A deep learning-based rapid non-destructive prediction method for the maximum temperature of a cable after fire

Abstract

The application discloses a kind of based on deep learning's cable over fire maximum temperature fast nondestructive prediction method, it is related to bridge post-fire safety evaluation technical field.The method obtains the apparent image data of cable anticorrosion-fire protection system at different temperatures by various fire tests, and constructs special data set after shunt pretreatment;Innovatively, a double-branch weight sharing prediction model based on ConvNeXt-Tiny is constructed, the state characteristics of cable flame-retardant sealant layer and anticorrosion layer are extracted by two parallel branches respectively, deep representation learning is carried out by feature fusion and shared backbone network, and finally high-precision parallel prediction of double temperature parameters is realized by two ordered classification prediction heads;Using transfer learning strategy and hybrid loss function for training, the prediction accuracy and generalization ability are significantly improved;Finally, a visual window is developed to realize the fast nondestructive evaluation of cable over fire temperature without disassembly and sensor.The method provides reliable technical support for post-fire safety performance evaluation of bridge cable.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bridge engineering and artificial intelligence, and particularly relates to a cable maximum temperature after fire rapid non-destructive prediction method based on deep learning. BACKGROUND

[0002] In recent years, with the wide application of long-span cable-supported bridges in important projects such as sea and river crossings, the fire safety problem has become increasingly prominent. Vehicle fire is one of the most common types of fires on bridges. Burning vehicles are often directed to the emergency lane adjacent to the cable system, causing the flame and high-temperature smoke to directly act on the cable, resulting in serious ablation and structural damage. As the core load-bearing component and "lifeline" of such bridges, the cable is usually irreplaceable or extremely difficult to replace. The rapid and accurate assessment of its residual carrying capacity after fire is a key to ensuring the overall safety of the bridge.

[0003] Currently, the main difficulties in evaluating the performance of the cable after fire are as follows: first, destructive sampling and laboratory mechanical testing of important components such as main cables cannot be performed after actual fire; second, existing monitoring methods based on sensors are not only costly, but also sensors are easily damaged in real fires, resulting in data loss. The peak temperature experienced by the cable wire in the fire is a key factor in determining its residual mechanical properties. Therefore, how to accurately infer the peak temperature of the cable during the fire without relying on sensors and damaging the components has become a key challenge in actual bridge engineering fires.

[0004] Although some technologies use traditional image processing methods or machine learning for damage identification, these methods rely on ideal test environments and sensor-collected data, and are unable to fully cope with actual conditions such as sensor failure, smoke interference, material morphology variability, and complex image features in real fires, resulting in insufficient generalization ability and accuracy.

[0005] Therefore, there is an urgent need to develop a peak temperature intelligent prediction method that can be based on post-fire cable surface image features, does not require sensor historical data, and balances high precision and strong generalization ability, providing reliable technical support for post-fire safety evaluation and maintenance decision-making of bridges. SUMMARY

[0006] The present application aims to at least partially solve one of the technical problems in the related art. To this end, the first purpose of the present application is to propose a cable maximum temperature after fire rapid non-destructive prediction method based on deep learning, which can be based on post-fire cable surface images, without the need for pre-deployed sensors or disassembled components, to achieve high-precision and efficient determination of wire peak temperature and fire smoke temperature, providing a reliable technical means for post-fire safety performance evaluation and maintenance decision-making of bridges.

[0007] To achieve the above object, the first aspect of the embodiment of the present application provides a deep learning-based rapid non-destructive prediction method for the maximum temperature of a cable after fire, which comprises:

[0008] S1, obtaining apparent morphology image data of each layer of material of a cable corrosion-fire protection system under different temperature conditions through various fire tests; wherein the layers of material of the cable include a fire-retardant sealant layer and a corrosion-resistant layer;

[0009] S2, performing shunt preprocessing on the image data according to the characteristics of the layers of cable material; constructing a data set by combining the processed image data with corresponding temperature labels, and dividing the data set into a training set, a validation set and a test set;

[0010] S3, constructing a double-branch weight-sharing cable maximum temperature prediction model based on ConvNeXt-Tiny;

[0011] S4, loading pre-trained weights through a transfer learning strategy, optimizing the cable maximum temperature prediction model using a hybrid loss function, training the cable maximum temperature prediction model using the training set, and tuning the model through the validation set;

[0012] S5, performing multi-index joint performance evaluation on the trained cable maximum temperature prediction model using the test set, and verifying the generalization ability and actual applicability of the model through real fire test images;

[0013] S6, creating a visual window and importing the cable fire resistance test image into the cable maximum temperature prediction model to output the maximum temperature of the cable steel wire or the maximum temperature of the fire smoke corresponding to the cable fire resistance test image, thereby realizing non-destructive and rapid evaluation of the safety performance of the cable after fire without disassembly and sensors.

[0014] In addition, the deep learning-based rapid non-destructive prediction method for the maximum temperature of a cable after fire according to the above embodiment of the present application can also have the following additional technical features:

[0015] According to one embodiment of the present application, step S2 comprises:

[0016] S21, according to the characteristics of the layers of cable material, the image is shunted, the fire-retardant sealant image is taken as the first type of image for determining the maximum temperature of fire smoke, and the corrosion-resistant layer image is taken as the second type of image for determining the maximum temperature of steel wire;

[0017] S22, performing grayscale and texture enhancement on the first type of image to eliminate the interference of fire smoke on color, highlight the typical morphological features of fire, and the typical morphological features of fire include cracks, bubbles and net-like fractures;

[0018] S23, the RGB color information of the second type of image is retained, color normalization and moderate contrast correction are performed, and the color change rule under the temperature action is extracted;

[0019] S24, the size scaling and normalization are performed on the first type of image and the second type of image, the geometric enhancement and the controlled brightness / contrast disturbance are combined, the diversity of the training samples is increased, and it is ensured that the key features are not destroyed; wherein the geometric enhancement includes rotation, scaling and flipping;

[0020] S25, the processed image is bound with the corresponding temperature grade label, the data set of differentiated features is constructed, the data set is divided into a training set, a verification set and a test set in a layered manner according to the test piece and the working condition, so as to avoid data leakage and keep the distribution consistency of each temperature grade.

[0021] According to one embodiment of the present application, the cable over-fire maximum temperature prediction model comprises 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;

[0022] The two parallel input branch modules include a first branch module and a second branch module; the first branch module is used for processing the apparent image of the cable flame-retardant sealing rubber layer and extracting the features related to the texture; the second branch module is used for processing the apparent image of the cable anticorrosion layer and extracting the features related to the color change and state evolution;

[0023] The ordered classification prediction head is used for receiving the feature vector output by the feature mapping layer and includes a first classification prediction head and a second classification prediction head; wherein the first classification prediction head outputs the ordered classification prediction of the maximum temperature of the cable area fire smoke, and the second classification prediction head outputs the ordered classification prediction of the peak temperature of the cable steel wire;

[0024] After the pre-processing, the first type of image and the second type of image are respectively input to the first branch module and the second branch module for feature extraction, the obtained features are fused through the channel splicing layer to form the input of the shared backbone network; the shared backbone network adopts the ConvNeXt-Tiny architecture, a plurality of structure units are stacked in the order of normalization, convolution operation and nonlinear activation inside, deep feature learning is performed on the fused input layer by layer, and a high-dimensional embedding vector is output; then the high-dimensional embedding vector is input to the feature mapping layer to be 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, and the prediction results of the fire smoke peak temperature and the cable steel wire peak temperature are output respectively.

[0025] According to one embodiment of the present application, the first branch module and the second branch module are both serialized feature extraction structures, and the serialized feature extraction structure comprises a plurality of convolution units connected in sequence, each of which comprises a convolution layer, a batch normalization layer and a SiLU activation function connected in sequence;

[0026] The feature mapping layer comprises an adaptive pooling layer, a flattening layer, a linear connection layer, a normalization layer and a GELU activation layer connected in sequence;

[0027] The first classification prediction head and the second classification prediction head adopt the same processing sequence, first pass through a linear layer for feature transformation, then pass through a normalization layer and a GELU activation layer for nonlinear processing, and then pass through a linear connection layer and a temperature output module for processing to output the final prediction result.

[0028] According to one embodiment of the present application, the temperature output module of the first classification prediction head and the second classification prediction head adopts a cumulative logistic regression structure, and K 1 binary classification unit learns the cumulative probability of each temperature threshold value, and maps the output feature of the double-task branch to a temperature level prediction result with an ordered relationship, wherein K represents the total number of temperature levels, K 1 is the number of binary classifiers to be established.

[0029] According to one embodiment of the present application, the transfer learning strategy in step S4 comprises:

[0030] S41, initializing the model using the pre-trained ConvNeXt-Tiny backbone network weight on a large-scale image dataset;

[0031] S42, inputting the fire-retardant sealant image and the anticorrosive layer image after shunt preprocessing into the model in a double-flow manner, and completing forward propagation through the shared backbone network and the double-task branch, and outputting K 1 cumulative probability from the ordered classification head and combining to form a prediction result of K temperature levels;

[0032] S43, using a hybrid loss function for optimization during the training process, the hybrid loss function comprising a cumulative logistic regression loss, a cross-entropy loss and an ordered distribution constraint loss, and combining through weighting to improve the classification accuracy and order of the model;

[0033] S44, learning parameters using the training set, and evaluating and optimizing the performance on the validation set to improve the generalization ability of the model under different fire conditions.

[0034] According to one embodiment of the present application, the hybrid loss function comprises a cumulative logistic regression loss, a cross-entropy loss and an ordered distribution constraint loss;

[0035] The cumulative logistic regression loss is used to supervise K 1binary classification unit learns the cumulative probability of each temperature threshold, and the calculation formula is:

[0036]

[0037] Wherein, L coral is the 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 level categories, k is the index number of temperature threshold k=(1, 2,..., K-1), z i,k is the output of the i-th sample at the k-th binary classification unit, σ(·) is the Sigmoid function, t i,k is the cumulative label of the i-th sample at the k-th threshold;

[0038] The cross entropy loss is used to supervise the final temperature level prediction result, and the overall classification accuracy is improved, and the calculation formula is:

[0039]

[0040] Wherein, L ce is the cross entropy loss, c is the temperature level category index c=(1, 2,..., K), y i,c is the real label distribution of the i-th sample in the c-th temperature category, p i,c is the prediction probability of the i-th sample in the c-th temperature category;

[0041] The ordered distribution constraint loss is used to measure the difference between the prediction distribution and the real label distribution and maintain the continuity between the temperature levels, and the calculation formula is:

[0042]

[0043] Wherein, L emd is the ordered distribution constraint loss, j is the cumulative index of the temperature level category j=(1, 2,..., K), p i,j is the prediction probability of the i-th sample in the j-th temperature category, y i,j is the real label distribution of the i-th sample in the j-th temperature category;

[0044] The cumulative logistic regression loss, the cross entropy loss and the ordered distribution constraint loss are combined by setting the weight to form a hybrid loss function, which simultaneously optimizes the classification accuracy and the order of the temperature level of the cable fire highest temperature prediction model, and the calculation formula of the hybrid loss function is:

[0045]

[0046] Wherein, the alpha, beta and gamma are preset weight coefficients, used for balancing the classification accuracy and the ordered discriminant ability.

[0047] According to an embodiment of the present application, the step S5 comprises:

[0048] S51, the two branch prediction results output by the cable maximum fire temperature prediction model are respectively decoded to obtain an ordered probability distribution, the cumulative probability distribution is converted into a discrete probability distribution of each temperature category, and the category with the maximum probability is selected as the predicted temperature grade;

[0049] S52, the accuracy, precision, recall and F1 score calculation module is called to calculate the classification performance indicators of the prediction tasks corresponding to the main cable wire peak temperature, the sling wire peak temperature and the fire smoke peak temperature;

[0050] S53, based on the prediction probability distribution and the real label, the EMD is calculated as the ordinal consistency evaluation index to measure the matching degree of the prediction distribution of the model on the ordered temperature category;

[0051] S54, the real HC fire test image data is loaded and input into the trained cable maximum fire temperature prediction model, the cross-dataset generalization accuracy of the main cable wire, the sling wire and the fire smoke temperature prediction tasks in the real scene is calculated, and the category confusion matrix and the class activation heat map are output.

[0052] According to an embodiment of the present application, in step S6, the creation of the visualization window comprises:

[0053] S61, based on the graphical user interface framework, a special visualization window is built, the main window interface is built and the trained cable maximum fire temperature prediction model is integrated, and the function areas for image display, temperature result display and user control are designed and laid out;

[0054] S62, an image input module is designed in the visualization window, the image reading and preprocessing functions are realized by using the image processing library, single or multiple cable apparent images after fire resistance test are imported through the file dialog box or the drag and drop mode, and the input images are automatically normalized in size and converted in channel format, so as to meet the tensor requirements of the cable maximum fire temperature prediction model input;

[0055] S63, the double-branch weight sharing cable maximum fire temperature prediction model is called to perform forward inference on the preprocessed images, the model weights are loaded by using the deep learning framework and the forward propagation calculation is performed, and the prediction values of the main cable wire peak temperature and the fire smoke peak temperature are respectively output;

[0056] S64, the original image, the predicted temperature value and the temperature grade label are synchronously displayed in the visualization interface by using the graphic drawing component.

[0057] Additional aspects and advantages of the present application will be partially given in the following description, partially will become apparent from the following description, or will be learned by practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 A flow chart of a deep learning based rapid non-destructive prediction method for the maximum temperature of a cable fire according to an embodiment of the present application;

[0059] Figure 2 A network architecture diagram of a double-branch weight sharing maximum temperature prediction model for a cable fire according to an embodiment of the present application;

[0060] Figure 3 A network architecture diagram of an input branch module according to an embodiment of the present application;

[0061] Figure 4 A network architecture diagram of an ordered classification prediction head according to an embodiment of the present application;

[0062] Figure 5 A network architecture diagram of a feature mapping layer according to an embodiment of the present application;

[0063] Figure 6 A diagram of the accuracy of a prediction task corresponding to the peak temperature of a steel wire according to an embodiment of the present application;

[0064] Figure 7 A diagram of the accuracy of a prediction task corresponding to the peak temperature of a smoke gas according to an embodiment of the present application;

[0065] Figure 8 A diagram of a confusion matrix corresponding to the peak temperature of a steel wire according to an embodiment of the present application;

[0066] Figure 9 A diagram of a confusion matrix corresponding to the peak temperature of a smoke gas according to an embodiment of the present application. DETAILED DESCRIPTION

[0067] Embodiments of the present application are described in detail below with reference to the accompanying drawings, in which the same or similar components have the same or similar reference numbers throughout. The embodiments described below are exemplary and are intended to explain the present application, and should not be understood as limiting the present application.

[0068] A deep learning based rapid non-destructive prediction method for the maximum temperature of a cable fire according to an embodiment of the present application is described below with reference to the accompanying drawings.

[0069] As Figure 1As shown, the deep learning-based rapid non-destructive prediction method for the maximum temperature of a cable after fire of an embodiment of the present application can include the following steps:

[0070] S1, obtaining apparent morphology image data of each layer of material of a cable corrosion-fire protection system under different temperature conditions through various fire tests; wherein the layers of material of the cable include a fire-retardant sealant layer and a corrosion-resistant layer.

[0071] S2, performing shunt preprocessing on the image data according to the layer characteristics of the cable material; constructing a data set by combining the processed image data with corresponding temperature labels, and dividing the data set into a training set, a validation set and a test set.

[0072] S3, constructing a double-branch weight-sharing cable maximum temperature prediction model based on ConvNeXt-Tiny.

[0073] S4, loading pre-trained weights through a transfer learning strategy, optimizing the cable maximum temperature prediction model using a hybrid loss function, training the cable maximum temperature prediction model using the training set, and tuning the model through the validation set.

[0074] S5, performing multi-index joint performance evaluation on the trained cable maximum temperature prediction model using the test set, and verifying the generalization ability and actual applicability of the model through real fire test images.

[0075] S6, creating a visualization window and importing the cable fire resistance test image into the cable maximum temperature prediction model to output the maximum temperature of the cable steel wire or the maximum temperature of the fire smoke corresponding to the cable fire resistance test image, thereby realizing non-destructive and rapid evaluation of the safety performance of the cable after fire without disassembly or sensors.

[0076] Specifically, the core principle of the method of an embodiment of the present application is based on the convolutional neural network (ConvNeXt-Tiny) in deep learning and multi-source image feature fusion technology, which realizes temperature prediction by analyzing the nonlinear mapping relationship between the apparent morphological changes of the cable protection layer (fire-retardant sealant layer and corrosion-resistant layer) after fire and the maximum temperature. The working process is as follows: first, image data of each layer of material under different temperatures is obtained through various fire tests, and shunt preprocessing is performed according to the layer characteristics to construct a data set with temperature labels; then a double-branch weight-sharing cable maximum temperature prediction model is constructed, pre-trained weights are loaded using transfer learning, model parameters are optimized through a hybrid loss function, and finally high-precision learning of the correlation between image features and temperature is realized, thereby avoiding the limitations of traditional sensors or disassembly detection.

[0077] Specifically, the workflow of the method of the embodiment of the present application can be divided into three stages of data preparation, model training and verification. In the data preparation stage, i.e. steps S1 and S2, image data of each layer of cable material at different temperatures is collected through a controllable fire test, and after pre-processing according to layer position, a labeled data set is constructed and divided into a training set, a verification set and a test set; in the model training stage, i.e. steps S3 and S4, a double-branch weight-sharing cable maximum temperature prediction model based on ConvNeXt-Tiny is constructed, the parameters are initialized through transfer learning, an end-to-end training is performed using a hybrid loss function (such as combining mean square error and perception loss), and the hyperparameters are adjusted using the verification set to prevent overfitting; in the verification and application stage, i.e. steps S5 and S6, the performance is evaluated through multiple indicators (such as MAE, RMSE) of the test set, and the generalization ability is verified using real fire images, and finally deployed to a visualization window to realize non-destructive temperature prediction.

[0078] Thus, the method of the embodiment of the present application realizes non-destructive and rapid prediction of the maximum temperature of the cable after fire, without the need to disassemble or install sensors, greatly reducing the evaluation cost and time; through the double-branch structure and the transfer learning strategy, the sensitivity of the model to small morphological changes and the prediction accuracy (the error can be controlled within ±5℃) are improved; it has strong generalization ability and can adapt to complex and variable scenes in actual fires, providing reliable technical support for the safety performance evaluation of cables after fire.

[0079] According to an embodiment of the present application, step S2 comprises:

[0080] S21, according to the layer position characteristics of the cable material, the image is divided into two types, the flame-retardant sealant image is taken as the first type of image for determining the maximum temperature of fire smoke, and the anticorrosion layer image is taken as the second type of image for determining the maximum temperature of steel wire;

[0081] S22, the first type of image is subjected to grayscale and texture enhancement to eliminate the interference of fire smoke on color, highlight the typical morphological features of fire, and the typical morphological features of fire include cracks, bubbles and net-like fractures;

[0082] S23, the second type of image is retained with RGB color information, subjected to color normalization and moderate contrast correction, and the color change rule under the action of temperature is extracted;

[0083] S24, the first type of image and the second type of image are subjected to size scaling and normalization, combined with geometric enhancement and controlled brightness / contrast disturbance, the diversity of training samples is increased, and it is ensured that the key features are not destroyed; wherein the geometric enhancement includes rotation, scaling and flipping;

[0084] S25, bind the processed image with the corresponding temperature level label, construct the data set of differentiated features, and divide the data set into a training set, a validation set and a test set in a hierarchical manner according to the test piece and the working condition, so as to avoid data leakage and keep the consistency of the distribution of each temperature level.

[0085] Thus, by means of the shunt processing and the differentiated preprocessing strategy, the correlation between the smoke shape feature and the anticorrosion layer color feature and the temperature is specifically strengthened, the accuracy and reliability of the subsequent model prediction are improved, secondly, by means of the geometric enhancement and the controlled disturbance operation, the diversity of the training samples is effectively expanded, and the generalization ability and the anti-interference of the model are enhanced, thirdly, by means of the hierarchical data set division manner, the data leakage risk is eliminated, the scientificity and the actual applicability of the model evaluation result are ensured, and a solid foundation is laid for the efficient training of the subsequent double-branch model.

[0086] According to one embodiment of the present application, as shown in Figure 2 the cable fire highest temperature prediction model comprises two parallel input branch modules, a channel splicing layer, a shared main network, a feature mapping layer and two task-specific ordered classification prediction heads;

[0087] The two parallel input branch modules comprise a first branch module and a second branch module; the first branch module is used for processing the apparent image of the cable flame-retardant sealing rubber layer and extracting the features related to the texture; the second branch module is used for processing the apparent image of the cable anticorrosion layer and extracting the features related to the color change and state evolution;

[0088] The ordered classification prediction head is used for receiving the feature vector output by the feature mapping layer and comprises a first classification prediction head and a second classification prediction head; wherein the first classification prediction head outputs the ordered classification prediction of the cable area fire smoke highest temperature, and the second classification prediction head outputs the ordered classification prediction of the cable steel wire peak temperature;

[0089] After the preprocessing, the first type of image and the second type of image are respectively input to the first branch module and the second branch module for feature extraction, the obtained features are fused through the channel splicing layer to form the input of the shared main network; the shared main network adopts the ConvNeXt-Tiny architecture, a plurality of structure units are stacked in the order of normalization, convolution operation and nonlinear activation inside, deep representation learning is performed on the fused input layer by layer, and a high-dimensional embedding vector is output; then the high-dimensional embedding vector is input to the feature mapping layer to be 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, and the prediction results of the fire smoke peak temperature and the cable steel wire peak temperature are output respectively.

[0090] According to one embodiment of the present application, the first branch module and the second branch module are both serialized feature extraction structures, and the serialized feature extraction structure comprises a plurality of convolution units connected in sequence, each of which comprises a convolution layer, a batch normalization layer and a SiLU activation function connected in sequence. As shown in Figure 3 each of which comprises a convolution layer, a batch normalization layer and a SiLU activation function connected in sequence; the parameters of each convolution layer include input channel number In_c, output channel number Out_c, convolution kernel size k, stride s and padding size p; the first convolution unit is used for primary feature extraction, the parameters of the first convolution layer of the first branch module are In_c=1, Out_c=32, k=3, s=1 and p=1, and the parameters of the first convolution layer of the second branch module are In_c=3, Out_c=32, k=3, s=1 and p=1; the second convolution unit is used for down-sampling and channel expansion, and the parameters of the second convolution layer of the first branch module and the second branch module are all In_c=32, Out_c=48, k=3, s=2 and p=1; the third convolution unit is used for feature refinement and output, and the parameters of the third convolution layer of the first branch module and the second branch module are all In_c=48, Out_c=48, k=3, s=1 and p=1. After processing by the three convolution units, the first branch module and the second branch module respectively output feature maps with consistent size and same number of channels, which are used for subsequent channel splicing and shared backbone network.

[0091] The feature mapping layer comprises an adaptive pooling layer, a flattening layer, a linear connection layer, a normalization layer and a GELU activation layer connected in sequence. As shown in Figure 5 the linear connection layer is used for channel dimension mapping, and the parameters thereof are input channel number In_c=96 and output channel number Out_c=768, realizing mapping from the spliced features to a high-dimensional embedding space.

[0092] The first classification prediction head and the second classification prediction head both adopt the same processing sequence, first performing feature transformation through a linear layer, then performing nonlinear processing through a normalization layer and a GELU activation layer, and finally performing processing through a linear connection layer and a temperature output module to output the final prediction result.

[0093] Specifically, as shown in Figure 4 the first classification prediction head and the second classification prediction head both adopt the same processing sequence, first performing feature transformation through a first-level linear connection layer, then performing nonlinear processing through a normalization layer and a GELU activation layer, and finally converting K 1 logits, output final prediction results; wherein, the input channel number of the first linear connection layer is In_c=768, the output channel number is Out_c=384, which is used for compressing high-dimensional embedding features and extracting intermediate representations; the input channel number of the second linear connection layer is In_c=384, and the output channel number is Out_c=6, which is used for outputting feature vectors corresponding to temperature levels.

[0094] According to an embodiment of the application, the temperature output modules of the first classification prediction head and the second classification prediction head both adopt a cumulative logistic regression structure, and the K 1 binary classification unit learns the cumulative probability of each temperature threshold, and maps the output features of the double-task branch to temperature level prediction results with an ordered relationship, wherein K represents the total number of temperature levels, K 1 is the number of binary discriminators to be established.

[0095] According to an embodiment of the application, the transfer learning strategy in step S4 includes:

[0096] S41, using the pre-trained ConvNeXt-Tiny backbone network weight on a large-scale image dataset to initialize the model;

[0097] S42, inputting the branched preprocessed fire-retardant sealant images and anticorrosive layer images into the model in a double-flow manner, and completing forward propagation through the shared backbone network and double-task branch, and outputting K 1 cumulative probabilities from the ordered classification head and combining to form prediction results of K temperature levels;

[0098] S43, using a hybrid loss function for optimization during the training process, the hybrid loss function including a cumulative logistic regression loss, a cross-entropy loss and an ordered distribution constraint loss, and improving the classification accuracy and order of the model through weighted combination;

[0099] S44, using the training set to learn parameters, and performing performance evaluation and optimization on the validation set to improve the generalization ability of the model under different fire conditions.

[0100] According to an embodiment of the application, the hybrid loss function includes a cumulative logistic regression loss, a cross-entropy loss and an ordered distribution constraint loss.

[0101] The cumulative logistic regression loss is used to supervise the K 1 binary classification units to learn the cumulative probability of each temperature threshold, and the calculation formula is:

[0102]

[0103] Wherein, L coralN is the total number of samples, i is the sample index number i = (1, 2,..., N), K is the total number of temperature level categories, k is the index number of the temperature level threshold value k = (1, 2,..., K-1), z i,k is the output of the i-th sample at the k-th binary unit, and σ(·) is the Sigmoid function, t i,k is the cumulative label of the i-th sample at the k-th threshold value.

[0104] The cross-entropy loss is used to supervise the final temperature level prediction result and improve the overall classification accuracy, and the calculation formula is:

[0105]

[0106] where L ce is the cross-entropy loss, c is the temperature level category index c = (1, 2,..., K), y i,c is the real label distribution of the i-th sample at the c-th temperature category, and p i,c is the prediction probability of the i-th sample at the c-th temperature category.

[0107] The ordered distribution constraint loss is used to measure the difference between the prediction distribution and the real label distribution and maintain the continuity between the temperature levels, and the calculation formula is:

[0108]

[0109] where L emd is the ordered distribution constraint loss, j is the cumulative index of the temperature level category j = (1, 2,..., K), p i,j is the prediction probability of the i-th sample at the j-th temperature category, and y i,j is the real label distribution of the i-th sample at the j-th temperature category.

[0110] The cumulative logistic regression loss, the cross-entropy loss and the ordered distribution constraint loss are combined by setting weights to form a hybrid loss function, which simultaneously optimizes the classification accuracy and the temperature level order of the cable fire highest temperature prediction model, and the calculation formula of the hybrid loss function is:

[0111]

[0112] where α, β, γ are preset weight coefficients for balancing the classification accuracy and the ordered discrimination ability.

[0113] According to an embodiment of the present application, step S5 comprises:

[0114] S51, the two branches of the cable maximum fire temperature prediction model output are respectively decoded to obtain ordered probability distribution, the cumulative probability distribution is converted into discrete probability distribution of each temperature category, and the category with the maximum probability is selected as the predicted temperature grade.

[0115] S52, an accuracy, precision, recall and F1 score calculation module is called to calculate the classification performance indicators of the predicted tasks corresponding to the main cable wire peak temperature, the sling wire peak temperature and the fire smoke peak temperature.

[0116] For example, the accuracy of the predicted task corresponding to the wire peak temperature is as shown in Table 1. Figure 6 The accuracy of the predicted task corresponding to the smoke peak temperature is as shown in Table 1. Figure 7 The classification indicators of the cable maximum fire temperature prediction model in the test set are as shown in Table 1.

[0117] Table 1

[0118]

[0119] S53, based on the predicted probability distribution and the real label, the EMD is calculated as the ordinal consistency evaluation indicator, and the matching degree of the prediction distribution of the model on the ordered temperature category is measured;

[0120] S54, the real HC fire test image data is loaded and input into the trained cable maximum fire temperature prediction model, the cross-dataset generalization accuracy of the main cable wire, the sling wire and the fire smoke temperature prediction tasks in the real scene is calculated, and the category confusion matrix and the class activation heat map are output.

[0121] For example, the confusion matrix corresponding to the wire peak temperature is as shown in Table 2. Figure 8 The confusion matrix corresponding to the smoke peak temperature is as shown in Table 2. Figure 9 The classification indicators of the cable maximum fire temperature prediction model on the real fire image information are as shown in Table 2.

[0122] Table 2

[0123]

[0124] According to one embodiment of the present application, in step S6, creating a visualization window comprises:

[0125] S61, a special visualization window is built based on a graphical user interface framework, a main window interface is built and the trained cable maximum fire temperature prediction model is integrated, and a functional area for image display, temperature result display and user control is designed and laid out;

[0126] S62, design an image input module in the visualization window, realize image reading and preprocessing functions by using an image processing library, support importing single or multiple apparent images after cable fire resistance test through a file dialog box or a drag-and-drop way, and perform automatic size normalization and channel format conversion on the input images, so that they meet the tensor requirements of the cable maximum temperature prediction model input;

[0127] S63, call the double-branch weight sharing cable maximum temperature prediction model to perform forward inference on the preprocessed images, load the model weights by using a deep learning framework and perform forward propagation calculation, and respectively output the predicted values of the peak temperature of the main cable steel wire and the peak temperature of the fire smoke;

[0128] S64, synchronously display the original images, the predicted temperature values and the temperature grade labels in the visualization interface through a graphical drawing component.

[0129] Specifically, a special visualization window can be built based on a Python graphical interface framework, a main window interface is built by using a PyQt or Tkinter framework, the trained cable maximum temperature prediction model is integrated, an image display area, a temperature result display panel, a control button group and a heat map visualization area are designed and laid out by using a QtDesigner tool. An image input module is designed in the window, an image reading and preprocessing function is realized by using an OpenCV library, single or multiple apparent images after cable fire resistance test are imported through a QFileDialog file dialog box or a drag-and-drop event response mechanism, and automatic size normalization and channel format conversion are performed on the input images, so that they meet the tensor requirements of the cable maximum temperature prediction model input. The cable maximum temperature prediction model is called to perform forward inference on the preprocessed images, the model weights are loaded by using a PyTorch framework, and the predicted values of the peak temperature of the main cable steel wire and the peak temperature of the fire smoke are respectively output by using a forward method. In the visualization interface, original images, predicted temperature values and temperature grade labels are synchronously displayed through Matplotlib or OpenCV drawing components.

[0130] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the described specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0131] In addition, the terms "first", "second", etc. are used only for the purpose of description, and should not be understood as indicating or implying relative importance or implying a number of the technical features indicated. Therefore, the features defined as "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise explicitly specified and limited.

[0132] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting", "fixing" and the like should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise explicitly limited. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0133] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A deep learning-based rapid non-destructive prediction method for the maximum temperature of a cable over fire, characterized in that, The method comprises: S1, obtaining apparent morphology image data of each layer material of the cable corrosion-fire protection system under different temperature conditions through a plurality of fire tests; wherein the layers of cable materials include a fire-retardant sealant layer and a corrosion-resistant layer; S2, pre-process the image data according to the characteristics of the layers of cable materials; construct a data set by combining the processed image data with the corresponding temperature labels, and divide the data set into a training set, a validation set and a test set; S3, constructing a double-branch weight sharing cable maximum temperature prediction model based on ConvNeXt-Tiny; S4, loading the pre-trained weights through a transfer learning strategy, optimizing the cable maximum temperature prediction model using a hybrid loss function, training the cable maximum temperature prediction model using the training set, and optimizing it through the validation set; S5, using the test set to perform multi-index joint performance evaluation on the trained cable maximum temperature prediction model, and verifying the generalization ability and actual applicability of the model through real fire test images; S6, creating a visualization window and importing the cable fire resistance test image into the cable maximum temperature prediction model to output the cable steel wire maximum temperature or the fire smoke maximum temperature corresponding to the cable fire resistance test image, thereby realizing non-destructive and rapid evaluation of the safety performance of the cable after the fire without disassembly and sensors; wherein the cable maximum temperature prediction model comprises two parallel input branch modules, a channel concatenation 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 apparent image of the cable fire-retardant sealant layer and extract texture-related features; the second branch module is used to process the apparent image of the cable corrosion-resistant layer and extract color change and state evolution-related features; The ordered classification prediction head is used to receive the feature vectors output by the feature mapping layer and includes a first classification prediction head and a second classification prediction head; wherein the first classification prediction head outputs an ordered classification prediction of the cable area fire smoke maximum temperature, and the second classification prediction head outputs an ordered classification prediction of the cable steel wire peak temperature; After pre-processing, the first type of image and the second type of image are input into the first branch module and the second branch module respectively for feature extraction, and the obtained features are fused through the channel concatenation layer to form the input of the shared backbone network; the shared backbone network adopts the ConvNeXt-Tiny architecture, which internally stacks multiple structural units in the order of normalization, convolution operation and nonlinear activation, and performs deep feature learning on the fused input layer by layer to output a high-dimensional embedding vector; then the high-dimensional embedding vector is input into the feature mapping layer to convert it into a low-dimensional compact feature vector; the low-dimensional compact feature vector is fed into the first classification prediction head and the second classification prediction head in parallel, and the prediction results of the fire smoke peak temperature and the cable steel wire peak temperature are output respectively; The transfer learning strategy in step S4 comprises: S41, initialize the model using the pre-trained ConvNeXt-Tiny backbone network weight on a large-scale image dataset; S42, input the flame-retardant sealant image and the anti-corrosion layer image after shunt preprocessing into the model in a double-flow manner, complete forward propagation through the shared backbone network and double-task branch, and output K 1 cumulative probability and combine to form the prediction result of K temperature grades; 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 order of the model are improved through weighted combination; S44, parameter learning is performed using the training set, and performance evaluation and optimization are performed on the validation set to improve the generalization ability of the model under different fire conditions; Step S5 includes: S51, the two branch prediction results output by the cable fire highest temperature prediction model are respectively decoded as ordered probability distribution, the cumulative probability distribution is converted into discrete probability distribution of each temperature category, and the category with the maximum probability is selected as the predicted temperature grade; S52, call the accuracy, precision, recall and F1 score calculation module to calculate the classification performance indicators of the prediction tasks corresponding to the main cable steel wire peak temperature, the sling steel wire peak temperature and the fire smoke peak temperature; S53, based on the prediction probability distribution and the real label, calculate EMD as the ordinal consistency evaluation index to measure the matching degree of the prediction distribution of the model on the ordered temperature category; S54, load the real HC fire test image data and input it into the trained cable fire highest temperature prediction model, calculate the cross-dataset generalization accuracy of the main cable steel wire, the sling steel wire and the fire smoke temperature prediction tasks in the real scene, and output the category confusion matrix and class activation heat map.

2. The deep learning based rapid non-destructive prediction method of the cable overfire maximum temperature according to claim 1, characterized in that, Step S2 includes: S21, according to the layer position characteristics of the cable material, the image is divided into two types, the first type of image is the flame-retardant sealant image, which is used for determining the highest temperature of fire smoke, and the second type of image is the anticorrosion layer image, which is used for determining the highest temperature of steel wire; S22, the first type of image is grayed and texture enhanced to eliminate the interference of fire smoke on color, highlight the typical morphological features of fire, including cracks, bubbles and net-like fractures; S23, the second type of image retains RGB color information, performs color normalization and moderate contrast correction, and extracts the color change rule under the action of temperature; S24, the first type of image and the second type of image are scaled and normalized, combined with geometric enhancement and controlled brightness / contrast disturbance to increase the diversity of training samples and ensure that key features are not damaged; wherein the geometric enhancement includes rotation, scaling and flipping; S25, the processed image and the corresponding temperature grade label are bound to construct a data set with differentiated features, and the data set is divided into training set, validation set and test set in a layered manner according to the test piece and working condition to avoid data leakage and maintain the consistency of the distribution of each temperature grade.

3. The deep learning based rapid non-destructive prediction method of the cable overfire maximum temperature according to claim 1, characterized in that, The first branch module and the second branch module are both serialized feature extraction structures, and the serialized feature extraction structure includes a plurality of convolution units connected in sequence, each convolution unit includes a convolution layer, a batch normalization layer and a SiLU activation function connected in sequence; The feature mapping layer comprises an adaptive pooling layer, a flattening layer, a linear connection layer, a normalization layer and a GELU activation layer connected in sequence. The first classification prediction head and the second classification prediction head adopt the same processing sequence, first pass through a linear layer for feature transformation, then pass through a normalization layer and a GELU activation layer for nonlinear processing, and then pass through a linear connection layer and a temperature output module for processing to output the final prediction result.

4. The deep learning based rapid non-destructive prediction method of the maximum temperature of a cable subjected to fire according to claim 3, characterized in that, The temperature output modules of the first classification prediction head and the second classification prediction head both adopt a cumulative logistic regression structure, and the outputs of the two task branches are mapped to temperature level prediction results with an ordered relationship through K 1binary classification units learn the cumulative probabilities of respective temperature thresholds, and map the output features of the double-task branches to temperature level prediction results with an ordered relationship, where K 1is the number of binary classification discriminators to be established.

5. The deep learning based rapid non-destructive prediction method of the cable overfire maximum temperature according to claim 1, characterized in that, The mixed loss function comprises a cumulative logistic regression loss, a cross-entropy loss and an ordered distribution constraint loss. The cumulative logistic regression loss is used to supervise K 1 binary classification unit learns the cumulative probability of each temperature threshold, and the calculation formula is: wherein L coral is the 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 categories of temperature levels, k is the index number of the threshold value of each temperature level k = (1, 2,..., K-1), z i,k is the output of the i-th sample at the k-th binary classification unit, σ(·) is a Sigmoid function, t i,k is the cumulative label of the i-th sample at the k-th threshold value; The cross-entropy loss is used to supervise the final temperature level prediction result and improve the overall classification accuracy, and the calculation formula is: wherein L ce is the cross-entropy loss, c is the temperature level class index number c = (1, 2, …, K), y i,c is the true label distribution of the i-th sample in the c-th temperature class, p i,c is the predicted probability of the i-th sample in the c-th temperature class. The ordered distribution constraint loss is used to measure the difference between the predicted distribution and the real label distribution and maintain the continuity between the temperature levels, and the calculation formula is: wherein L emd is an ordered distribution constraint loss, j is an accumulated index j = (1, 2, …, K) of temperature level categories, p i,j is a predicted probability of the i-th sample on the j-th temperature category, y i,j is a real label distribution of the i-th sample on the j-th temperature category; The cumulative logistic regression loss, the cross-entropy loss and the ordered distribution constraint loss are combined by setting weights to form a mixed loss function, which simultaneously optimizes the classification accuracy and the ordered nature of the temperature level of the cable fire highest temperature prediction model, and the calculation formula of the mixed loss function is: Wherein, α, β, γ are preset weight coefficients for balancing the classification accuracy and the ordered nature of the temperature level.

6. The deep learning based rapid non-destructive prediction method of the maximum cable fire temperature according to claim 1, characterized in that, In step S6, the creation of the visualization window comprises: S61, based on the graphical user interface framework, build a special visualization window, build a main window interface and integrate the trained cable fire highest temperature prediction model, design and layout the function areas for image display, temperature result display and user control; S62, design an image input module in the visualization window, use an image processing library to realize image reading and preprocessing functions, support importing single or multiple cable apparent images after fire resistance test through a file dialog box or a drag-and-drop method, and perform automatic size normalization and channel format conversion on the input image to make it meet the tensor requirements of the cable fire highest temperature prediction model input; S63, call the double-branch weight-sharing cable fire highest temperature prediction model to perform forward inference on the preprocessed image, load the model weight through a deep learning framework and perform forward propagation calculation, and output the predicted values of the main cable wire peak temperature and the fire smoke peak temperature respectively; S64, display the original image, the predicted temperature value and the temperature level label synchronously in the visualization interface through a graphical drawing component.

Images