Rapid lossless prediction method for maximum overfire temperature of cable based on deep learning

Abstract

The invention discloses a fast lossless prediction method for the highest overfire temperature of a cable based on deep learning, and relates to the technical field of safety assessment after a bridge fire. According to the method, apparent image data of a cable anticorrosion-fireproof protection system at different temperatures are obtained through various fire tests, and a special data set is constructed after shunting preprocessing; a ConvNeXt-Tiny-based double-branch weight sharing prediction model is innovatively constructed, state features of a cable flame-retardant sealant layer and an anticorrosive layer are extracted through two parallel branches respectively, deep representation learning is performed through feature fusion and a shared backbone network, and finally high-precision parallel prediction of double temperature parameters is realized through two ordered classification prediction heads. A transfer learning strategy and a mixed loss function are used for training, so that the prediction precision and generalization ability are remarkably improved; finally, a visual window is developed, and cable fire temperature rapid nondestructive evaluation without dismounting the sensor is realized. The method provides reliable technical support for safety performance evaluation of the bridge cable after the fire disaster.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of bridge engineering and artificial intelligence, and in particular to a rapid and non-destructive method for predicting the maximum temperature of cable fires based on deep learning. Background Technology

[0002] In recent years, with the widespread application of long-span cable-stayed bridges in important projects such as those spanning seas and rivers, fire safety issues have become increasingly prominent. Vehicle fires are one of the most common types of fires on bridges. Burning vehicles are often directed to the emergency lanes adjacent to the cable system, causing flames and high-temperature smoke to directly impact the cables, resulting in severe ablation and structural damage. As the core load-bearing components and "lifeline" of such bridges, cables are usually irreplaceable or extremely difficult to replace. Therefore, the rapid and accurate assessment of their remaining load-bearing capacity after a fire is crucial to ensuring the overall safety of the bridge.

[0003] Currently, the main challenges in assessing cable performance after a fire are: firstly, in actual fires, it is impossible to perform destructive sampling and laboratory mechanical testing on critical components such as the main cable, which cannot be replaced; secondly, existing sensor-based monitoring methods are not only costly, but the sensors are also easily damaged in real fires, leading to data loss. The peak temperature experienced by the cable wires in a fire is the core factor determining its residual mechanical properties. Therefore, accurately inferring the peak temperature of the cable after a fire without relying on sensors or damaging the components has become a key challenge in actual bridge engineering fires.

[0004] Although some technologies have adopted traditional image processing methods or machine learning for damage identification, these methods mostly rely on ideal experimental environments and data collected by sensors. They fail to adequately address real-world conditions such as sensor malfunction, smoke interference, variable material morphology, and complex image features in real fires, resulting in insufficient generalization ability and accuracy.

[0005] Therefore, there is an urgent need to develop an intelligent peak temperature prediction method that can be based on the apparent image features of cables after a fire, without the need for historical sensor data, and that combines high accuracy with strong generalization ability, so as to provide reliable technical support for safety assessment and maintenance decisions after bridge fires. Summary of the Invention

[0006] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the first objective of this invention is to propose a rapid and non-destructive method for predicting the maximum temperature of cables after a fire based on deep learning. This method can accurately and efficiently determine the peak temperature of the steel wires and the temperature of the fire smoke based on post-fire images of the cables, without the need for pre-deployment of sensors or disassembly of components, providing a reliable technical means for assessing the safety performance of bridges after a fire and making maintenance decisions.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a rapid and non-destructive method for predicting the maximum temperature of a cable catching fire based on deep learning, the method comprising: S1, through various fire tests, obtain image data of the apparent morphology of each layer of the cable anti-corrosion-fire protection system under different temperature conditions; among which, each layer of the cable material includes a flame-retardant sealant layer and an anti-corrosion layer; S2, preprocess the image data by splitting it according to the layer characteristics of the cable material; construct a dataset by combining the processed image data with the corresponding temperature labels, and divide the dataset into training set, validation set and test set; S3, Construct a dual-branch weight-shared cable fire maximum temperature prediction model based on ConvNeXt-Tiny; S4. Pre-trained weights are loaded through transfer learning strategy, and a hybrid loss function is used to optimize the cable fire maximum temperature prediction model. The cable fire maximum temperature prediction model is trained using the training set and fine-tuned using the validation set. S5 uses the test set to perform a multi-index joint performance evaluation of the trained cable fire maximum temperature prediction model, and verifies the model's generalization ability and practical applicability through real fire test images. S6 creates a visualization window and imports the cable fire resistance test image into the cable fire maximum temperature prediction model to output the maximum temperature of the cable wire or the maximum temperature of the fire smoke corresponding to the cable fire resistance test image, thereby achieving non-destructive and rapid assessment of the cable's safety performance after a fire without disassembly or sensors.

[0008] In addition, the deep learning-based method for rapid and non-destructive prediction of the maximum temperature of cable fires according to the above embodiments of the present invention may also have the following additional technical features: According to an embodiment of the present invention, step S2 includes: S21. Based on the layer characteristics of cable materials, the images are divided into two categories: the flame-retardant sealant image is used as the first category image for determining the highest temperature of fire smoke; the anti-corrosion layer image is used as the second category image for determining the highest temperature of steel wire. S22, perform grayscale conversion and texture enhancement on the first type of image to eliminate the interference of fire smoke on color and highlight the typical morphological features of fire, including cracks, bubbling and network fractures. S23, retain RGB color information for the second type of image, perform color normalization and appropriate contrast correction, and extract the color change pattern under temperature. S24, the first and second class images are scaled and normalized, and combined with geometric enhancement and controlled brightness / contrast perturbation, the diversity of training samples is increased and key features are not destroyed; among them, geometric enhancement includes rotation, scaling and flipping. S25. The processed images are bound to the corresponding temperature level labels to construct a dataset with differentiated features. The dataset is divided into training set, validation set and test set according to the specimen and working condition stratification method to avoid data leakage and maintain the distribution consistency of each temperature level.

[0009] According to one embodiment of the present invention, the maximum temperature prediction model for cable fires includes two parallel input branch modules, a channel splicing layer, a shared backbone network, a feature mapping layer, and two task-specific ordered classification prediction heads; The two parallel input branch modules include a first branch module and a second branch module; the first branch module is used to process the appearance image of the cable flame-retardant sealant layer and extract texture-related features; the second branch module is used to process the appearance image of the cable anti-corrosion layer and extract features related to color change and state evolution. The ordered classification prediction head is used to receive the feature vector output by the feature mapping layer, including a first classification prediction head and a second classification prediction head; wherein, the first classification prediction head outputs the ordered classification prediction of the highest temperature of the fire smoke in the cable area, and the second classification prediction head outputs the ordered classification prediction of the peak temperature of the cable wire. After preprocessing, the first and second class images are input to the first and second branch modules respectively for feature extraction. The obtained features are fused by the channel splicing layer to form the input of the shared backbone network. The shared backbone network adopts the ConvNeXt-Tiny architecture, which stacks multiple structural units in the order of normalization, convolution operation and nonlinear activation. It performs deep representation learning on the fused input layer by layer and outputs a high-dimensional embedding vector. Then, the high-dimensional embedding vector is input to the feature mapping layer and converted into a low-dimensional compact feature vector. The low-dimensional compact feature vector is fed in parallel to the first and second classification prediction heads and outputs the prediction results of the peak temperature of fire smoke and the peak temperature of cable wire respectively.

[0010] According to an embodiment of the present invention, both the first branch module and the second branch module are serialized feature extraction structures. The serialized feature extraction structure includes multiple convolutional units connected in sequence. Each convolutional unit includes a convolutional layer, a batch normalization layer and a SiLU activation function connected in sequence. The feature mapping layer consists of an adaptive pooling layer, a flattening layer, a linear connection layer, a normalization layer, and a GELU activation layer connected in sequence. Both the first and second classification prediction heads use the same processing sequence. First, feature transformation is performed through a linear layer, then nonlinear processing is performed through a normalization layer and a GELU activation layer, and finally, the prediction result is output through a linear connection layer and a temperature output module.

[0011] According to one embodiment of the present invention, the temperature output modules of both the first classification prediction head and the second classification prediction head adopt a cumulative logistic regression structure, through K... Each binary classification unit learns the cumulative probability of each temperature threshold and maps the output features of the dual-task branch to temperature level prediction results with ordered relationships, where K refers to the total number of temperature level categories. 1 represents the number of binary classifiers that need to be built.

[0012] According to an embodiment of the present invention, the transfer learning strategy in step S4 includes: S41, the model is initialized using the ConvNeXt-Tiny backbone network weights pre-trained on a large-scale image dataset; S42, the flame-retardant sealant image and the anti-corrosion layer image, after split preprocessing, are input into the model in a two-stream manner. Forward propagation is completed through a shared backbone network and dual-task branches, and K is output by the ordered classification head. One cumulative probability is combined to form a prediction result of K temperature levels; S43 uses a hybrid loss function for optimization during training. The hybrid loss function includes cumulative logistic regression loss, cross-entropy loss and ordered distribution constraint loss, and improves the classification accuracy and orderliness of the model by weighted combination. S44 utilizes the training set for parameter learning and performs performance evaluation and tuning on the validation set to improve the model's generalization ability under different fire conditions.

[0013] According to one embodiment of the present invention, the hybrid loss function includes cumulative logistic regression loss, cross-entropy loss, and ordered distribution constraint loss; Cumulative logistic regression loss is used to supervise the K in the ordered classification head. The cumulative probability of a binary classification unit learning each temperature threshold is calculated using the following formula: Among them, L coral For cumulative logistic regression loss, N is the total number of samples, i is the sample index number i=(1,2,......,N), K is the total number of temperature levels, k is the index number of the threshold for each temperature level k=(1,2,......,K-1), z i,k Let σ(·) be the output of the i-th sample in the k-th binary classification unit, and t be the output of the ith sample in the k-th binary classification unit. i,k The cumulative label of the i-th sample at the k-th threshold; Cross-entropy loss is used to supervise the final temperature level prediction results and improve the overall classification accuracy. Its calculation formula is as follows: Among them, Lce The cross-entropy loss is given by c, where c is the temperature level category index (c=(1,2,......,K)). i,c Let p be the true label distribution of the i-th sample in the c-th temperature category. i,c Let be the predicted probability of the i-th sample in the c-th temperature category; The ordered distribution constraint loss is used to measure the difference between the predicted distribution and the true label distribution while maintaining the continuity between temperature levels. Its calculation formula is as follows: Among them, L emd For ordered distribution constraint loss, j is the cumulative index of temperature level category j=(1,2,......,K), p i,j Let y be the predicted probability of the i-th sample in the j-th temperature category. i,j Let be the true label distribution of the i-th sample in the j-th temperature category; The cumulative logistic regression loss, cross-entropy loss, and ordered distribution constraint loss are weighted and combined to form a hybrid loss function, which simultaneously optimizes the classification accuracy and temperature level ordering of the cable fire maximum temperature prediction model. The formula for calculating the hybrid loss function is as follows: Among them, α, β, and γ are preset weight coefficients used to balance classification accuracy and order discrimination ability.

[0014] According to an embodiment of the present invention, step S5 includes: S51, the two branches of the prediction results output by the cable fire maximum temperature prediction model are decoded by ordered probability distribution, the cumulative probability distribution is converted into the discrete probability distribution of each temperature category, and the category with the highest probability is selected as the predicted temperature level. S52 calls the accuracy, precision, recall and F1 score calculation modules to calculate the classification performance indicators of the prediction task corresponding to the peak temperature of the main cable wire, the peak temperature of the suspension cable wire and the peak temperature of the fire smoke, respectively. S53. Based on the predicted probability distribution and the real label, EMD is calculated as an ordinal consistency evaluation index to measure the degree of matching of the model's predicted distribution on the ordered temperature category. S54: Load real HC fire test image data and input it into the trained cable fire maximum temperature prediction model. Calculate the cross-dataset generalization accuracy of the main cable wire, suspension cable wire, and fire smoke temperature prediction tasks in real scenarios, and output the confusion matrix and class activation heatmap for each category.

[0015] According to an embodiment of the present invention, in step S6, creating a visualization window includes: S61 is based on a graphical user interface framework to build a dedicated visualization window, build the main window interface and integrate the trained cable fire maximum temperature prediction model, and design and lay out functional areas for image display, temperature result display and user control. S62, an image input module is designed in the visualization window. The image processing library is used to realize the image reading and preprocessing functions. It supports the import of one or more appearance images after the cable fire resistance test through the file dialog box or drag and drop. The input image is automatically normalized in size and converted in channel format to meet the tensor requirements of the cable fire maximum temperature prediction model. S63, the dual-branch weight shared cable fire maximum temperature prediction model is called to perform forward inference on the preprocessed image, the model weights are loaded through the deep learning framework and forward propagation calculation is performed, and the predicted values ​​of the peak temperature of the main cable steel wire and the peak temperature of the fire smoke are output respectively. S64 displays the original image, predicted temperature value, and temperature level label simultaneously in the visualization interface through a graphics drawing component.

[0016] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0017] Figure 1 A flowchart of a deep learning-based method for rapid and non-destructive prediction of the maximum temperature of cable fires according to an embodiment of the present invention; Figure 2 This is a network architecture diagram of a dual-branch weight-shared cable fire maximum temperature prediction model according to an embodiment of the present invention; Figure 3 This is a network architecture diagram of an input branch module according to an embodiment of the present invention; Figure 4 This is a network architecture diagram of an ordered classification prediction head according to an embodiment of the present invention; Figure 5 This is a network architecture diagram of the feature mapping layer according to an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the accuracy of a prediction task corresponding to the peak temperature of the steel wire according to an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating the accuracy of a prediction task corresponding to the peak temperature of flue gas according to an embodiment of the present invention. Figure 8 This is a schematic diagram of the confusion matrix corresponding to the peak temperature of the steel wire according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the confusion matrix corresponding to the peak temperature of flue gas according to an embodiment of the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0019] The following description, with reference to the accompanying drawings, describes a method for rapid and non-destructive prediction of the maximum temperature of cable fires based on deep learning, as proposed in an embodiment of the present invention.

[0020] like Figure 1 As shown in the figure, the method for rapid and non-destructive prediction of the maximum temperature of a cable fire based on deep learning according to an embodiment of the present invention may include the following steps: S1, through various fire tests, obtain the appearance morphology image data of each layer of the cable anti-corrosion-fire protection system under different temperature conditions; among them, each layer of cable material includes flame-retardant sealant layer and anti-corrosion layer.

[0021] S2, perform split preprocessing on the image data according to the layer characteristics of the cable material; construct a dataset by combining the processed image data with the corresponding temperature labels, and divide the dataset into training set, validation set and test set.

[0022] S3. Construct a dual-branch weight-shared cable fire maximum temperature prediction model based on ConvNeXt-Tiny.

[0023] S4. Pre-trained weights are loaded through transfer learning strategy, and a hybrid loss function is used to optimize the cable fire maximum temperature prediction model. The cable fire maximum temperature prediction model is trained using the training set and fine-tuned using the validation set.

[0024] S5 uses a test set to perform a multi-index joint performance evaluation of the trained cable fire maximum temperature prediction model, and verifies the model's generalization ability and practical applicability through real fire test images.

[0025] S6 creates a visualization window and imports the cable fire resistance test image into the cable fire maximum temperature prediction model to output the maximum temperature of the cable wire or the maximum temperature of the fire smoke corresponding to the cable fire resistance test image, thereby achieving non-destructive and rapid assessment of the cable's safety performance after a fire without disassembly or sensors.

[0026] Specifically, the core principle of the method in this invention is based on convolutional neural networks (ConvNeXt-Tiny) and multi-source image feature fusion technology in deep learning. It achieves temperature prediction by analyzing the nonlinear mapping relationship between the apparent morphological changes of the cable protection layers (flame-retardant sealant layer and anti-corrosion layer) after a fire and the maximum temperature. The process involves first acquiring image data of each layer of material at different temperatures through various fire tests, and then performing preprocessing based on layer characteristics to construct a temperature-labeled dataset. Subsequently, a dual-branch weight-sharing cable fire maximum temperature prediction model is constructed. Transfer learning is used to load pre-trained weights, and the model parameters are optimized through a hybrid loss function. Ultimately, this achieves high-precision learning of the correlation between image features and temperature, thus avoiding the limitations of traditional sensors or disassembly detection.

[0027] Specifically, the workflow of the method in this embodiment of the invention can be divided into three stages: data preparation, model training, and validation. In the data preparation stage, namely steps S1 and S2, image data of each layer of cable material at different temperatures are collected through controlled fire tests. After preprocessing by layer-by-layer splitting, a labeled dataset is constructed and divided into training, validation, and test sets. In the model training stage, namely steps S3 and S4, a dual-branch weight-sharing cable maximum fire temperature prediction model based on ConvNeXt-Tiny is constructed. Parameters are initialized through transfer learning, and end-to-end training is performed using a hybrid loss function (such as combining mean squared error and perceptual loss). Hyperparameters are adjusted using the validation set to prevent overfitting. In the validation and application stage, namely steps S5 and S6, performance is evaluated using multiple indicators (such as MAE and RMSE) on the test set, and the generalization ability is verified using real fire images. Finally, it is deployed to a visualization window to achieve lossless temperature prediction.

[0028] Therefore, the method of this invention enables non-destructive and rapid prediction of the maximum temperature of cable fires without the need to disassemble or install sensors, significantly reducing assessment costs and time. Through a dual-branch structure and transfer learning strategy, the model's sensitivity to minute morphological changes and prediction accuracy are improved (errors can be controlled within ±5℃). It has strong generalization ability and can adapt to complex and ever-changing scenarios in actual fires, providing reliable technical support for the safety performance assessment of cable fires.

[0029] According to an embodiment of the present invention, step S2 includes: S21. Based on the layer characteristics of cable materials, the images are divided into two categories: the flame-retardant sealant image is used as the first category image for determining the highest temperature of fire smoke; the anti-corrosion layer image is used as the second category image for determining the highest temperature of steel wire. S22, perform grayscale conversion and texture enhancement on the first type of image to eliminate the interference of fire smoke on color and highlight the typical morphological features of fire, including cracks, bubbling and network fractures. S23, retain RGB color information for the second type of image, perform color normalization and appropriate contrast correction, and extract the color change pattern under temperature. S24, the first and second class images are scaled and normalized, and combined with geometric enhancement and controlled brightness / contrast perturbation, the diversity of training samples is increased and key features are not destroyed; among them, geometric enhancement includes rotation, scaling and flipping. S25. The processed images are bound to the corresponding temperature level labels to construct a dataset with differentiated features. The dataset is divided into training set, validation set and test set according to the specimen and working condition stratification method to avoid data leakage and maintain the distribution consistency of each temperature level.

[0030] Therefore, by employing a diversion and differentiated preprocessing strategy, the correlation between flue gas morphology characteristics, anti-corrosion layer color characteristics, and temperature was specifically strengthened, improving the accuracy and reliability of subsequent model predictions. Secondly, through geometric augmentation and controlled perturbation operations, the diversity of training samples was effectively expanded, enhancing the model's generalization ability and anti-interference capability. Thirdly, by adopting a hierarchical dataset partitioning method, the risk of data leakage was eliminated, ensuring the scientific validity and practical applicability of the model evaluation results, laying a solid foundation for the efficient training of the subsequent dual-branch model.

[0031] According to one embodiment of the present invention, such as Figure 2 As shown, the cable fire maximum temperature prediction model includes two parallel input branch modules, a channel splicing layer, a shared backbone network, a feature mapping layer, and two task-specific ordered classification prediction heads; The two parallel input branch modules include a first branch module and a second branch module; the first branch module is used to process the appearance image of the cable flame retardant sealant layer and extract texture-related features; the second branch module is used to process the appearance image of the cable anti-corrosion layer and extract features related to color change and state evolution. The ordered classification prediction head is used to receive the feature vector output by the feature mapping layer, including a first classification prediction head and a second classification prediction head; wherein, the first classification prediction head outputs the ordered classification prediction of the highest temperature of the fire smoke in the cable area, and the second classification prediction head outputs the ordered classification prediction of the peak temperature of the cable wire. After preprocessing, the first and second class images are input to the first and second branch modules respectively for feature extraction. The obtained features are fused by the channel splicing layer to form the input of the shared backbone network. The shared backbone network adopts the ConvNeXt-Tiny architecture, which stacks multiple structural units in the order of normalization, convolution operation and nonlinear activation. It performs deep representation learning on the fused input layer by layer and outputs a high-dimensional embedding vector. Then, the high-dimensional embedding vector is input to the feature mapping layer and converted into a low-dimensional compact feature vector. The low-dimensional compact feature vector is fed in parallel to the first and second classification prediction heads and outputs the prediction results of the peak temperature of fire smoke and the peak temperature of cable wire respectively.

[0032] According to one embodiment of the present invention, both the first branch module and the second branch module are serialized feature extraction structures. The serialized feature extraction structure includes multiple convolutional units connected in sequence. Each convolutional unit includes a convolutional layer, a batch normalization layer, and a SiLU activation function connected in sequence. For example, ... Figure 3 As shown, both the first and second branch modules employ a sequential feature extraction structure consisting of three convolutional units. Each convolutional unit sequentially includes a convolutional layer, a batch normalization layer, and a SiLU activation function. The control parameters for each convolutional layer include the number of input channels In_c, the number of output channels Out_c, the kernel size k, the stride s, and the padding size p. The first convolutional unit is used for primary feature extraction. The parameters of the first convolutional layer in the first branch module are In_c=1, Out_c=32, k=3, s=1, and p=1. The parameters of the first convolutional layer of the branch module are In_c=3, Out_c=32, k=3, s=1, p=1; the second convolutional unit is used for downsampling and channel expansion, and the parameters of the second convolutional layer of both the first and second branch modules are In_c=32, Out_c=48, k=3, s=2, p=1; the third convolutional unit is used for feature refinement and output, and the parameters of the third convolutional layer of both the first and second branch modules are In_c=48, Out_c=48, k=3, s=1, p=1. After processing by the three convolutional units, the first and second branch modules output feature maps of the same size and with the same number of channels, which are then used for subsequent channel concatenation and sharing with the backbone network.

[0033] The feature mapping layer comprises, in sequence, an adaptive pooling layer, a flattening layer, a linear connection layer, a normalization layer, and a GELU activation layer. For example, ... Figure 5 As shown, the linear connection layer is used for channel dimension mapping, with parameters In_c=96 input channels and Out_c=768 output channels, realizing the mapping from spliced ​​features to a high-dimensional embedding space.

[0034] Both the first and second classification prediction heads use the same processing sequence. First, feature transformation is performed through a linear layer, then nonlinear processing is performed through a normalization layer and a GELU activation layer, and finally, the prediction result is output through a linear connection layer and a temperature output module.

[0035] Specifically, such as Figure 4 As shown, both the first and second classification prediction heads use the same processing sequence. First, feature transformation is performed through a first-level linear connection layer, followed by nonlinear processing through a normalization layer and a GELU activation layer, and then converted to K-order values ​​via a second-level linear connection layer and a temperature ordinal output module. One logits is output to produce the final prediction result. The first-level linear connection layer has 768 input channels and 384 output channels, which is used to compress high-dimensional embedded features and extract intermediate representations. The second-level linear connection layer has 384 input channels and 6 output channels, which is used to output feature vectors corresponding to temperature levels.

[0036] According to one embodiment of the present invention, the temperature output modules of both the first classification prediction head and the second classification prediction head adopt a cumulative logistic regression structure, through K... Each binary classification unit learns the cumulative probability of each temperature threshold and maps the output features of the dual-task branch to temperature level prediction results with ordered relationships, where K refers to the total number of temperature level categories. 1 represents the number of binary classifiers that need to be built.

[0037] According to an embodiment of the present invention, the transfer learning strategy in step S4 includes: S41, the model is initialized using the ConvNeXt-Tiny backbone network weights pre-trained on a large-scale image dataset; S42, the flame-retardant sealant image and the anti-corrosion layer image, after split preprocessing, are input into the model in a two-stream manner. Forward propagation is completed through a shared backbone network and dual-task branches, and K is output by the ordered classification head. One cumulative probability is combined to form a prediction result of K temperature levels; S43 uses a hybrid loss function for optimization during training. The hybrid loss function includes cumulative logistic regression loss, cross-entropy loss and ordered distribution constraint loss, and improves the classification accuracy and orderliness of the model by weighted combination. S44 utilizes the training set for parameter learning and performs performance evaluation and tuning on the validation set to improve the model's generalization ability under different fire conditions.

[0038] According to one embodiment of the present invention, the hybrid loss function includes cumulative logistic regression loss, cross-entropy loss, and ordered distribution constraint loss; Cumulative logistic regression loss is used to supervise the K in the ordered classification head. The cumulative probability of a binary classification unit learning each temperature threshold is calculated using the following formula: Among them, L coral For cumulative logistic regression loss, N is the total number of samples, i is the sample index number i=(1,2,......,N), K is the total number of temperature levels, k is the index number of the threshold for each temperature level k=(1,2,......,K-1), z i,k Let σ(·) be the output of the i-th sample in the k-th binary classification unit, and t be the output of the ith sample in the k-th binary classification unit. i,k The cumulative label of the i-th sample at the k-th threshold; Cross-entropy loss is used to supervise the final temperature level prediction results and improve the overall classification accuracy. Its calculation formula is as follows: Among them, L ce The cross-entropy loss is given by c, where c is the temperature level category index (c=(1,2,......,K)). i,c Let p be the true label distribution of the i-th sample in the c-th temperature category. i,c Let be the predicted probability of the i-th sample in the c-th temperature category; The ordered distribution constraint loss is used to measure the difference between the predicted distribution and the true label distribution while maintaining the continuity between temperature levels. Its calculation formula is as follows: Among them, L emd For ordered distribution constraint loss, j is the cumulative index of temperature level category j=(1,2,......,K), p i,j Let y be the predicted probability of the i-th sample in the j-th temperature category. i,j Let be the true label distribution of the i-th sample in the j-th temperature category; The cumulative logistic regression loss, cross-entropy loss, and ordered distribution constraint loss are weighted and combined to form a hybrid loss function, which simultaneously optimizes the classification accuracy and temperature level ordering of the cable fire maximum temperature prediction model. The formula for calculating the hybrid loss function is as follows: Among them, α, β, and γ are preset weight coefficients used to balance classification accuracy and order discrimination ability.

[0039] According to an embodiment of the present invention, step S5 includes: S51, the two branches of the prediction results output by the cable fire maximum temperature prediction model are decoded by ordered probability distribution, the cumulative probability distribution is converted into the discrete probability distribution of each temperature category, and the category with the highest probability is selected as the predicted temperature level.

[0040] S52 calls the accuracy, precision, recall, and F1 score calculation modules to calculate the classification performance indicators of the prediction task corresponding to the peak temperature of the main cable wire, the peak temperature of the suspension cable wire, and the peak temperature of the fire smoke, respectively.

[0041] For example, the accuracy of the prediction task corresponding to the peak temperature of the steel wire is as follows: Figure 6 As shown, the accuracy of the prediction task corresponding to the peak flue gas temperature is as follows: Figure 7 As shown in Table 1, the classification index of the cable fire maximum temperature prediction model on the test set is as follows.

[0042] Table 1 S53. Based on the predicted probability distribution and the real label, EMD is calculated as an ordinal consistency evaluation index to measure the degree of matching of the model's predicted distribution on the ordered temperature category. S54: Load real HC fire test image data and input it into the trained cable fire maximum temperature prediction model. Calculate the cross-dataset generalization accuracy of the main cable wire, suspension cable wire, and fire smoke temperature prediction tasks in real scenarios, and output the confusion matrix and class activation heatmap for each category.

[0043] For example, the confusion matrix corresponding to the peak temperature of the steel wire is as follows: Figure 8 As shown, the confusion matrix corresponding to the peak flue gas temperature is as follows: Figure 9 As shown in Table 2, the classification index of the cable fire maximum temperature prediction model on real fire image information is as follows.

[0044] Table 2 According to an embodiment of the present invention, in step S6, creating a visualization window includes: S61 is based on a graphical user interface framework to build a dedicated visualization window, build the main window interface and integrate the trained cable fire maximum temperature prediction model, and design and lay out functional areas for image display, temperature result display and user control. S62, an image input module is designed in the visualization window. The image processing library is used to realize the image reading and preprocessing functions. It supports the import of one or more appearance images after the cable fire resistance test through the file dialog box or drag and drop. The input image is automatically normalized in size and converted in channel format to meet the tensor requirements of the cable fire maximum temperature prediction model. S63, the dual-branch weight shared cable fire maximum temperature prediction model is called to perform forward inference on the preprocessed image, the model weights are loaded through the deep learning framework and forward propagation calculation is performed, and the predicted values ​​of the peak temperature of the main cable steel wire and the peak temperature of the fire smoke are output respectively. S64 displays the original image, predicted temperature value, and temperature level label simultaneously in the visualization interface through a graphics drawing component.

[0045] Specifically, a dedicated visualization window can be built based on the Python graphical interface framework. The main window interface is constructed using PyQt or Tkinter, integrating a trained cable fire maximum temperature prediction model. The image display area, temperature result display panel, control button group, and heat map visualization area are designed and laid out using QtDesigner. An image input module is designed within the window, using the OpenCV library to implement image reading and preprocessing functions. It supports importing one or more apparent images of the cable fire resistance test results via the QFileDialog file dialog box or drag-and-drop event response mechanism, and automatically normalizes the size and converts the channel format of the input images to meet the tensor requirements of the cable fire maximum temperature prediction model. The cable fire maximum temperature prediction model is then called to perform forward inference on the preprocessed images. The model weights are loaded using the PyTorch framework, and the predicted values ​​of the peak temperature of the main cable wires and the peak temperature of the fire smoke are output using the forward method. The original image, predicted temperature values, and temperature level labels are simultaneously displayed in the visualization interface using Matplotlib or OpenCV plotting components.

[0046] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0047] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0048] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0049] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A rapid and non-destructive method for predicting the maximum temperature of cables after fire based on deep learning, characterized in that, The method includes: S1, through various fire tests, obtain image data of the apparent morphology of each layer of the cable anti-corrosion-fire protection system under different temperature conditions; among which, each layer of the cable material includes a flame-retardant sealant layer and an anti-corrosion layer; S2, preprocess the image data by splitting it according to the layer characteristics of the cable material; construct a dataset by combining the processed image data with the corresponding temperature labels, and divide the dataset into training set, validation set and test set; S3, Construct a dual-branch weight-shared cable fire maximum temperature prediction model based on ConvNeXt-Tiny; S4. Pre-trained weights are loaded through transfer learning strategy, and a hybrid loss function is used to optimize the cable fire maximum temperature prediction model. The cable fire maximum temperature prediction model is trained using the training set and fine-tuned using the validation set. S5 uses the test set to perform a multi-index joint performance evaluation of the trained cable fire maximum temperature prediction model, and verifies the model's generalization ability and practical applicability through real fire test images. S6 creates a visualization window and imports the cable fire resistance test image into the cable fire maximum temperature prediction model to output the maximum temperature of the cable wire or the maximum temperature of the fire smoke corresponding to the cable fire resistance test image, thereby achieving non-destructive and rapid assessment of the cable's safety performance after a fire without disassembly or sensors.

2. The method for rapid and non-destructive prediction of the maximum fire temperature of cables based on deep learning according to claim 1, characterized in that, Step S2 includes: S21. Based on the layer characteristics of cable materials, the images are divided into two categories: the flame-retardant sealant image is used as the first category image for determining the highest temperature of fire smoke; the anti-corrosion layer image is used as the second category image for determining the highest temperature of steel wire. S22, perform grayscale conversion and texture enhancement on the first type of image to eliminate the interference of fire smoke on color and highlight the typical fire morphological features, including cracks, bubbling and network fractures. S23, retain the RGB color information of the second type of image, perform color normalization and appropriate contrast correction, and extract the color change pattern under the influence of temperature; S24, the first type of image and the second type of image are scaled and normalized, and combined with geometric enhancement and controlled brightness / contrast perturbation, the diversity of training samples is increased and key features are not destroyed; wherein, geometric enhancement includes rotation, scaling and flipping. S25. The processed images are bound to the corresponding temperature level labels to construct a dataset with differentiated features. The dataset is divided into training set, validation set and test set according to the specimen and working condition stratification method to avoid data leakage and maintain the distribution consistency of each temperature level.

3. The method for rapid and non-destructive prediction of the maximum fire temperature of cables based on deep learning according to claim 2, characterized in that, The cable fire maximum temperature prediction model includes two parallel input branch modules, a channel splicing layer, a shared backbone network, a feature mapping layer, and two task-specific ordered classification prediction heads; The two parallel input branch modules include a first branch module and a second branch module; the first branch module is used to process the appearance image of the cable flame-retardant sealant layer and extract texture-related features; the second branch module is used to process the appearance image of the cable anti-corrosion layer and extract features related to color change and state evolution. The ordered classification prediction head is used to receive the feature vector output by the feature mapping layer, including a first classification prediction head and a second classification prediction head; wherein, the first classification prediction head outputs the ordered classification prediction of the highest temperature of the fire smoke in the cable area, and the second classification prediction head outputs the ordered classification prediction of the peak temperature of the cable wire. After preprocessing, the first and second type images are input to the first and second branch modules respectively for feature extraction. The obtained features are fused by the channel splicing layer to form the input of the shared backbone network. The shared backbone network adopts the ConvNeXt-Tiny architecture, which stacks multiple structural units in the order of normalization, convolution operation and nonlinear activation. It performs deep representation learning on the fused input layer by layer and outputs a high-dimensional embedding vector. Then, the high-dimensional embedding vector is input to the feature mapping layer and converted into a low-dimensional compact feature vector. The low-dimensional compact feature vector is fed in parallel to the first classification prediction head and the second classification prediction head to output the prediction results of the peak temperature of fire smoke and the peak temperature of cable wire, respectively.

4. The method for rapid and non-destructive prediction of the maximum temperature of a cable fire based on deep learning according to claim 3, characterized in that, Both the first branch module and the second branch module are serialized feature extraction structures. The serialized feature extraction structure includes multiple convolutional units connected in sequence. Each convolutional unit includes a convolutional layer, a batch normalization layer and a SiLU activation function connected in sequence. The feature mapping layer includes an adaptive pooling layer, a flattening layer, a linear connection layer, a normalization layer, and a GELU activation layer connected in sequence. Both the first and second classification prediction heads use the same processing sequence. First, feature transformation is performed through a linear layer, then nonlinear processing is performed through a normalization layer and a GELU activation layer, and finally, the prediction result is output through a linear connection layer and a temperature output module.

5. The method for rapid and non-destructive prediction of the maximum fire temperature of cables based on deep learning according to claim 4, characterized in that, Both the temperature output modules of the first and second classification prediction heads employ a cumulative logistic regression structure, using K... Each binary classification unit learns the cumulative probability of each temperature threshold and maps the output features of the dual-task branch to temperature level prediction results with ordered relationships, where K refers to the total number of temperature level categories. 1 represents the number of binary classifiers that need to be built.

6. The method for rapid and non-destructive prediction of the maximum temperature of a cable ignition point based on deep learning according to claim 1, characterized in that, The transfer learning strategies in step S4 include: S41, the model is initialized using the ConvNeXt-Tiny backbone network weights pre-trained on a large-scale image dataset; S42, the flame-retardant sealant image and the anti-corrosion layer image, after split preprocessing, are input into the model in a two-stream manner. Forward propagation is completed through a shared backbone network and dual-task branches, and K is output by the ordered classification head. One cumulative probability is combined to form a prediction result of K temperature levels; S43, During the training process, a hybrid loss function is used for optimization. The hybrid loss function includes cumulative logistic regression loss, cross-entropy loss and ordered distribution constraint loss, and the classification accuracy and orderliness of the model are improved by weighted combination. S44 utilizes the training set for parameter learning and performs performance evaluation and tuning on the validation set to improve the model's generalization ability under different fire conditions.

7. The method for rapid and non-destructive prediction of the maximum fire temperature of cables based on deep learning according to claim 6, characterized in that, The hybrid loss function includes cumulative logistic regression loss, cross-entropy loss, and ordered distribution constraint loss. The cumulative logistic regression loss is used to supervise the K in the ordered classification head. The cumulative probability of a binary classification unit learning each temperature threshold is calculated using the following formula: Among them, L coral For cumulative logistic regression loss, N is the total number of samples, i is the sample index number i=(1,2,......,N), K is the total number of temperature levels, k is the index number of the threshold for each temperature level k=(1,2,......,K-1), z i,k Let σ(·) be the output of the i-th sample in the k-th binary classification unit, and t be the output of the ith sample in the k-th binary classification unit. i,k The cumulative label of the i-th sample at the k-th threshold; The cross-entropy loss is used to supervise the final temperature level prediction results and improve the overall classification accuracy. Its calculation formula is as follows: Among them, L ce The cross-entropy loss is given by c, where c is the temperature level category index (c=(1,2,......,K)). i,c Let p be the true label distribution of the i-th sample in the c-th temperature category. i,c Let be the predicted probability of the i-th sample in the c-th temperature category; The ordered distribution constraint loss is used to measure the difference between the predicted distribution and the true label distribution while maintaining the continuity between temperature levels. Its calculation formula is as follows: Among them, L emd For ordered distribution constraint loss, j is the cumulative index of temperature level category j=(1,2,......,K), p i,j Let y be the predicted probability of the i-th sample in the j-th temperature category. i,j Let be the true label distribution of the i-th sample in the j-th temperature category; The cumulative logistic regression loss, the cross-entropy loss, and the ordered distribution constraint loss are weighted and combined to form a hybrid loss function, which simultaneously optimizes the classification accuracy and temperature level ordering of the cable fire maximum temperature prediction model. The calculation formula for the hybrid loss function is as follows: Among them, α, β, and γ are preset weight coefficients used to balance classification accuracy and order discrimination ability.

8. The method for rapid and non-destructive prediction of the maximum temperature of a cable fire based on deep learning according to claim 1, characterized in that, Step S5 includes: S51, the two branches of the prediction results output by the cable fire maximum temperature prediction model are decoded by ordered probability distribution, the cumulative probability distribution is converted into the discrete probability distribution of each temperature category, and the category with the highest probability is selected as the predicted temperature level. S52 calls the accuracy, precision, recall and F1 score calculation modules to calculate the classification performance indicators of the prediction task corresponding to the peak temperature of the main cable wire, the peak temperature of the suspension cable wire and the peak temperature of the fire smoke, respectively. S53. Based on the predicted probability distribution and the real label, EMD is calculated as an ordinal consistency evaluation index to measure the degree of matching of the model's predicted distribution on the ordered temperature category. S54: Load real HC fire test image data and input it into the trained cable fire maximum temperature prediction model. Calculate the cross-dataset generalization accuracy of the main cable wire, suspension cable wire, and fire smoke temperature prediction tasks in real scenarios, and output the confusion matrix and class activation heatmap for each category.

9. The method for rapid and non-destructive prediction of the maximum fire temperature of cables based on deep learning according to claim 1, characterized in that, In step S6, creating the visualization window includes: S61 is based on a graphical user interface framework to build a dedicated visualization window, build the main window interface and integrate the trained cable fire maximum temperature prediction model, and design and lay out functional areas for image display, temperature result display and user control. S62, an image input module is designed in the visualization window. The image processing library is used to realize the image reading and preprocessing functions. It supports the import of one or more appearance images after the cable fire resistance test through the file dialog box or drag and drop. The input image is automatically normalized in size and converted in channel format to meet the tensor requirements of the cable fire maximum temperature prediction model. S63, the dual-branch weight shared cable fire maximum temperature prediction model is called to perform forward inference on the preprocessed image, the model weights are loaded through the deep learning framework and forward propagation calculation is performed, and the predicted values ​​of the peak temperature of the main cable steel wire and the peak temperature of the fire smoke are output respectively. S64 displays the original image, predicted temperature value, and temperature level label simultaneously in the visualization interface through a graphics drawing component.

Images