention is to provide a deep learning-based method for predicting relative stress heatmaps in the neighborhood of cracks and training prediction models. Through an end-to-end pixel-level regression network, a high-resolution relative stress distribution heatmap can be quickly predicted directly from the crack grayscale image, enabling accurate positioning of high stress concentration areas such as crack tips, thereby improving the efficiency and engineering practicality of structural health assessment.
[0010] To achieve the above objectives, the present invention provides the following technical solution: This invention first proposes a deep learning-based method for training a crack neighborhood relative stress thermogram prediction model, comprising: S1: Obtain training samples: The training samples include crack neighborhood structure images and stress field data corresponding to the structure images; S2: Data Processing The crack neighborhood structure image is converted into a crack grayscale image and adjusted to a preset input size; The stress field data is preprocessed, including at least outlier handling, spatial alignment, and scale unification; the preprocessed stress field data is then normalized to obtain a normalized relative stress heatmap for supervised learning. S3: Model Training The grayscale image of the crack is input into an end-to-end pixel-level regression network, and a single-channel prediction result is output. The composite loss is calculated based on the normalized relative stress heatmap and the single-channel prediction results, and the network parameters are iteratively updated to complete the training.
[0011] Furthermore, the method for normalizing the stress field data is as follows: Mini-maximum normalization is performed on the stress field data of a single sample, mapping the stress value range of each sample to the [0,1] interval, expressed as: in: This is the result after normalization; This is the preprocessed stress field data; and These are the minimum and maximum values within the sample, respectively. This represents a matrix consisting entirely of zeros.
[0012] Furthermore, the end-to-end pixel-level regression network includes an encoder, a bridging multi-scale context fusion module, and a decoder; The encoder is used to extract multi-scale features and form a hierarchical pyramid representation; The bridging multi-scale context fusion module is used to perform multi-scale fusion on the features output by the encoder. The decoder is used to upsample the output of the bridging multi-scale context fusion module step by step to restore the spatial resolution, and to achieve multi-scale feature fusion through skip connections of features of the same scale as the encoder, and finally output the single-channel prediction result.
[0013] Furthermore, the encoder uses a compact variant of the ConvNeXt V2 series as the backbone network, and its initialization parameters are obtained by self-supervised pre-training based on a mask autoencoder. The bridging multi-scale context fusion module adopts a hollow spatial pyramid pooling structure, which includes multi-branch feature transformation. The features of each branch are concatenated in the channel dimension and then fused through 1×1 convolution to obtain the bridging result.
[0014] Furthermore, the composite loss includes at least a global regression loss term and a hotspot priority loss term; the global regression loss term is used to constrain the overall prediction error; and the hotspot priority loss term is used to assign higher weights to the high-stress regions of the normalized relative stress heatmap.
[0015] Furthermore, the global regression loss term includes at least one of mean squared error, mean absolute error, and Soft Dice loss; Mean square error is expressed as: The mean absolute error is expressed as: The soft Dice loss is expressed as: in: Batch size; and The height and width of the heatmap space; and The first Each sample at pixel coordinates The predicted value and the label value; To prevent extremely small constants with a denominator of 0.
[0016] Furthermore, the hotspot priority loss term includes at least one of a hotspot overprediction penalty term and a peak intensity alignment term; The over-prediction penalty term for hotspots is used to apply a weighted penalty to high-stress areas where the predicted value is greater than the actual value, and is expressed as follows: in: and The first Each sample at pixel coordinates The predicted value and label value; B is the batch size; and The height and width of the heatmap space; To base on the actual heatmap label values Dynamically calculated pixel weights; and The weighting coefficient for hotspots; The peak intensity alignment term is used to constrain the peak intensity of the predicted heatmap and the supervised heatmap to remain consistent, and is expressed as: in, Center front The set of the largest pixels; for Center front The set of the largest pixels, and , The percentage of peak intensity pixels; For smoothing L1 loss function; To predict the average intensity value of the top k largest pixels in the heatmap; This represents the average intensity value of the top k largest pixel regions in the actual heatmap.
[0017] Furthermore, the predicted value The activation mapping is obtained from the original model output and aligned with the normalized supervised heatmap in numerical range, and is represented as follows: Where Z is the original output value of the model.
[0018] Furthermore, the composite loss also includes a leakage penalty term to suppress the outward expansion of the predicted high-response region beyond the high-stress region defined by the monitoring heatmap; the leakage penalty term is calculated through the following steps: Define high-stress zones: in: For the first Each sample in pixels The indicated value at the location; when hour, =1, otherwise 0; This is a binary high-stress zone indicator diagram; This indicates the pixel coordinates of the b-th sample. The label value at the location; This indicates an indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise. High stress threshold; right Allowed region for morphological dilation: in: This represents the allowable region matrix obtained after morphological dilation, used to tolerate slight outward expansion of high-stress regions. Indicates the kernel size or expansion radius of the morphological expansion operation; For morphological dilation operations; Calculate the leak area: in: This refers to the predicted stress response value in the leak area, i.e., the predicted value outside the permissible area; For the first Each sample at pixel coordinates The predicted value; Calculate leakage losses: in: B represents the loss due to leakage; B represents the batch size. and The height and width of the heatmap space; The value is the p-th power of the predicted value for the leaked area.
[0019] This invention also proposes a deep learning-based method for predicting the relative stress thermogram of the crack neighborhood, comprising the following steps: Step 1: Obtain a grayscale image of the cracks in the structure to be evaluated and adjust it to the preset input size; Step 2: Input the crack grayscale image into the crack neighborhood relative stress heat map prediction model obtained according to the training method described above, and output the single-channel prediction result; Step 3: Map the single-channel prediction results to obtain a relative stress thermogram, which is used to determine the high stress concentration area.
[0020] The beneficial effects of this invention are as follows: This invention presents a deep learning-based training method for predicting relative stress heatmaps in the crack neighborhood. Through an end-to-end deep learning architecture, it achieves a fast and accurate mapping from crack images to stress field distributions, yielding significant technical results. (1) Efficiency breakthrough: By directly using crack images as input, the complex and time-consuming finite element modeling and iterative calculation are avoided, and the stress field prediction is realized in real time or near real time. The calculation efficiency is improved by several orders of magnitude, which meets the needs of rapid evaluation on the engineering site. (2) Enhanced accuracy and practicality: The pixel-level regression network and multi-scale feature fusion mechanism are adopted to restore the stress concentration details of key areas such as crack tips while maintaining the global stress distribution trend; the introduced composite loss function specifically improves the fitting accuracy of high stress areas and effectively suppresses background pseudo-response and unreasonable expansion of the predicted area, and the output heat map has strong engineering interpretability. (3) Expanded applicability: The input image is preprocessed by grayscale conversion and normalization, and the pre-trained backbone network is used to improve the robustness of the model to complex real scenes (such as texture and lighting changes), get rid of the dependence on abstract mechanical parameters, and only requires image data, which greatly reduces the threshold of use and the cost of data acquisition. (4) Technological integration and innovation: The end-to-end learning paradigm of computer vision is deeply integrated with the full-field stress analysis task in solid mechanics, opening up a new technological path of "visual perception-mechanical inversion" and providing a new paradigm for intelligent structural health monitoring.
[0021] In summary, this invention has made substantial progress in terms of efficiency, accuracy, practicality, and technological integration. Attached Figure Description
[0022] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a flowchart of the training method for the crack neighborhood relative stress thermogram prediction model based on deep learning according to the present invention. Figure 2 This is the overall architecture diagram of the relative stress thermogram prediction model in the crack neighborhood; Figure 3 A schematic diagram of a pyramid pooling module for hollow spaces; Figure 4 This is a partial display of the original data; Figure 5 This is a partial display of training results; Figure 6 A performance comparison chart of model training results; Figure 7 This is a flowchart of the deep learning-based method for predicting relative stress thermograms in the crack neighborhood according to the present invention. Detailed Implementation
[0023] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.
[0024] Example 1: Training Method for Crack Neighborhood Relative Stress Thermogram Prediction Model Based on Deep Learning like Figure 1 As shown in the figure, the training method for the crack neighborhood relative stress thermogram prediction model based on deep learning in this embodiment includes the following steps.
[0025] S1: Obtain training samples.
[0026] The training samples include structural images of the crack neighborhood and stress field data corresponding to the structural images. In this embodiment, to make the training set more closely resemble actual engineering practice, the training sample set... Generated through finite element numerical analysis, wherein: For crack neighborhood structure images (e.g., grayscale or RGB images containing crack geometry and texture information generated by finite element software); This is the result obtained by finite element analysis and structural image correspondence under the same working condition, and... The corresponding stress field data in space. To improve the model's generalization ability, the training set should cover different crack morphologies, sizes, orientations, and load conditions.
[0027] Specifically, in Example 1, the training set sample establishment includes: establishing a finite element model of a concrete component with cracks; setting material parameters, boundary conditions, and load cases; calculating the stress field distribution in the crack neighborhood; and rendering or exporting the geometric / texture information of the crack neighborhood as a grayscale structural image. Simultaneously, the stress field data corresponding to the image in the spatial coordinate system is exported. This results in a one-to-one correspondence of training data pairs at the sample level.
[0028] S2: Data processing.
[0029] S2.1: Image preprocessing of crack neighborhood structure.
[0030] The crack neighborhood structure image is converted into a crack grayscale image and adjusted to a preset input size. In this embodiment, the preset input size is H×W. Specifically, the method steps for preprocessing the crack neighborhood structure image are as follows.
[0031] 1) Convert the structured image to grayscale to obtain a single-channel grayscale image. ; 2) Resample the grayscale image to a preset size (e.g., using bilinear interpolation) to obtain... .
[0032] Furthermore, since the crack grayscale image is a single-channel image, and the encoder uses a pre-trained feature extraction module and typically requires multi-channel input, it is preferable to convert the single-channel grayscale image into a multi-channel image by channel copying / mapping before inputting it into the encoder, for example: In this embodiment, the number of channels C=3 to match the input of the three-channel pre-trained encoder.
[0033] S2.2: Preprocessing of stress field data The stress field data is preprocessed, including at least outlier handling, spatial alignment, and scale uniformity.
[0034] 1) Outlier handling: Replace non-numerical elements (e.g., NaN, Inf) in the stress field data with preset values or neighborhood statistics to avoid numerical instability during training. Preferably, non-numerical elements are replaced with 0 or statistical values related to the mean / median of their neighborhood.
[0035] 2) Spatial alignment: Based on the coordinate system relationship between stress field data and crack images, spatial alignment is performed... Perform at least one of rotation, flipping, or transposition to align the stress field with the crack grayscale image in spatial location. Preferably, a rotation transformation can be used to correct the difference between the simulation coordinates and the image coordinates.
[0036] 3) Scale unification: The spatially aligned stress field data is resampled to the same resolution as the input image or a preset resolution to obtain... Preferably, bilinear interpolation is used for resampling to ensure the continuity of the stress field and the smoothness of the numerical values.
[0037] Normalization is performed on the preprocessed stress field data to obtain a normalized relative stress thermogram for supervised learning, such as... Figure 4 As shown. In this embodiment, the method for normalizing the stress field data is as follows: Mini-maximum normalization is performed on the stress field data of a single sample, so that the stress value range of each sample is mapped to the [0,1] interval, expressed as: in: This is the result after normalization; This is the preprocessed stress field data; and These are the minimum and maximum values within the sample, respectively. This represents a matrix consisting entirely of zeros.
[0038] This embodiment enables the stress value range of samples under different working conditions to be uniformly mapped to [0,1], which is beneficial to training stability and comparability between samples.
[0039] S3: Model Training S3.1: Input the grayscale image of the crack into an end-to-end pixel-level regression network and output a single-channel prediction result.
[0040] like Figure 2As shown, the end-to-end pixel-level regression network includes an encoder, a bridging multi-scale context fusion module, and a decoder. Specifically, the encoder performs downsampling extraction during operation; the downsampling result from the encoder is fused at multiple scales using a dilated spatial pyramid pooling module, and then the features of each branch are concatenated along the channel dimension and fused through 1×1 convolution to obtain a bridging result, which is then input into the decoder; the decoder upsamples the bridging result and concatenates the result of each upsampling with the output features of the encoder at the same scale along the channel dimension, and then fuses them through a convolution module, and then introduces ConvNeXtBlock to refine the residuals of the fused features, improving the local structural expressiveness and output smoothness; a final module is set at the end of the decoder to perform a second upsampling of the final upsampling result from the decoder, and further recover high-resolution details through convolution and residual refinement.
[0041] Specifically, the encoder is used to extract multi-scale features and form a hierarchical pyramid representation. In this embodiment, the encoder uses a compact variant of the ConvNeXt V2 series as the backbone network, and its initialization parameters are obtained by self-supervised pre-training based on a mask autoencoder. Specifically, the encoder of the compact variant of the ConvNeXt V2 series includes GRN layers; the encoder's initialization parameters are obtained by self-supervised pre-training of FCMAE (or by self-supervised pre-training based on a mask autoencoder). The GRN layers are used to enhance inter-channel feature competition and prevent the "feature collapse" phenomenon; the fully convolutional mask autoencoder (FCMAE) self-supervised pre-training framework is used to pre-train the encoder on a large dataset to prevent overfitting on small datasets.
[0042] In this embodiment, the encoder employs a convolutional backbone network with multi-scale feature extraction capabilities, outputting a multi-scale feature set {F0, F1, F2, F3}. F0 has high spatial resolution, used to preserve crack edges and detailed textures; F3 has lower resolution but a larger receptive field, used to characterize the overall structure and stress distribution trend of the crack neighborhood. The encoder can be loaded with pre-trained weights to improve generalization ability and convergence speed. In this embodiment, a bridging multi-scale context fusion module is set between the deepest feature F3 of the encoder and the decoder to fuse multi-scale contextual information.
[0043] Specifically, the bridging multi-scale context fusion module is used to perform multi-scale fusion of the features output by the encoder. In this embodiment, the bridging multi-scale context fusion module adopts a dilated spatial pyramid pooling structure, includes multi-branch feature transformation, and concatenates the features of each branch in the channel dimension before fusing them through 1×1 convolution to obtain the bridging result. Figure 3As shown, the bridging multi-scale context fusion module uses dilated spatial pyramid pooling (ASPP), which includes: (1) a 1×1 convolution branch; (2) dilated convolution branches with different dilation rates; and (3) a global pooling branch. The outputs of each branch are concatenated and then fused by convolution to obtain the bridging features. The bridging multi-scale context fusion module expands the effective receptive field on low-resolution semantic features, which is beneficial to improving the overall modeling ability of stress strip morphology and hotspot regions.
[0044] Specifically, the decoder upsamples the output of the bridging multi-scale context fusion module at each level, then concatenates the upsampled results with the encoder's output features at the same scale in the channel dimension. This concatenation is then fused using a convolution module, and a residual refinement module is introduced to enhance the local structural representation. A final upsampling module is set at the end of the decoder to further upsample and refine the decoding results to restore spatial resolution. Multi-scale feature fusion is achieved through skip connections with the encoder's features at the same scale, ultimately outputting the single-channel prediction result.
[0045] In this embodiment, the method steps for the decoder to finally output the single-channel prediction result are as follows: (1) The upsampled features are concatenated with the encoder features of the corresponding scale in the channel dimension; (2) Extract fusion features through the convolutional fusion module; (3) Further enhance local expressive power and improve output smoothness through residual refinement module; The final result is a high-resolution feature of the same size as the input image or that can be aligned to the same size, and a single-channel logits is output.
[0046] If there is a difference between the output resolution and the input resolution at the decoding end, interpolation can be used to align it to the input size H×W to ensure that the output heatmap corresponds to the input crack image at the pixel level. Some training results are as follows: Figure 5 As shown.
[0047] S3.2: Calculate the composite loss based on the normalized relative stress heatmap and the single-channel prediction results, and iteratively update the network parameters to complete the training.
[0048] Based on predicted heatmap With monitoring heatmap Calculate composite loss and network parameters Iterative updates are performed to complete training. The composite loss includes at least a global regression loss term and a hotspot priority loss term, and may optionally include background suppression, leakage penalty, and strip geometry constraint terms to improve engineering usability and interpretability.
[0049] (1) Global regression loss term The global regression loss term is used to constrain the overall prediction error and includes at least: mean squared error (MSE), mean absolute error (L1), and soft Dice loss (which suppresses overall bias and enhances structural consistency), calculated using the following formula.
[0050] Mean square error is expressed as: The mean absolute error is expressed as: The soft Dice loss is expressed as: in: This refers to the batch size (the number of samples per iteration). and The height and width of the heatmap (or input image) spatial dimensions; and The first Each sample at pixel coordinates The predicted value and the label value; To prevent extremely small constants with a denominator of 0.
[0051] (2) Hotspot priority loss items The hotspot priority loss term is used to impose stronger constraints on high-stress areas in the supervised heatmap, thereby improving the fitting accuracy of stress concentration areas such as crack tips. The hotspot priority loss term includes at least: hotspot overprediction penalty and peak intensity alignment (top-k), calculated using the following formula.
[0052] The over-prediction penalty term for hotspots is used to apply a weighted penalty to high-stress areas where the predicted value is greater than the actual value, and is expressed as follows: in: and The first Each sample at pixel coordinates The predicted value and label value; B is the batch size; and The height and width of the heatmap space; To base on the actual heatmap label values Dynamically calculated pixel weights; and The weighting coefficient for hotspots.
[0053] The peak intensity alignment term is used to constrain the peak intensity of the predicted heatmap and the supervised heatmap to remain consistent, and is expressed as: in, Center front The set of the largest pixels; for Center front The set of the largest pixels, and , The percentage of peak intensity pixels; For smoothing L1 loss function; To predict the average intensity value of the top k largest pixels in the heatmap; This represents the average intensity value of the top k largest pixel regions in the actual heatmap.
[0054] Preferably, the composite loss may include a background suppression term, a leakage penalty term, and a boundary background constraint term, which are used to suppress false highlighting in low-stress background areas, limit the outward expansion of high-response areas, and prevent the model from producing undesirable brightening in image boundary areas.
[0055] In this embodiment, a leakage penalty term is introduced into the composite loss to penalize the expansion of the predicted high-response region beyond the supervised allowable region. The allowable region can be obtained based on the supervised heatmap threshold and can be further constructed using morphological expansion to balance spatial micro-bias and expansion suppression. Specifically, the leakage penalty term is calculated through the following steps.
[0056] Define high-stress zones: in: For the first Each sample in pixels The indicated value at the location; when hour, =1, otherwise 0; This is a binary high-stress zone indicator diagram; This indicates the pixel coordinates of the b-th sample. The label value at the location; This indicates an indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise. High stress threshold; right Allowed region for morphological dilation: in: This represents the allowable region matrix obtained after morphological dilation, used to tolerate slight outward expansion of high-stress regions. Indicates the kernel size or expansion radius of the morphological expansion operation; This is a morphological dilation operation.
[0057] Calculate the leak area: in: This refers to the predicted stress response value in the leak area, i.e., the predicted value outside the permissible area; For the first Each sample at pixel coordinates The predicted value.
[0058] Calculate leakage losses: in: B represents the loss due to leakage; B represents the batch size. and The height and width of the heatmap space; The value is the p-th power of the predicted value for the leaked area.
[0059] In a preferred embodiment of this example, the composite loss may further include a background suppression term, a leakage penalty term, and a boundary background constraint term, which are used to suppress false highlighting in low-stress background areas, limit the outward expansion of high-response areas, and prevent the model from generating undesirable brightening in image boundary areas.
[0060] In this embodiment, the predicted value The activation mapping is obtained from the original model output and aligned with the normalized supervised heatmap in numerical range, and is represented as follows: Where Z is the original output value of the model.
[0061] In this embodiment, to verify the effectiveness of the trained prediction model, the error and structural consistency between the predicted stress heatmap and the normalized relative stress heatmap can be evaluated on the validation dataset, and the following steps can be performed: Figure 6 The performance comparisons shown are as follows. Specifically, the evaluation metrics used include: root mean square error (RMSE), mean absolute error (MAE), and structural similarity index (SSIM).
[0062] Let the supervised heatmap of the k-th sample in the validation set be... The predicted heat map is Flatten it to a length of After the vector, we have: (1) RMSE: (2) MAE: (3) SSIM is used to measure the consistency of the predicted heatmap and the label heatmap in terms of brightness, contrast and structural information. Its calculation can be done using the above structural similarity index formula, where the pixel dynamic range L is preferably 1 (corresponding to normalization to [0,1]).
[0063] It should be understood that the evaluation indicators and visualization verification are used to illustrate the feasibility and effectiveness of the technical solution of the present invention, and do not constitute a limitation on the scope of protection of the present invention; those skilled in the art can also use other error measures or structural similarity indicators for equivalent verification.
[0064] like Figure 6 The data shows that this model has a significant performance advantage compared to existing network models (such as U-Net, Res-U-Net, and ConvNeXt-U Net). Specifically, this model exhibits lower errors in both RMSE and MAE metrics, at 0.0865 and 0.0507 respectively, outperforming U-Net (0.110 and 0.085) and Res-U-Net (0.095 and 0.072). This indicates that this model deviates less from the true values during prediction and achieves higher accuracy.
[0065] Furthermore, this model achieved a SSIM score of 0.8673, significantly higher than U-Net's 0.78 and close to ConvNeXt-U Net (no pretrain)'s 0.83, further demonstrating the model's superiority in preserving the structure and capturing details of the stress heatmap.
[0066] It is worth noting that although this model has a similar number of parameters to other models (28M parameters), its lower error and higher SSIM indicate that this model can effectively reduce the computational burden while ensuring high performance, and has better engineering application value.
[0067] Compared with the prior art, this embodiment has the following advantages: (1) By using the crack neighborhood image as input and employing an end-to-end pixel-level regression network to output a relative stress heat map, the high stress concentration area in the crack neighborhood can be quickly located, reducing the dependence on the complex finite element solution process, thereby reducing computational costs and improving engineering evaluation efficiency. (2) By performing outlier processing, spatial alignment and scale unification on the supervised stress field and normalizing it, the training labels and input images are aligned in spatial position and numerical range, which helps to improve training stability and prediction consistency. (3) By adopting an encoder with multi-scale feature extraction capability, and combining a multi-scale context fusion bridging module with a step-by-step upsampling decoder, the spatial details of key areas such as crack tips are restored while maintaining global semantic consistency, thereby improving the fitting accuracy of stress concentration hotspot areas. (4) By introducing a composite loss that includes global regression and hotspot priority, and further combining background suppression, leakage penalty and boundary constraints, the risk of false background highlighting and hotspot expansion can be reduced, thereby improving the engineering usability and interpretability of the output stress heat map. In summary, this embodiment achieves rapid and high-precision prediction of stress heatmaps by aligning and normalizing crack neighborhood images and relative stress field data, and by employing an end-to-end pixel-level regression network of multi-scale encoder-bridge fusion-decoder and hotspot-priority composite loss. This reduces finite element dependence, suppresses false highlights and extrapolation, and improves engineering evaluation efficiency and interpretability.
[0068] Example 2: Deep Learning-Based Relative Stress Thermographic Map Prediction Method for Crack Neighborhood like Figure 7 As shown in the figure, the crack neighborhood relative stress thermogram prediction method based on deep learning in this embodiment includes the following steps.
[0069] Step 1: Acquire and preprocess the image to be tested.
[0070] Images of cracked areas on the surface of the concrete structure to be evaluated are acquired, the images are preprocessed, and adjusted to a preset input size.
[0071] Step 2: Forward prediction of the model.
[0072] The crack region image is input into a trained end-to-end pixel-level regression network. The network automatically performs feature extraction, multi-scale fusion, and upsampling reconstruction, and finally outputs a single-channel predicted logits. After Sigmoid activation, a normalized relative stress heatmap is obtained, which is the single-channel prediction result.
[0073] Step 3: Result Analysis and Application.
[0074] The single-channel prediction results are mapped to obtain a relative stress heatmap, which is used to identify high stress concentration areas. The predicted heatmap can be directly used for visualization; brighter colors indicate higher relative stress. High stress concentration areas can be automatically extracted by setting a threshold, or it can be overlaid with the original image for display. This result provides engineers with an intuitive and quantitative reference for judging the severity of cracks and determining reinforcement priorities, significantly improving assessment efficiency.
[0075] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.
Claims
1. A method for training a deep learning-based prediction model of relative stress thermograms in the crack neighborhood, characterized in that: include: S1: Obtain training samples: The training samples include crack neighborhood structure images and stress field data corresponding to the structure images; S2: Data Processing The crack neighborhood structure image is converted into a crack grayscale image and adjusted to a preset input size; The stress field data is preprocessed, including at least outlier handling, spatial alignment, and scale unification; the preprocessed stress field data is then normalized to obtain a normalized relative stress heatmap for supervised learning. S3: Model Training The grayscale image of the crack is input into an end-to-end pixel-level regression network, and a single-channel prediction result is output. The composite loss is calculated based on the normalized relative stress heatmap and the single-channel prediction results, and the network parameters are iteratively updated to complete the training.
2. The training method for the crack neighborhood relative stress thermogram prediction model based on deep learning according to claim 1, characterized in that: The method for normalizing stress field data is as follows: Mini-maximum normalization is performed on the stress field data of a single sample, mapping the stress value range of each sample to the [0,1] interval, expressed as: in: This is the result after normalization; This is the preprocessed stress field data; and These are the minimum and maximum values within the sample, respectively. This represents a matrix consisting entirely of zeros.
3. The training method for the crack neighborhood relative stress thermogram prediction model based on deep learning according to claim 1, characterized in that: The end-to-end pixel-level regression network includes an encoder, a bridging multi-scale context fusion module, and a decoder; The encoder is used to extract multi-scale features and form a hierarchical pyramid representation; The bridging multi-scale context fusion module is used to perform multi-scale fusion on the features output by the encoder. The decoder is used to upsample the output of the bridging multi-scale context fusion module step by step to restore the spatial resolution, and to achieve multi-scale feature fusion through skip connections of features of the same scale as the encoder, and finally output the single-channel prediction result.
4. The training method for the crack neighborhood relative stress thermogram prediction model based on deep learning according to claim 3, characterized in that: The encoder uses a compact variant of the ConvNeXt V2 series as its backbone network, and its initialization parameters are obtained by self-supervised pre-training based on a mask autoencoder. The bridging multi-scale context fusion module adopts a hollow spatial pyramid pooling structure, which includes multi-branch feature transformation. The features of each branch are concatenated in the channel dimension and then fused through 1×1 convolution to obtain the bridging result.
5. The training method for the crack neighborhood relative stress thermogram prediction model based on deep learning according to claim 1, characterized in that: The composite loss includes at least a global regression loss term and a hotspot priority loss term; the global regression loss term is used to constrain the overall prediction error; the hotspot priority loss term is used to assign higher weights to the high-stress regions of the normalized relative stress heatmap.
6. The training method for the crack neighborhood relative stress thermogram prediction model based on deep learning according to claim 5, characterized in that: The global regression loss term includes at least one of mean squared error, mean absolute error, and Soft Dice loss; Mean square error is expressed as: The mean absolute error is expressed as: The soft Dice loss is expressed as: in: Batch size; and The height and width of the heatmap space; and The first Each sample at pixel coordinates The predicted value and the label value; To prevent extremely small constants with a denominator of 0.
7. The training method for the crack neighborhood relative stress thermogram prediction model based on deep learning according to claim 5, characterized in that: The hotspot priority loss term includes at least one of a hotspot overprediction penalty term and a peak intensity alignment term; The hotspot over-prediction penalty term is used to apply a weighted penalty to high-stress areas where the predicted value is greater than the actual value, and is expressed as follows: in: and The first Each sample at pixel coordinates The predicted value and label value; B is the batch size; and The height and width of the heatmap space; To base on the actual heatmap label values Dynamically calculated pixel weights; and The weighting coefficient for hotspots; The peak intensity alignment term is used to constrain the peak intensity of the predicted heatmap and the supervised heatmap to remain consistent, and is expressed as: in, Center front The set of the largest pixels; for Center front The set of the largest pixels, and , The percentage of peak intensity pixels; For smoothing L1 loss function; To predict the average intensity value of the top k largest pixels in the heatmap; This represents the average intensity value of the top k largest pixel regions in the actual heatmap.
8. The training method for the crack neighborhood relative stress thermogram prediction model based on deep learning according to claim 5, characterized in that: Predicted value The activation mapping is obtained from the original model output and aligned with the normalized supervised heatmap in numerical range, and is represented as follows: Where Z is the original output value of the model.
9. The method for training a deep learning-based relative stress thermogram prediction model for crack neighborhoods according to any one of claims 1, 5-8, characterized in that: The composite loss also includes a background suppression term, a leakage penalty term, and a boundary background constraint term; the leakage penalty term is used to suppress the expansion of the predicted high-response region beyond the high-stress region defined by the supervised heatmap; the leakage penalty term is calculated through the following steps: Define high-stress zones: in: For the first Each sample in pixels The indicated value at the location; when hour, =1, otherwise 0; This is a binary high-stress zone indicator diagram; This indicates the pixel coordinates of the b-th sample. The label value at the location; This indicates an indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise. High stress threshold; right Allowed region for morphological dilation: in: This represents the allowable region matrix obtained after morphological dilation, used to tolerate slight outward expansion of high-stress regions. Indicates the kernel size or expansion radius of the morphological expansion operation; For morphological dilation operations; Calculate the leak area: in: This refers to the predicted stress response value in the leak area, i.e., the predicted value outside the permissible area; For the first Each sample at pixel coordinates The predicted value; Calculate the leakage loss: in: B represents the loss due to leakage; B represents the batch size. and The height and width of the heatmap space; The value is the p-th power of the predicted value for the leaked area.
10. A method for predicting relative stress thermal maps in the crack neighborhood based on deep learning, characterized in that: Includes the following steps: Step 1: Obtain a grayscale image of the cracks in the structure to be evaluated and adjust it to the preset input size; Step 2: Input the crack grayscale image into the crack neighborhood relative stress heat map prediction model obtained according to the training method described in any one of claims 1-9, and output the single-channel prediction result; Step 3: Map the single-channel prediction results to obtain a relative stress thermogram, which is used to determine the high stress concentration area.