Multi-scale adaptive cable fire temperature identification method based on microstructure evolution
By constructing a multi-scale adaptive intelligent identification model based on microscopic electron microscope images, the stability and user threshold issues of temperature assessment after cable fires were solved, and high-precision, low-interference identification of cable fire temperature was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH
- Filing Date
- 2026-04-30
- Publication Date
- 2026-06-26
AI Technical Summary
Existing methods for assessing temperature after cable fires are susceptible to external interference, struggle to maintain stability across the entire temperature range, and have high user barriers, failing to automatically match microstructural changes in different materials.
By using microscopic electron microscope scan images of each layer of materials in the cable protection system, a multi-scale adaptive intelligent identification model is constructed. Through local-to-global collaborative feature extraction, material-guided adaptive fusion, and temperature zone adaptive criterion switching, high-precision identification of material category and overfire temperature is achieved.
It improves identification accuracy and anti-interference ability, lowers the user threshold, achieves high robustness identification and automatic material identification in the entire temperature range, and has good generalization ability.
Smart Images
Figure CN122289248A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary fields of bridge engineering and artificial intelligence, and in particular to a multi-scale adaptive cable fire temperature identification method based on microstructure evolution. Background Technology
[0002] Cables are core load-bearing components of long-span bridges such as cable-stayed bridges and suspension bridges. Their damage state and remaining load-bearing capacity after a fire directly affect the overall safety of the bridge structure. Therefore, quickly and accurately identifying the fire temperature experienced by the cables after a fire is a key technical issue in bridge fire damage assessment.
[0003] Existing methods for assessing temperature after cable fires mainly include sensor monitoring, numerical inversion, and image recognition-based methods. Sensor monitoring requires pre-embedded components, making construction complex and unsuitable for rapid post-disaster assessment; numerical inversion is highly dependent on fire boundary conditions and material parameters, making accurate acquisition difficult in practical engineering. In recent years, artificial intelligence methods have been gradually applied to this field, among which image recognition-based methods are characterized by non-contact operation and high processing efficiency. For example, Chinese patent application CN121010870B discloses a rapid, non-destructive prediction method for the maximum temperature of cable fires based on deep learning, which achieves temperature prediction by acquiring an apparent image of the cable protection system. However, this method still has the following shortcomings:
[0004] First, the apparent image is easily affected by factors such as smoke adhesion, flame blackening, and surface contamination, making it difficult to stably characterize the true evolution state of the material at high temperatures and affecting the accuracy of identification.
[0005] Secondly, different protective materials exhibit varying sensitivities to microstructural changes across different temperature ranges: some materials show clear evolutionary characteristics at low temperatures, while others show insignificant changes at low temperatures and only develop clear criteria at high temperatures. Existing methods lack a mechanism based on dynamic temperature-range switching criteria, making it difficult to guarantee identification stability across the entire temperature range.
[0006] Third, existing models are mainly designed for appearance images. The algorithms cannot automatically match and identify the predicted target with the required materials. Users need to have basic professional knowledge of bridge protection systems and actively provide the algorithm with the correct appearance images of the protective layer materials that need to be identified, which greatly raises the user threshold. In addition, it is difficult to take into account the feature differences of different materials and different observation scales.
[0007] Under the influence of fire, the multi-layered protective materials in cable protection systems undergo not only surface changes but also internal evolution at the microstructural level, including melting, cracking, collapse, and shrinkage. These microstructural changes can reveal the fire temperature history experienced by the material from the perspective of its temperature evolution mechanism, while effectively avoiding interference from external factors such as blackening and dust contamination in the surface images. Therefore, it is necessary to propose a cable fire temperature identification method that uses microscopic electron microscopy images as the identification basis, integrates the microscopic evolution characteristics of multiple material sources, and possesses intelligent material type identification and temperature zone adaptive identification capabilities to improve the accuracy, reliability, and anti-interference ability of the identification. Summary of the Invention
[0008] To address the aforementioned technical shortcomings, the present invention aims to provide a multi-scale adaptive cable fire temperature identification method based on microstructural evolution. This method uses microscopic electron microscopy images of each layer of the cable protection system as the identification basis. By integrating the microscopic evolution characteristics of multiple materials in different temperature ranges, and constructing a multi-scale adaptive intelligent identification model with intelligent material type identification and temperature range adaptive criterion switching capabilities, it achieves high-precision and robust identification of cable fire temperature, effectively avoiding interference from external factors such as smoke on the appearance image identification results.
[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0010] This invention provides a multi-scale adaptive cable fire temperature identification method based on microstructure evolution, comprising the following steps:
[0011] S1. Obtain microscopic electron microscopy images of each layer of the cable protection system at different fire temperatures, and establish the correspondence between microscopic image features and fire temperatures;
[0012] S2. Perform differentiated preprocessing on the micro-electron microscope scan images for different material properties to construct a dataset containing material category labels and temperature labels;
[0013] S3. Construct a multi-scale adaptive intelligent identification model based on MaxVit-Swin, and adopt a local-global collaborative feature extraction, material-guided adaptive fusion and temperature zone adaptive criterion switching mechanism to simultaneously realize material category identification and overheating temperature identification.
[0014] S4. The training and optimization of the multi-scale adaptive intelligent identification model are completed by adopting a transfer learning strategy and a multi-objective collaborative constraint function;
[0015] S5. Verify the effectiveness of the trained multi-scale adaptive intelligent identification model and evaluate its robustness and generalization ability in material category identification and overheating temperature determination under different material types, different observation scales and different imaging regions.
[0016] S6. Design a GUI interactive interface to integrate a pre-trained multi-scale adaptive intelligent identification model, enabling batch import of images to be tested and visualization of temperature and material category identification output.
[0017] Preferably, the cable protection system comprises different materials including PVF wrapping tape, aerogel felt, and PE sheath, and the differentiated pretreatment for the properties of different materials specifically includes:
[0018] For the microscopic images of PVF wrapping tape, pore feature enhancement processing is used to highlight the micropore evolution characteristics; for the microscopic images of aerogel felt, skeleton structure enhancement processing is used to quantify the degree of structural degradation; for the microscopic images of PE sheath, melt texture enhancement processing is used to characterize the phase transition evolution process.
[0019] The micro-electron microscope scanned images that have undergone differential preprocessing are uniformly subjected to scale normalization and grayscale standardization to form the input tensor;
[0020] The processed images are bound to corresponding material category labels and temperature labels, and training sets, validation sets, and test sets are constructed in layers.
[0021] Preferably, the multi-scale adaptive intelligent identification model includes an encoding module for extracting microstructural response characterization, a positioning module for screening micro-information-rich regions, a local-global collaborative characterization module for extracting global tissue state features and local damage state features, a material-guided adaptive fusion module for fusing features and identifying material categories, a temperature zone adaptive criterion switching module for dynamically switching the dominant temperature identification material according to a preset high and low temperature boundary threshold, and a temperature identification module for outputting the overfire temperature identification result.
[0022] Preferably, the local-global collaborative characterization module adopts a dual-backbone extraction structure, including a global characterization branch and a local characterization branch; wherein, the global characterization branch is used to extract global organizational state characteristics of the overall organizational continuity, structural integrity and macroscopic distribution relationship of the cable protection system material, and the local characterization branch is used to extract local damage state characteristics reflecting pore evolution, fiber breakage, network collapse, melt flow marks, boundary ambiguity and residue distribution.
[0023] Preferably, the material-guided adaptive fusion module is used to fuse and map global tissue state features and local damage state features to form a multi-scale joint representation, and output a material category probability vector and material feature fusion weights.
[0024] Preferably, the temperature zone adaptive criterion switching module is used to receive the multi-scale joint characterization and dynamically switch the dominant material basis for temperature identification according to the preset high and low temperature boundary threshold. In the low temperature zone, the micromorphological evolution of PVF wrapping tape and PE sheath is the dominant criterion, and in the high temperature zone, the degradation of the micro-skeleton and network structure of aerogel felt is the dominant criterion.
[0025] Preferably, the temperature identification module is used to fuse the multi-scale joint characterization, the material feature fusion weights, and the temperature zone adaptive criterion switching results to output the final overheating temperature identification result.
[0026] Preferably, step S4 specifically includes:
[0027] S41. Use a transfer learning strategy to initialize the pre-training weights of the multi-scale adaptive intelligent identification model, and input training samples to obtain material category identification results, temperature zone adaptive criterion switching results, and overfire temperature identification results.
[0028] S42. Construct a material identification loss term based on material category supervision information, construct a high-low temperature boundary discrimination loss term based on high-low temperature boundary supervision information, construct an ordered temperature identification loss term based on temperature supervision information, and combine a multi-scale characterization consistency constraint term and a temperature distribution constraint loss term to construct a multi-objective collaborative constraint function to optimize and solve the parameters of the multi-scale adaptive intelligent identification model.
[0029] S43. By using validation samples, the connection weights and gating parameters of the material-guided adaptive fusion module and the boundary threshold of the temperature-zone adaptive criterion switching module are adjusted to improve the recognition stability and cross-scene adaptability of the multi-scale adaptive intelligent identification model under different material types, different observation scales and different imaging regions.
[0030] Preferably, in step S42, the multi-objective collaborative constraint function is in a weighted combination form, specifically expressed as follows:
[0031] ;
[0032] in, This represents the total loss function used to optimize the parameters of the multi-scale adaptive intelligent identification model. This is the material identification loss term, used to supervise the accuracy of the multi-scale adaptive intelligent identification model in predicting the material category (PVF wrapping tape, aerogel felt, or PE sheath) of the input microscopic image. This is a loss term for high and low temperature boundary discrimination, used to supervise the multi-scale adaptive intelligent identification model in binary classification to determine whether the temperature corresponding to the input image is higher than a preset boundary threshold (e.g., 500℃). The ordered temperature identification loss term is used to supervise the multi-scale adaptive intelligent identification model's ordered regression prediction of overfire temperature while maintaining the ordinal relationship between temperature levels. This is a temperature distribution constraint loss term, used to constrain the cumulative distribution difference between the predicted and actual temperature distributions. This is a multi-scale representation consistency constraint term, used to constrain the semantic consistency between local features and global features on the same sample. and The weight coefficients for the corresponding loss terms are adjusted using validation set performance to balance the contribution of each loss term to the optimization of the multi-scale adaptive intelligent identification model.
[0033] The material identification loss term employs a multi-class cross-entropy loss function to quantify the difference between the probability distribution of the material category prediction by the multi-scale adaptive intelligent identification model and the actual material label. Its expression is as follows:
[0034] ;
[0035] in, The total number of training samples, This represents the total number of material categories, corresponding to three types: PVF wrapping tape, aerogel felt, and PE sheath. For the first The actual material label of the sample and if the sample belongs to the first... For materials of the same type, use 1; otherwise, use 0. The multi-scale adaptive intelligent identification model predicts that this sample belongs to the first... The probability of a material being classified as a class of materials.
[0036] The high and low temperature boundary discrimination loss term adopts a binary classification cross-entropy loss function, which is used to supervise the discrimination accuracy of the multi-scale adaptive intelligent identification model in determining whether the temperature corresponding to the input image is higher than the preset boundary threshold. Its expression is:
[0037] ;
[0038] in, For the first The true high and low temperature labels of each sample are set, with a value of 1 if the sample temperature is higher than a preset threshold, and 0 otherwise. This is the probability that the temperature of the sample is higher than the threshold value, predicted by the multi-scale adaptive intelligent identification model.
[0039] The ordered temperature identification loss term adopts the cumulative logistic regression loss function, through... Each binary classification task learns the cumulative probability of each temperature threshold to maintain the order between temperature levels, and its expression is:
[0040] ;
[0041] in, The total number of temperature rating levels. For the first The sample at the th The cumulative label at the temperature threshold and if the actual temperature of the sample is greater than or equal to the threshold... If a threshold is reached, the value is 1; otherwise, it is 0. Let be the log-probability of the k-th threshold output by the multi-scale adaptive intelligent identification model. for Activation function.
[0042] The temperature distribution constraint loss term, based on the squared error of the cumulative distribution function, is used to constrain the global similarity between the temperature probability distribution predicted by the multi-scale adaptive intelligent identification model and the true temperature distribution, thereby enhancing the multi-scale adaptive intelligent identification model's ability to characterize the temperature distribution. Its expression is as follows:
[0043] ;
[0044] in, For the first The true temperature cumulative distribution function of the sample on the th... The value at each temperature level This represents the cumulative distribution function predicted by the multi-scale adaptive intelligent identification model.
[0045] The multi-scale representation consistency constraint term is used to constrain the semantic consistency between local features and global features on the same sample, so as to promote the mutual verification and collaborative expression of the microstructural information of the material by the features extracted by the multi-scale adaptive intelligent identification model at different observation scales. Its expression is:
[0046] ;
[0047] in, The first The local feature vector and global feature vector of each sample.
[0048] Through the weighted collaborative optimization of the above five loss terms, the multi-objective collaborative constraint function can guide the multi-scale adaptive intelligent identification model to learn simultaneously in multiple dimensions such as material identification, temperature zone discrimination, ordered temperature regression, distribution consistency and multi-scale feature consistency, so as to achieve high-precision and high-robust identification of cable fire temperature.
[0049] Preferably, in step S5, the effectiveness of the trained multi-scale adaptive intelligent identification model is verified, specifically including:
[0050] S51. Construct a validation sample set containing different material types, different observation scales, and different imaging regions. Input the validation sample set into the trained multi-scale adaptive intelligent identification model to obtain the material category identification results and overheating temperature identification results corresponding to each sample.
[0051] S52. Based on the material category identification results and real material labels, calculate the classification evaluation index of material category identification to evaluate the identification stability and heterogeneous material adaptability of the multi-scale adaptive intelligent identification model under different material type conditions;
[0052] S53. Based on the fire temperature identification results and the actual temperature label, calculate the temperature evaluation index of the temperature identification to evaluate the accuracy of the multi-scale adaptive intelligent identification model under different observation scale conditions;
[0053] S54. For samples in different imaging regions, calculate the classification evaluation index and the temperature evaluation index respectively, and evaluate the generalization ability of the multi-scale adaptive intelligent identification model in different imaging fields by comparing and analyzing the performance differences between different imaging regions.
[0054] S55. Based on the above evaluation results, the effectiveness, robustness and generalization ability of the multi-scale adaptive intelligent identification model in complex real-world application scenarios are verified.
[0055] Preferably, the GUI user interface designed in step S6 specifically includes:
[0056] S61. Construct a main window interface based on a graphical user interface framework. Set up an image input unit, a result display unit, a model calling unit, and a result export unit in the main window interface, and integrate a multi-scale adaptive intelligent identification model that has been trained. The model calling unit includes an identification start control and a parameter setting control, and the result export unit includes a result saving control.
[0057] S62. Import the microscopic electron microscope scan image of the cable protection system under test through the image input unit. The image under test includes a global image containing complete tissue information and a local image containing local damage information. Perform automatic size normalization and format conversion on the imported image to make it meet the input requirements of the multi-scale adaptive intelligent identification model.
[0058] S63. In response to the user's trigger operation on the identification start control in the model calling unit, the trained multi-scale adaptive intelligent identification model is called to perform forward inference on the preprocessed image, automatically perform material phase identification and temperature identification, and output the material category identification result and overheating temperature identification result corresponding to each image.
[0059] S64. The original image and its corresponding material category and overheating temperature are displayed synchronously through the result display unit, and the identification results are exported as a specified format file and saved through the trigger operation of the result save control in the result export unit.
[0060] Beneficial effects:
[0061] 1. Improved accuracy and anti-interference capability: This invention abandons the traditional method relying on surface images and instead uses microscopic electron microscopy images of each layer of the cable protection system as the basis for identification. Microstructure (such as pore evolution, fiber breakage, and melt flow marks) can reveal the fire temperature history experienced by the material from the mechanism of temperature evolution, effectively avoiding interference from external factors such as smoke adhesion, flame blackening, and surface contamination on the surface images, significantly improving the accuracy, reliability, and anti-interference capability of fire temperature identification.
[0062] 2. Achieving High Robustness in Temperature Range Identification: This invention constructs a multi-scale adaptive intelligent identification model with temperature-adaptive criterion switching capability. Addressing the differences in the sensitivity of different protective materials to microstructural changes across different temperature ranges, it can dynamically switch the dominant material basis for temperature identification based on preset high and low temperature boundary thresholds (e.g., in the low-temperature range, the micromorphological evolution of PVF wrapping tape and PE sheath is dominant, while in the high-temperature range, the structural degradation of aerogel felt is dominant). This mechanism overcomes the insufficient applicability of single-material criteria across the entire temperature range, ensuring that the multi-scale adaptive intelligent identification model maintains stable and accurate identification performance throughout the entire temperature range from low to high temperatures.
[0063] 3. Integrating Multi-Scale Microscopic Features to Enhance Characterization Capabilities: This invention designs a dual-backbone extraction structure that combines local and global approaches. The global characterization branch extracts global features such as the overall material structure continuity, while the local characterization branch focuses on local damage features such as pores and fiber fractures. By integrating information from these two scales, the multi-scale adaptive intelligent identification model can more comprehensively and accurately characterize the microstructural evolution of materials under high temperatures, providing a richer and more discriminative feature base for subsequent category recognition and temperature determination.
[0064] 4. Achieving Intelligent Material Identification and Automated Judgment: This invention, through a material-guided adaptive fusion module, enables a multi-scale adaptive intelligent judgment model to automatically identify the protective material category (such as PVF wrapping tape, aerogel felt, or PE sheath) of the input image, and generates targeted feature fusion weights based on the recognition results. This mechanism achieves synchronous identification of the correlation between material type and fire temperature. Users do not need professional knowledge of bridge protection systems to distinguish and input images of specific materials, greatly reducing the usage threshold and realizing end-to-end automated intelligent judgment.
[0065] 5. Enhancing the overall performance of the multi-scale adaptive intelligent identification model through multi-objective collaborative optimization: This invention constructs a multi-objective collaborative constraint function encompassing multiple dimensions, including material identification, temperature zone discrimination, ordered temperature regression, distribution consistency, and multi-scale feature consistency. Through joint optimization, the multi-scale adaptive intelligent identification model can not only accurately identify material categories and determine temperatures, but also maintain the ordinal relationships between temperature levels, constrain the global similarity of predicted temperature distributions, and promote semantic consistency between local and global features. This improves the prediction accuracy, robustness, and physical interpretability of the multi-scale adaptive intelligent identification model across multiple dimensions.
[0066] 6. Excellent generalization ability and engineering applicability: During the model training and verification stages, this invention fully considers factors such as different material types, different observation scales, and different imaging regions, ensuring the robustness and generalization ability of the multi-scale adaptive intelligent identification model in complex real-world application scenarios. Furthermore, through the GUI interface of the multi-scale adaptive intelligent identification model, batch import of test images, one-click identification, and visualized export of results are achieved. The operation is convenient, allowing even non-professional users to quickly identify the post-disaster cable fire temperature, demonstrating high engineering application value. Attached Figure Description
[0067] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.
[0068] Figure 1 A flowchart of a multi-scale adaptive cable fire temperature identification method based on microstructure evolution provided in an embodiment of the present invention;
[0069] Figure 2 This is an architecture diagram of the multi-scale adaptive intelligent identification model based on MaxVit-Swin provided in an embodiment of the present invention;
[0070] Figure 3 This is a diagram of the multi-scale adaptive intelligent identification model architecture of the microstructure feature encoding module provided in this embodiment of the invention.
[0071] Figure 4 This is a diagram illustrating the architecture of a multi-scale adaptive intelligent identification model for a local salient region localization module provided in an embodiment of the present invention.
[0072] Figure 5 This is a diagram of the multi-scale adaptive intelligent identification model architecture of the local-global collaborative representation module provided in this embodiment of the invention.
[0073] Figure 6 This is a diagram illustrating the architecture of a multi-scale adaptive intelligent identification model for a material-guided adaptive fusion module provided in an embodiment of the present invention.
[0074] Figure 7 This is a diagram of the multi-scale adaptive intelligent identification model architecture of the temperature zone adaptive criterion switching module provided in an embodiment of the present invention.
[0075] Figure 8 This is a diagram of the multi-scale adaptive intelligent identification model architecture of the temperature identification module provided in an embodiment of the present invention. Detailed Implementation
[0076] 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.
[0077] like Figure 1 As shown, this embodiment of the invention provides a multi-scale adaptive cable fire temperature identification method based on microstructure evolution, including the following steps:
[0078] S1. Through high-temperature tests on the cable protection system, obtain microscopic electron microscopy images of each layer of the cable protection system at different fire temperatures, and establish the correspondence between microscopic image features and fire temperatures; the different materials of the cable protection system include PVF wrapping tape, aerogel felt, and PE sheath.
[0079] In this embodiment, the test platform disclosed in Chinese invention patent application number 2025111956407, "Real-time Mixed Test Platform and Implementation Method for Vehicle-Fire-Wind Force of Substructure of Long-Span Bridge," is used to conduct fire resistance tests on cable protection materials under controllable parameters of wind, fire, and force coupling. This accurately simulates different fire scenarios and loading conditions, obtains samples of various materials after different fire temperatures, and then performs micro-electron microscopy scanning on these samples to obtain micro-electron microscopy images at different magnifications. This provides high-fidelity and diverse data support for the subsequent training and verification of multi-scale adaptive intelligent identification models.
[0080] S2. Perform differentiated preprocessing on the micro-electron microscope scan images for different material properties to construct a dataset containing material category labels and temperature labels;
[0081] For the microscopic images of PVF wrapping tape, pore feature enhancement processing is adopted. High-frequency filtering is used to highlight the pore edges and enhance the texture information of the pore distribution area, while suppressing the background noise in the non-pore area, so as to highlight the micropore evolution characteristics of PVF wrapping tape near the temperature threshold of 250℃.
[0082] For the microscopic images of aerogel mat, a skeleton structure enhancement process is adopted. Directional filtering is used to enhance the continuity of glass fibers and strengthen the sparsity characteristics of aerogel network pores. At the same time, local anomaly features of fiber break points and network collapse areas are extracted to quantify the degree of structural degradation of aerogel mat in the temperature range above 500℃.
[0083] For the microscopic images of PE sheath, melt texture enhancement processing is adopted. Local variance analysis is used to highlight the melt flow mark area and enhance the gradient features of the blurred boundary area. At the same time, the melt flow direction and residue distribution features are extracted to characterize the phase transformation evolution process of PE sheath in the temperature range of 250℃ to 500℃.
[0084] The micro-electron microscope scan images that have undergone differential preprocessing are uniformly normalized in scale and grayscale to form input tensors of consistent size. Data augmentation strategies are then used to randomly rotate, crop, and perturb the brightness of the training set images to expand the sample diversity.
[0085] The processed images are bound to the corresponding material category labels and temperature labels. Training sets, validation sets, and test sets are constructed hierarchically according to material type and temperature range to ensure that the sample distribution of different material types and different temperature ranges in each data subset is balanced, avoiding data leakage and class imbalance problems.
[0086] S3. Construct a multi-scale adaptive intelligent identification model based on MaxVit-Swin, namely MRLG-MicroTempNet (hereinafter referred to as MRLG-MicroTempNet). It adopts a local-global collaborative feature extraction, material-guided adaptive fusion, and temperature-zone adaptive criterion switching mechanism to simultaneously realize material category identification and overheating temperature identification. In MRLG-MicroTempNet, MRLG represents the combination of four multi-scale adaptive intelligent identification model structures: material-guided (Material-Routed), local-global collaborative (Local-Global), microstructure encoding (Micro), and temperature identification (Temperature).
[0087] See Figure 2 The multi-scale adaptive intelligent identification model includes:
[0088] The microstructure feature encoding module is used to convert the input micro-electron microscope scan image into a structural response characterization suitable for material phase identification and temperature determination;
[0089] A local salient region localization module is used to automatically filter out regions rich in pores, fiber skeletons, melt flow marks and other microscopic information from the structural response characterization.
[0090] The local-global collaborative characterization module adopts a dual-backbone extraction structure, including a global characterization branch and a local characterization branch. The global characterization branch is used to extract global organizational state features of the overall material structure continuity, structural integrity and macroscopic distribution relationship of the cable protection system. The local characterization branch is used to extract local damage state features reflecting pore evolution, fiber breakage, network collapse, melt flow marks, boundary ambiguity and residue distribution.
[0091] The material-guided adaptive fusion module performs feature fusion, outputs material category identification results, and generates corresponding material feature fusion weights.
[0092] The temperature zone adaptive criterion switching module is used to dynamically switch the dominant material basis for temperature identification according to the preset high and low temperature boundary threshold. In the low temperature zone, the micromorphological evolution of PVF wrapping tape and PE sheath is the dominant criterion, and in the high temperature zone, the degradation of the micro-skeleton and network structure of aerogel felt is the dominant criterion.
[0093] The temperature identification module is used to integrate the multi-scale joint characterization, the material feature fusion weights, and the temperature zone adaptive criterion switching results to output the overheating temperature identification result.
[0094] The connection relationships between the modules in MRLG-MicroTempNet are as follows:
[0095] The input microscopic electron microscope scan image is processed sequentially by the microstructure feature encoding module and the local salient region localization module to complete the microstructure response construction and local salient region screening;
[0096] The filtered features are input into the local-global collaborative representation module, which extracts global tissue state features and local damage state features through a dual-backbone extraction structure. The material-guided adaptive fusion module then fuses and maps the global tissue state features and the local damage state features to form a multi-scale joint representation, while outputting the material category identification result.
[0097] The material guidance adaptive fusion module outputs a material category probability vector based on the multi-scale joint characterization, and generates corresponding material feature fusion weights based on the material category probability vector; wherein, the material category probability vector is used to characterize the degree to which the input micro-electron microscope scan image belongs to PVF wrapping tape, aerogel felt or PE sheath, and the material feature fusion weights are used to guide the temperature identification module to perform differentiated temperature identification for the micro-morphological features of different materials.
[0098] The multi-scale joint characterization is simultaneously input into the temperature zone adaptive criterion switching module and the temperature identification module, and outputs the final overheating temperature identification result.
[0099] The microstructure feature encoding module adopts a multi-level convolutional encoding structure, which includes multiple convolutional units connected in series. Each convolutional unit consists of a convolutional layer, a batch normalization layer, and a non-linear activation layer. The output of the previous convolutional unit serves as the input of the next convolutional unit, and after step-by-step feature extraction, an encoded feature map is output.
[0100] See Figure 3 Before performing convolutional encoding on the input microscopic electron microscope scan image, the microstructure feature encoding module first extracts the grayscale component, gradient magnitude component, and second-order texture component to characterize the brightness distribution, edge changes, and high-order texture response in the material's microstructure, respectively. Then, the grayscale component, gradient magnitude component, and second-order texture component are concatenated along the channel dimension to form a three-channel basic microstructure feature map, which is then input into the channel concatenation and projection mapping structure for convolutional encoding. The channel splicing and projection mapping structure includes a first convolutional unit, a second convolutional unit, and an output mapping layer. The first convolutional unit includes a convolutional layer with 3 input channels, 16 output channels, a 3×3 kernel size, a stride of 1, and padding of 1, a batch normalization layer, and a GELU activation layer. The second convolutional unit includes a convolutional layer with 16 input channels, 16 output channels, a 3×3 kernel size, a stride of 1, and padding of 1, a batch normalization layer, and a GELU activation layer. The output mapping layer is a convolutional layer with 16 input channels, 3 output channels, a 1×1 kernel size, a stride of 1, and padding of 0. It is used to map the microstructural features extracted step by step into an encoded feature map of size 224×224.
[0101] See Figure 4The local salient region localization module includes a salient response generation layer, a local response sorting layer, and a region clipping layer connected in sequence. The encoded feature map is input to the salient response generation layer to generate a local response map. After the local response sorting layer selects multiple candidate regions with high response values, the region clipping layer outputs fixed-size local salient region image patches. Further, the local salient region localization module takes the encoded feature map as input and outputs four fixed-size local salient region image patches; each local image patch has a size of 96×96, and the output tensor shape is [B,4,3,96,96].
[0102] See Figure 5 In the local-global collaborative representation module, the global representation branch extraction unit adopts a MaxVit backbone structure, consisting of a feature embedding layer, multiple global feature extraction units, an adaptive pooling layer, and a linear mapping layer connected in sequence. The encoded feature map is input into the global representation branch and outputs a global feature vector after step-by-step processing. The local representation branch feature extraction unit adopts a SwinTransformer backbone structure, consisting of a local feature embedding layer, multiple window attention feature extraction units, an instance aggregation layer, and a linear mapping layer connected in sequence. The local salient region image patch is input into the local representation branch and outputs a local feature vector after step-by-step processing. Further, the global representation branch receives an encoded feature map of size 224×224 and outputs a global feature vector global_feat, while the local representation branch receives four local salient region image patches of size 96×96 and outputs a local feature vector local_bag_feat. The global feature vector and the local feature vector are concatenated and linearly mapped to obtain a multi-scale joint representation with a dimension of 512.
[0103] See Figure 6 The material-guided adaptive fusion module includes a feature splicing layer, a joint mapping layer, a normalization layer, a gated weighting layer, and a material classification head. The global feature vector and local feature vector are concatenated by the feature splicing layer and then sequentially input into the joint mapping layer and the normalization layer to form a multi-scale joint representation. This multi-scale joint representation is simultaneously input into the gated weighting layer and the material classification head, generating material feature fusion weights and material category probability vectors, respectively. Further, the material-guided adaptive fusion module takes a 512-dimensional multi-scale joint representation as input and obtains a 256-dimensional shared adaptive feature after joint mapping. The material classification head outputs a 3-dimensional material category probability vector [p_PVF, p_AG, p_PE], and the gated weighting layer outputs 3-dimensional material feature fusion weights.
[0104] See Figure 7The temperature zone adaptive criterion switching module includes a high-low temperature boundary discrimination branch, a low temperature criterion branch, and a high temperature criterion branch. The multi-scale joint representation inputs the high-low temperature boundary discrimination branch and outputs the probability that the sample temperature is higher than a preset boundary threshold. The low temperature criterion branch receives the material feature fusion weights and generates low temperature criterion features, and the high temperature criterion branch receives the material category probability vectors and generates high temperature criterion features. Based on the probability, the corresponding temperature zone scheduling weights are generated, and the temperature zone adaptive criterion switching result is output.
[0105] See Figure 8 The temperature identification module includes a shared temperature characterization head, a material-specific temperature characterization head, and an output fusion layer. Both the shared temperature characterization head and the material-specific temperature characterization head are composed of a linear layer, a normalization layer, and an activation layer connected in sequence. The shared temperature characterization head receives the multi-scale joint characterization, and the material-specific temperature characterization head receives the material criterion features after weighted fusion based on the material category probability vector. The outputs of the shared temperature characterization head, the material-specific temperature characterization head, and the temperature zone adaptive criterion switching results are input to the output fusion layer, and the final overheating temperature identification result is output after weighted fusion.
[0106] S4. The training and optimization of the multi-scale adaptive intelligent identification model are completed using a transfer learning strategy and a multi-objective collaborative constraint function; specifically including:
[0107] S41. Use a transfer learning strategy to initialize the pre-training weights of the multi-scale adaptive intelligent identification model, and input training samples to obtain material category identification results, temperature zone adaptive criterion switching results, and overfire temperature identification results.
[0108] S42. Construct a material identification loss term based on material category supervision information, construct a high-low temperature boundary discrimination loss term based on high-low temperature boundary supervision information, construct an ordered temperature identification loss term based on temperature supervision information, and combine a multi-scale characterization consistency constraint term and a temperature distribution constraint loss term to construct a multi-objective collaborative constraint function to optimize and solve the parameters of the multi-scale adaptive intelligent identification model.
[0109] The multi-objective collaborative constraint function is in weighted combination form, specifically expressed as follows:
[0110] ;
[0111] in, This represents the total loss function used to optimize the parameters of the multi-scale adaptive intelligent identification model. This is the material identification loss term, used to supervise the accuracy of the multi-scale adaptive intelligent identification model in predicting the material category (PVF wrapping tape, aerogel felt, or PE sheath) of the input microscopic image. This is a loss term for high and low temperature boundary discrimination, used to supervise the multi-scale adaptive intelligent identification model in binary classification to determine whether the temperature corresponding to the input image is higher than a preset boundary threshold (e.g., 500℃). The ordered temperature identification loss term is used to supervise the multi-scale adaptive intelligent identification model's ordered regression prediction of overfire temperature while maintaining the ordinal relationship between temperature levels. This is a temperature distribution constraint loss term, used to constrain the cumulative distribution difference between the predicted and actual temperature distributions. This is a multi-scale representation consistency constraint term, used to constrain the semantic consistency between local features and global features on the same sample. and The weight coefficients for the corresponding loss terms are adjusted using validation set performance to balance the contribution of each loss term to the optimization of the multi-scale adaptive intelligent identification model.
[0112] The material identification loss term employs a multi-class cross-entropy loss function to quantify the difference between the probability distribution of the material category prediction by the multi-scale adaptive intelligent identification model and the actual material label. Its expression is as follows:
[0113] ;
[0114] in, The total number of training samples, This represents the total number of material categories, corresponding to three types: PVF wrapping tape, aerogel felt, and PE sheath. For the first The actual material label of the sample and if the sample belongs to the first... For materials of the same type, use 1; otherwise, use 0. The multi-scale adaptive intelligent identification model predicts that this sample belongs to the first... The probability of a material being classified as a class of materials.
[0115] The high and low temperature boundary discrimination loss term adopts a binary classification cross-entropy loss function, which is used to supervise the discrimination accuracy of the multi-scale adaptive intelligent identification model in determining whether the temperature corresponding to the input image is higher than the preset boundary threshold. Its expression is:
[0116] ;
[0117] in, For the first The true high and low temperature labels of each sample are set, with a value of 1 if the sample temperature is higher than a preset threshold, and 0 otherwise. This is the probability that the temperature of the sample is higher than the threshold value, predicted by the multi-scale adaptive intelligent identification model.
[0118] The ordered temperature identification loss term adopts the cumulative logistic regression loss function, through... Each binary classification task learns the cumulative probability of each temperature threshold to maintain the order between temperature levels, and its expression is:
[0119] ;
[0120] in, The total number of temperature rating levels. For the first The sample at the th The cumulative label at the temperature threshold and if the actual temperature of the sample is greater than or equal to the threshold... If a threshold is reached, the value is 1; otherwise, it is 0. Let be the log-probability of the k-th threshold output by the multi-scale adaptive intelligent identification model. for Activation function.
[0121] The temperature distribution constraint loss term, based on the squared error of the cumulative distribution function, is used to constrain the global similarity between the temperature probability distribution predicted by the multi-scale adaptive intelligent identification model and the true temperature distribution, thereby enhancing the multi-scale adaptive intelligent identification model's ability to characterize the temperature distribution. Its expression is as follows:
[0122] ;
[0123] in, For the first The true temperature cumulative distribution function of the sample on the th... The value at each temperature level This represents the cumulative distribution function predicted by the multi-scale adaptive intelligent identification model.
[0124] The multi-scale representation consistency constraint term is used to constrain the semantic consistency between local features and global features on the same sample, so as to promote the mutual verification and collaborative expression of the microstructural information of the material by the features extracted by the multi-scale adaptive intelligent identification model at different observation scales. Its expression is:
[0125] ;
[0126] in, The first The local feature vector and global feature vector of each sample.
[0127] Through the weighted collaborative optimization of the above five loss terms, the multi-objective collaborative constraint function can guide the multi-scale adaptive intelligent identification model to learn simultaneously in multiple dimensions such as material identification, temperature zone discrimination, ordered temperature regression, distribution consistency and multi-scale feature consistency, so as to achieve high-precision and high-robust identification of cable fire temperature.
[0128] S43. By using validation samples, the connection weights and gating parameters of the material-guided adaptive fusion module and the boundary threshold of the temperature-zone adaptive criterion switching module are adjusted to improve the recognition stability and cross-scene adaptability of the multi-scale adaptive intelligent identification model under different material types, different observation scales and different imaging regions.
[0129] S5. Validate the effectiveness of the trained multi-scale adaptive intelligent identification model, evaluating its robustness and generalization ability in material category identification and overheating temperature determination under different material types, observation scales, and imaging regions; specifically including:
[0130] S51. Construct a validation sample set containing different material types, different observation scales, and different imaging regions. Input the validation sample set into the trained multi-scale adaptive intelligent identification model to obtain the material category identification results and overheating temperature identification results corresponding to each sample.
[0131] S52. Based on the material category identification results and real material labels, calculate the classification evaluation index of material category identification to evaluate the identification stability and heterogeneous material adaptability of the multi-scale adaptive intelligent identification model under different material type conditions;
[0132] S53. Based on the fire temperature identification results and the actual temperature label, calculate the temperature evaluation index of the temperature identification to evaluate the accuracy of the multi-scale adaptive intelligent identification model under different observation scale conditions;
[0133] S54. For samples in different imaging regions, calculate the classification evaluation index and the temperature evaluation index respectively, and evaluate the generalization ability of the multi-scale adaptive intelligent identification model in different imaging fields by comparing and analyzing the performance differences between different imaging regions.
[0134] S55. Based on the above evaluation results, the effectiveness, robustness and generalization ability of the multi-scale adaptive intelligent identification model in complex real-world application scenarios are verified.
[0135] S6. Design a GUI interface to integrate a pre-trained multi-scale adaptive intelligent identification model, enabling batch import of images to be tested and visual identification output of temperature and material category, specifically including:
[0136] S61. Construct a main window interface based on a graphical user interface framework. Set up an image input unit, a result display unit, a model calling unit, and a result export unit in the main window interface, and integrate a multi-scale adaptive intelligent identification model that has been trained. The model calling unit includes an identification start control and a parameter setting control, and the result export unit includes a result saving control.
[0137] S62. Import the microscopic electron microscope scan image of the cable protection system under test through the image input unit. The image under test includes a global image containing complete tissue information and a local image containing local damage information. Perform automatic size normalization and format conversion on the imported image to make it meet the input requirements of the multi-scale adaptive intelligent identification model.
[0138] S63. In response to the user's trigger operation on the identification start control in the model calling unit, the trained multi-scale adaptive intelligent identification model is called to perform forward inference on the preprocessed image, automatically perform material phase identification and temperature identification, and output the material category identification result and overheating temperature identification result corresponding to each image.
[0139] S64. The original image and its corresponding material category and overheating temperature are displayed synchronously through the result display unit, and the identification results are exported as a specified format file and saved through the trigger operation of the result save control in the result export unit.
[0140] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A multi-scale adaptive cable fire temperature identification method based on microstructure evolution, characterized in that, Includes the following steps: S1. Obtain microscopic electron microscopy images of each layer of the cable protection system at different fire temperatures, and establish the correspondence between microscopic image features and fire temperatures; S2. Perform differentiated preprocessing on the micro-electron microscope scan images for different material properties to construct a dataset containing material category labels and temperature labels; S3. Construct a multi-scale adaptive intelligent identification model based on MaxVit-Swin, and adopt a local-global collaborative feature extraction, material-guided adaptive fusion and temperature zone adaptive criterion switching mechanism to simultaneously realize material category identification and overheating temperature identification. S4. The training and optimization of the multi-scale adaptive intelligent identification model are completed by adopting a transfer learning strategy and a multi-objective collaborative constraint function; S5. Verify the effectiveness of the trained multi-scale adaptive intelligent identification model and evaluate its robustness and generalization ability in material category identification and overheating temperature determination under different material types, different observation scales and different imaging regions. S6. Design a GUI interactive interface to integrate a pre-trained multi-scale adaptive intelligent identification model, enabling batch import of images to be tested and visualization of temperature and material type identification output.
2. The multi-scale adaptive cable fire temperature identification method based on microstructure evolution according to claim 1, characterized in that, The cable protection system uses different materials, including PVF wrapping tape, aerogel felt, and PE sheath. The specific pretreatment for the different material properties is as follows: For the microscopic images of PVF wrapping tape, pore feature enhancement processing is used to highlight the micropore evolution characteristics; for the microscopic images of aerogel felt, skeleton structure enhancement processing is used to quantify the degree of structural degradation. For the microscopic images of the PE sheath, melt texture enhancement processing is used to characterize the phase transition evolution process; The micro-electron microscope scanned images that have undergone differential preprocessing are uniformly subjected to scale normalization and grayscale standardization to form the input tensor; The processed images are bound to corresponding material category labels and temperature labels, and training sets, validation sets, and test sets are constructed in layers.
3. The multi-scale adaptive cable fire temperature identification method based on microstructure evolution according to claim 1, characterized in that, The multi-scale adaptive intelligent identification model includes an encoding module for extracting microstructural response characterization, a positioning module for screening micro-information-rich regions, a local-global collaborative characterization module for extracting global tissue state features and local damage state features, a material-guided adaptive fusion module for fusing features and identifying material categories, a temperature zone adaptive criterion switching module for dynamically switching the dominant temperature identification material according to a preset high and low temperature boundary threshold, and a temperature identification module for outputting the overfire temperature identification result.
4. The multi-scale adaptive cable fire temperature identification method based on microstructure evolution according to claim 3, characterized in that, The local-global collaborative characterization module adopts a dual-backbone extraction structure, including a global characterization branch and a local characterization branch. The global characterization branch is used to extract global organizational state characteristics of the overall material structure continuity, structural integrity, and macroscopic distribution relationship of the cable protection system. The local characterization branch is used to extract local damage state characteristics reflecting pore evolution, fiber breakage, network collapse, melt flow marks, boundary ambiguity, and residue distribution. The material-guided adaptive fusion module is used to fuse and map global tissue state features and local damage state features to form a multi-scale joint representation, and outputs a material category probability vector and material feature fusion weights.
5. The multi-scale adaptive cable fire temperature identification method based on microstructure evolution according to claim 4, characterized in that, The temperature zone adaptive criterion switching module is used to receive the multi-scale joint characterization and dynamically switch the dominant material basis for temperature identification according to the preset high and low temperature boundary threshold. In the low temperature zone, the micromorphological evolution of PVF wrapping tape and PE sheath is the dominant criterion, and in the high temperature zone, the degradation of the micro-skeleton and network structure of aerogel felt is the dominant criterion.
6. The multi-scale adaptive cable fire temperature identification method based on microstructure evolution according to claim 4, characterized in that, The temperature identification module is used to integrate the multi-scale joint characterization, the material feature fusion weights, and the temperature zone adaptive criterion switching results to output the final overheating temperature identification result.
7. The multi-scale adaptive cable fire temperature identification method based on microstructure evolution according to claim 1, characterized in that, Step S4 specifically includes: S41. Use a transfer learning strategy to initialize the pre-training weights of the multi-scale adaptive intelligent identification model, and input training samples to obtain material category identification results, temperature zone adaptive criterion switching results, and overfire temperature identification results. S42. Construct a material identification loss term based on material category supervision information, construct a high-low temperature boundary discrimination loss term based on high-low temperature boundary supervision information, construct an ordered temperature identification loss term based on temperature supervision information, and combine a multi-scale characterization consistency constraint term and a temperature distribution constraint loss term to construct a multi-objective collaborative constraint function to optimize and solve the parameters of the multi-scale adaptive intelligent identification model. S43. By using validation samples, the connection weights and gating parameters of the material-guided adaptive fusion module and the boundary threshold of the temperature-zone adaptive criterion switching module are adjusted to improve the recognition stability and cross-scene adaptability of the multi-scale adaptive intelligent identification model under different material types, different observation scales and different imaging regions.
8. The multi-scale adaptive cable fire temperature identification method based on microstructure evolution according to claim 7, characterized in that, In step S42, the multi-objective collaborative constraint function is in weighted combination form, specifically expressed as follows: ； in, This represents the total loss function used to optimize the parameters of the multi-scale adaptive intelligent identification model. For material identification loss items, To determine the loss term at the high and low temperature boundary, To identify the loss term for ordered temperature, For temperature distribution constraint loss term, This is a multi-scale representation of consistency constraints. and The weight coefficients for the corresponding loss terms are adjusted using validation set performance to balance the contribution of each loss term to the optimization of the multi-scale adaptive intelligent identification model.
9. The multi-scale adaptive cable fire temperature identification method based on microstructure evolution according to claim 1, characterized in that, In step S5, the effectiveness of the trained multi-scale adaptive intelligent identification model is verified, specifically including: S51. Construct a validation sample set containing different material types, different observation scales, and different imaging regions. Input the validation sample set into the trained multi-scale adaptive intelligent identification model to obtain the material category identification results and overheating temperature identification results corresponding to each sample. S52. Based on the material category identification results and real material labels, calculate the classification evaluation index of material category identification to evaluate the identification stability and heterogeneous material adaptability of the multi-scale adaptive intelligent identification model under different material type conditions; S53. Based on the fire temperature identification results and the actual temperature label, calculate the temperature evaluation index of the temperature identification to evaluate the accuracy of the multi-scale adaptive intelligent identification model under different observation scale conditions; S54. For samples in different imaging regions, calculate the classification evaluation index and the temperature evaluation index respectively, and evaluate the generalization ability of the multi-scale adaptive intelligent identification model in different imaging fields by comparing and analyzing the performance differences between different imaging regions. S55. Based on the above evaluation results, the effectiveness, robustness and generalization ability of the multi-scale adaptive intelligent identification model in complex real-world application scenarios are verified.
10. The multi-scale adaptive cable fire temperature identification method based on microstructure evolution according to claim 1, characterized in that, The GUI user interface designed in step S6 specifically includes: S61. Construct a main window interface based on a graphical user interface framework. Set up an image input unit, a result display unit, a model calling unit, and a result export unit in the main window interface, and integrate a multi-scale adaptive intelligent identification model that has been trained. The model calling unit includes an identification start control and a parameter setting control, and the result export unit includes a result saving control. S62. Import the microscopic electron microscope scan image of the cable protection system under test through the image input unit. The image under test includes a global image containing complete tissue information and a local image containing local damage information. Perform automatic size normalization and format conversion on the imported image to make it meet the input requirements of the multi-scale adaptive intelligent identification model. S63. In response to the user's trigger operation on the identification start control in the model calling unit, the trained multi-scale adaptive intelligent identification model is called to perform forward inference on the preprocessed image, automatically perform material phase identification and temperature identification, and output the material category identification result and overheating temperature identification result corresponding to each image. S64. The original image and its corresponding material category and overheating temperature are displayed synchronously through the result display unit, and the identification results are exported as a specified format file and saved through the trigger operation of the result save control in the result export unit.