Machine learning based test method for fire-retardant, heat-insulating, and electrically insulating cable fire blankets

Abstract

The application discloses a kind of based on machine learning cable fireproof blanket flame-retardant heat-insulating electric insulating test method, the method includes: obtaining image data and electric heat data in the combustion process of cable fireproof blanket, after noise filtration, integrated as multimodal test data set according to time stamp order;Multimodal test data set is reduced to dimension, space-time feature screening is carried out to multiple sources, cross-dimension feature extraction and weight distribution are completed by lightweight convolutional neural network with attention mechanism, and generate multimodal fusion feature;Multimodal fusion deep learning model is constructed, and adaptive learning rate optimization algorithm is used to complete iterative training and parameter optimization, and the optimal test model is obtained;Using optimal test model, directly outputting flame-retardant, heat-insulating, electric insulating performance test result.The application realizes multimodal data fusion and three performances synchronous evaluation, solves the problem that existing test independent evaluation, precision is insufficient, depends on artificial interpretation, improves test objectivity, precision and efficiency.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a test method for flame-retardant, heat-insulating, and electrical insulation of cable fire blankets based on machine learning. Background Technology

[0002] Cable fire blankets are passive fire protection materials wrapped around cables. Their function is to delay or prevent damage to cables from flames during a fire by utilizing their flame-retardant, heat-insulating, and electrical-insulating properties. This ensures the continued operation of power and signal transmission systems under fire conditions, buying valuable time for personnel evacuation and fire rescue. Therefore, accurate and comprehensive testing and evaluation of the flame-retardant, heat-insulating, and electrical-insulating properties of cable fire blankets are crucial for ensuring their reliability and engineering applicability.

[0003] Currently, performance testing of cable fire blankets primarily relies on combustion test methods specified in national standard GB / T 18380 or international standard IEC 60332. During testing, the fire blanket is typically wrapped around a standard cable and placed in a flame at a specified temperature for combustion. Individual indicators of flame retardancy, heat insulation, and electrical insulation performance are then obtained through manual observation or single-sensor measurements. However, these three properties are coupled and synergistic in actual fire environments. Evaluating them separately lacks a systematic method that can simultaneously output combined test results for multiple properties. This results in test results that fail to accurately reflect the comprehensive protective effectiveness of the fire blanket in fire scenarios. Reliance on manual interpretation and experience-based assessment results in insufficient objectivity and accuracy. Furthermore, it lacks consideration of the changes in material performance over time.

[0004] Therefore, it is necessary to propose a solution to improve one or more problems existing in the above-mentioned related technical solutions.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this disclosure is to provide a test method for the flame retardant, heat-insulating, and electrical insulation properties of cable fire blankets based on machine learning, thereby overcoming, at least to some extent, one or more problems caused by the limitations and defects of related technologies.

[0007] A machine learning-based testing method for the flame retardant, heat-insulating, and electrical insulation properties of cable fire blankets, according to an embodiment of this disclosure, includes the following steps: Image data and electrothermal data of the cable fire blanket during combustion were acquired, noise was filtered out, and the data were integrated into a multimodal test dataset in time stamp order. Feature extraction is performed on the multimodal test dataset to generate multimodal fusion features; wherein, the feature extraction includes multi-source data dimensionality reduction, spatiotemporal feature filtering, cross-dimensional feature extraction through a lightweight convolutional neural network with attention mechanism, and weight allocation through attention mechanism; A multimodal fusion deep learning model is constructed. The multimodal fusion features are input into the model, and an adaptive learning rate optimization algorithm is used to iteratively train and optimize the model to obtain the optimal test model. After noise filtering and feature extraction, the image data and electrothermal data of the fireproof blanket for the cable under test are input into the optimal test model, and the performance test results of the fireproof blanket for the cable under test are output.

[0008] In an exemplary embodiment of this application, the image data includes infrared thermal imaging data and high-speed camera data during the combustion of the cable fire blanket; the electrothermal data includes insulation resistance, withstand voltage, thermal conductivity, surface temperature, and internal temperature during the combustion of the cable fire blanket.

[0009] In an exemplary embodiment of this application, the step of acquiring image data and electrothermal data during the combustion process of the cable fire blanket, filtering out noise from them, and integrating them into a multimodal test dataset in time-stamp order includes: Infrared thermal imaging data is acquired using an infrared thermal imager, and high-speed video data is acquired using a high-speed camera. Insulation resistance and withstand voltage values ​​were collected using an insulation resistance tester and a withstand voltage tester, respectively. Thermal conductivity was collected using a heat flow meter. Surface temperature and internal temperature were collected using multiple thermocouples installed on the surface and inside the cable fire blanket, respectively.

[0010] In an exemplary embodiment of this application, the step of acquiring image data and electrothermal data during the combustion process of the cable fire blanket, filtering out noise from them, and integrating them into a multimodal test dataset in time-stamp order further includes: A variational autoencoder model is constructed for the infrared thermal imaging data, high-speed camera data, and electrothermal data consisting of insulation resistance, withstand voltage, thermal conductivity, surface temperature, and internal temperature, respectively. Each variational autoencoder model encodes and decodes its corresponding data to reconstruct it. Isolated data points with reconstruction errors greater than a preset threshold are filtered out as noise components to obtain clean image data and clean electrothermal data. Using the ignition time of the combustion test as the base timestamp, the frame rate and sampling rate of the pure image data and pure electrothermal data are time-aligned using an interpolation algorithm and then stitched together along the time dimension to form the multimodal test dataset.

[0011] In an exemplary embodiment of this application, the step of constructing a variational autoencoder model for the infrared thermal imaging data, high-speed camera data, and electrothermal data composed of insulation resistance, withstand voltage, thermal conductivity, surface temperature, and internal temperature respectively includes: A first variational autoencoder model is constructed for the infrared thermal imaging data. The encoder of the first variational autoencoder model is composed of a multi-layer convolutional neural network. It extracts the spatial temperature distribution features in the infrared thermal image step by step and compresses the features into a low-dimensional latent vector. The decoder of the first variational autoencoder model is composed of a multi-layer deconvolutional neural network. It is used to reconstruct an infrared thermal image with the same size as the input from its low-dimensional latent vector. A second variational autoencoder model is constructed for the high-speed camera data. The encoder of the second variational autoencoder model is composed of a multi-layer convolutional neural network. It extracts texture, edge and deformation features in the visible light image and compresses the features into a low-dimensional latent vector. The decoder of the second variational autoencoder model is composed of a multi-layer deconvolutional neural network. It is used to reconstruct a visible light image with the same size as the input from its low-dimensional latent vector. A third variational autoencoder model is constructed for the electrothermal data. The encoder and decoder of the third variational autoencoder model are both composed of multi-layer fully connected networks. By capturing the crossover and temporal evolution between insulation resistance, withstand voltage, thermal conductivity, surface temperature and internal temperature in the electrothermal data and compressing them into a low-dimensional latent vector, the electrothermal data consistent with the input dimension is reconstructed from its low-dimensional latent vector.

[0012] In an exemplary embodiment of this application, feature extraction is performed on the multimodal test dataset to generate multimodal fusion features; wherein the feature extraction includes the steps of multi-source data dimensionality reduction, spatiotemporal feature filtering, cross-dimensional feature extraction using a lightweight convolutional neural network with an attention mechanism, and weight allocation using an attention mechanism, including: Principal component analysis was used to reduce the dimensionality of the image modal data in the multimodal test dataset, and an autoencoder was used to reduce the dimensionality of the electrothermal modal data to obtain a low-dimensional feature set. A slow feature analysis algorithm based on a sliding time window is used to process the low-dimensional feature set and extract spatiotemporal synchronous features that characterize the slow changes in material properties over time. A lightweight convolutional neural network based on an attention mechanism is constructed. The lightweight convolutional neural network is composed of multiple depthwise separable convolutional layers and point convolutional layers stacked alternately. The spatiotemporal synchronization features are input into the lightweight convolutional neural network to obtain initial fusion features. A channel weight map is generated through a channel attention mechanism, and a spatial weight map is generated through a spatial attention mechanism. The initial fusion feature is then multiplied element-wise with the channel weight map and the spatial weight map to obtain the multimodal fusion feature.

[0013] In an exemplary embodiment of this application, the loss function of the slow feature analysis algorithm is... for: in, For time step index, The total number of time steps. and The first Time and the The low-dimensional feature vector at time step 1. It is a feature mapping function; The initial fusion features are multiplied element-wise with the channel attention weight map and the spatial attention weight map to obtain the multimodal fusion features. The calculation method is as follows: in, Indicates the initial fusion features, This represents the channel attention weight map. Represents the spatial attention weight map. This indicates an element-wise multiplication operation.

[0014] In an exemplary embodiment of this application, the step of constructing a multimodal fusion deep learning model, inputting the multimodal fusion features into the model, and using an adaptive learning rate optimization algorithm to iteratively train and optimize the model to obtain the optimal test model includes: A temporal dual-stream hybrid network is constructed as the multimodal fusion deep learning model. The temporal dual-stream hybrid network includes a recurrent neural network stream composed of gated recurrent units and a temporal convolutional network stream composed of temporal convolutional networks. The recurrent neural network stream and the temporal convolutional network stream are set in parallel. The multimodal fusion features are simultaneously input into the recurrent neural network stream and the temporal convolutional network stream. The recurrent neural network stream is used to extract temporal dependent features, and the temporal convolutional network stream is used to capture long-term temporal patterns. The hidden state sequence output by the recurrent neural network stream is concatenated with the temporal feature map output by the temporal convolutional network stream to obtain a hybrid temporal feature map. The hybrid temporal feature map is input into the fully connected layer and the Softmax layer, and the joint probability distribution of flame retardant rating, thermal insulation rating and electrical insulation rating is output. An Adam-based adaptive learning rate optimization algorithm is used to minimize the cross-entropy loss between the joint probability distribution and the true label as the training objective. The weight parameters of the temporal two-stream hybrid network are iteratively updated until the model converges, thus obtaining the optimal test model.

[0015] In an exemplary embodiment of this application, the cross-entropy loss function for: in, Indicates the total number of performance categories. Indicates performance category, Indicates the level number within the performance category. Indicates the first The total number of performance levels for each class. The one-hot encoded value of the real label. This represents the predicted probability value output by the temporal two-stream hybrid network.

[0016] In an exemplary embodiment of this application, the step of inputting the image data and electrothermal data of the fire blanket for the cable under test, after noise filtering and feature extraction, into the optimal test model and outputting the performance test results of the fire blanket for the cable under test includes: The multimodal fusion features of the fireproof blanket of the cable under test are input into the optimal test model, and a joint test result vector containing flame retardant rating, heat insulation rating and electrical insulation rating is output. When any level in the joint test result vector is lower than the preset safety threshold, an early warning signal is generated, and the spatiotemporal synchronization feature corresponding to that moment is reversed and mapped to the defect area in the original image data for visual marking.

[0017] This application proposes a machine learning-based testing method for the flame retardant, heat-insulating, and electrical insulation properties of cable fire blankets. On the one hand, by integrating denoised multimodal data and extracting fusion features, the coupling correlation and spatiotemporal variation of the three performance parameters are realized, solving the problem of single test data and independent performance evaluation. On the other hand, by optimizing the deep learning model, intelligent evaluation is achieved, replacing manual interpretation, improving test accuracy and efficiency, and truly reflecting the comprehensive protective effectiveness of the fire blanket. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0019] Figure 1 The diagram illustrates the steps of a machine learning-based test method for flame-retardant, heat-insulating, and electrical insulation of cable fire blankets in an exemplary embodiment of this application. Figure 2 This diagram illustrates a test method flow chart for flame-retardant, heat-insulating, and electrical insulation of cable fire blankets based on machine learning, as shown in an exemplary embodiment of this application. Detailed Implementation

[0020] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0021] Furthermore, the accompanying drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0022] This example implementation first provides a machine learning-based testing method for the flame retardant, heat-insulating, and electrical insulation properties of cable fire blankets. This method can be applied to a terminal device, such as a mobile terminal like a mobile phone, desktop computer, personal digital assistant, laptop, tablet, or smartwatch. (Reference) Figure 1 , Figure 2 As shown, the method may include the following steps: Step S101: Acquire image data and electrothermal data during the burning process of the cable fire blanket, filter out noise, and integrate them into a multimodal test dataset in time stamp order.

[0023] Step S102: Extract features from the multimodal test dataset to generate multimodal fusion features; wherein, the feature extraction includes multi-source data dimensionality reduction, spatiotemporal feature filtering, cross-dimensional feature extraction through a lightweight convolutional neural network with attention mechanism, and weight allocation through attention mechanism.

[0024] Step S103: Construct a multimodal fusion deep learning model, input the multimodal fusion features into the model, and use an adaptive learning rate optimization algorithm to iteratively train and optimize the model to obtain the optimal test model.

[0025] Step S104: After the image data and electrothermal data of the fireproof blanket of the cable to be tested are filtered out by noise and extracted by feature extraction, they are input into the optimal test model and the performance test results of the fireproof blanket of the cable to be tested are output.

[0026] This application proposes a machine learning-based testing method for the flame retardant, heat-insulating, and electrical insulation properties of cable fire blankets. On the one hand, by integrating denoised multimodal data and extracting fusion features, the coupling correlation and spatiotemporal variation of the three performance parameters are realized, solving the problem of single test data and independent performance evaluation. On the other hand, by optimizing the deep learning model, intelligent evaluation is achieved, replacing manual interpretation, improving test accuracy and efficiency, and truly reflecting the comprehensive protective effectiveness of the fire blanket.

[0027] Below, as Figures 1-2 As shown, a more detailed explanation will be given of the test method for flame retardant, heat insulation and electrical insulation of cable fire blankets based on machine learning proposed in this example embodiment.

[0028] Step S101: Acquire image data and electrothermal data during the burning process of the cable fire blanket, filter out noise, and integrate them into a multimodal test dataset in time stamp order.

[0029] It is important to understand that during the combustion process, the changes in the flame-retardant, heat-insulating, and electrical insulation properties of cable fire blankets are manifested through various physical fields. For example, visible light images reflect the flame morphology and material deformation, infrared thermal images reflect the temperature field distribution, and electrothermal parameters reflect changes in the internal electrical and thermal states. These physical quantities are inherently causally related over time, therefore, they need to be collected synchronously at the same time reference to fully capture the entire process of performance degradation.

[0030] Combustion testing environments are harsh, and the collected data inevitably contains various noises, such as dead pixels from infrared detectors, momentary smoke obstruction from high-speed camera lenses, and pulse interference caused by poor sensor contact. These noises statistically manifest as isolated outliers deviating from the normal data manifold. A variational autoencoder (VAE) is a deep generative model that learns the distribution patterns of a large amount of normal data, compressing input data into a low-dimensional latent space and reconstructing it from there. For a VAE that has learned the normal data distribution, normal data points can be reconstructed with high fidelity, while noise points, because they do not conform to the learned distribution, will have significantly increased reconstruction errors. By setting a reconstruction error threshold, noise components can be effectively identified and filtered out.

[0031] Different sensors have different sampling frequencies. Image data typically has a high frame rate, while electrothermal data has a relatively low sampling rate. In order to establish a correspondence between visual phenomena occurring at the same moment and changes in electrothermal state, it is necessary to use the ignition moment of the combustion test as a global reference timestamp. By using an interpolation algorithm, data with different sampling rates are unified onto a common time grid, thereby forming a multimodal data unit containing complete image and electrothermal information at each time point.

[0032] Specifically, a multimodal data synchronous acquisition system is constructed. The fire blanket for the cable to be tested is wrapped around a standard cable according to standard requirements and placed in a combustion test apparatus conforming to GB / T 18380 or IEC 60332 standards. Infrared thermal imagers and high-speed cameras are deployed around the test apparatus. The infrared thermal imager is aimed at the fire-exposed and unexposed sides of the fire blanket to acquire infrared thermal imaging data of the two-dimensional temperature field distribution. The high-speed camera is used to record high-speed video data of visible light visual information such as flame morphology, material carbonization, cracking, and deformation. Simultaneously, multiple surface thermocouples are deployed at different positions on the unexposed side of the cable fire blanket, and multiple internal thermocouples are pre-embedded in its internal interlayer to collect changes in surface and internal temperatures over time. Insulation resistance testers and withstand voltage testers are installed to continuously measure the insulation resistance and withstand voltage values ​​of the fire blanket. A steady-state heat flow meter is installed to measure the thermal conductivity passing through the fire blanket. All of the above-mentioned acquisition devices are connected to a synchronous trigger controller, which issues a unified trigger signal at the moment of ignition during the combustion test, causing each device to start recording synchronously and marking all data frames and sampling points with a unified timestamp with the ignition time as zero.

[0033] After completing one or more standard combustion tests and collecting sufficient infrared thermal imaging data, high-speed camera data, and electrothermal data consisting of insulation resistance, withstand voltage, thermal conductivity, surface temperature, and internal temperature, the noise filtering stage begins. A variational autoencoder model is constructed for each of these three types of data: the first variational autoencoder model, the second variational autoencoder model, and the third variational autoencoder model.

[0034] The first variational autoencoder model targets infrared thermal imaging data. Its encoder consists of a multi-layer convolutional neural network, extracting spatial temperature distribution features from the infrared thermal image layer by layer, ultimately compressing them into a low-dimensional latent vector. Its decoder consists of a multi-layer deconvolutional neural network, reconstructing an infrared thermal image with the same size as the input from this low-dimensional latent vector. During training, a large number of unprocessed infrared thermal images collected at the test site are used as the training set. The joint training objective is to minimize the mean square error between the original and reconstructed infrared thermal images, as well as the KL divergence between the latent distribution and the standard normal distribution. Backpropagation optimizes the network parameters, enabling the variational autoencoder to learn the normal evolution pattern of the temperature field of the cable fire blanket under combustion conditions, such as the smooth diffusion of heat from the fire-exposed side to the unexposed side and the continuous change of the temperature gradient. After training, the original infrared thermal imaging data is input frame by frame into the first variational autoencoder model. The model encodes and decodes each frame, reconstructing the image and calculating the absolute error of each pixel position between the input and reconstructed frames. Pixels with an absolute error greater than a first preset threshold are identified as single-pixel abnormal high temperature caused by detector bad pixels or stripe noise caused by electromagnetic interference. These pixels are removed from the original frame or filled with neighborhood interpolation to obtain clean infrared thermal imaging data.

[0035] The second variational autoencoder model is designed for high-speed camera data. Its encoder also consists of a multi-layer convolutional neural network, but the number of layers and channels is adjusted according to the texture complexity of the visible light image to fully extract visual features such as flame edges, carbonized material regions, and smoke diffusion patterns. The decoder structure is similar to the first variational autoencoder model. During training, unprocessed normal high-speed camera frames are used to teach the model visual manifolds such as stable flame morphology, carbonized material surface texture, and normal smoke diffusion. After training, the original high-speed camera data is input frame by frame into the second variational autoencoder model. The reconstruction error is calculated pixel by pixel. Pixel regions with errors greater than a second preset threshold are identified as noise regions caused by momentary smoke obstruction, lens reflection, or signal transmission errors and are filtered out to recover a clear high-speed camera frame sequence.

[0036] The third variational autoencoder model targets electrothermal data. Since electrothermal data is a vector composed of multiple one-dimensional time series and lacks spatial translation invariance, both the encoder and decoder are composed of multi-layer fully connected networks. This model takes the insulation resistance, withstand voltage, thermal conductivity, surface temperature, and internal temperature at the same moment as a multi-dimensional vector input. The encoder learns the cross-correlation between these parameters and their temporal evolution, compressing the input into a low-dimensional latent vector. The decoder then reconstructs the electrothermal vector with the same dimension as the input from the latent vector. During training, the model learns the coordinated changes in various electrothermal parameters under standard combustion conditions. For example, as combustion progresses, the surface temperature and internal temperature rise at a specific rate while the temperature difference remains within a certain range; the insulation resistance gradually decreases with increasing temperature and material carbonization; and the thermal conductivity changes due to material phase transitions. After training, the original electrothermal data is input into the third variational autoencoder model time by time. The overall reconstruction error between the input vector and the reconstructed vector is calculated. The electrothermal data at the time when the overall reconstruction error is greater than the third preset threshold is determined to be pulse noise caused by poor sensor contact or instantaneous thermal shock. The electrothermal vector at that time is removed from the data sequence to obtain a continuous, smooth, and pure electrothermal data sequence.

[0037] After noise filtering, the clean image data and clean electrothermal data are time-aligned and integrated. Using the ignition moment of the combustion test as the baseline timestamp zero, the time axes of the infrared thermal imaging data, high-speed camera data, and electrothermal data are unified to this baseline. Since the frame rate of image data is usually higher than the sampling rate of electrothermal data, an interpolation algorithm is used to interpolate the electrothermal data onto a dense time grid matching the image frame rate, ensuring that each image frame has a corresponding set of interpolated electrothermal data values. Alternatively, all data can be uniformly interpolated onto a preset common target time series. After time alignment, each aligned frame of image data is concatenated with the corresponding electrothermal data vector along the time dimension to form a structured multimodal test dataset. This dataset contains an infrared thermal image, a high-speed camera image, and an electrothermal data vector at each discrete time point, fully encapsulating the visual state, temperature field distribution, and electrothermal physical state information of the cable fire blanket at that moment. This provides high-quality, spatiotemporally aligned foundational data for multi-source data dimensionality reduction and synchronous extraction of spatiotemporal features.

[0038] Step S102: Extract features from the multimodal test dataset to generate multimodal fusion features; wherein, the feature extraction includes multi-source data dimensionality reduction, spatiotemporal feature filtering, cross-dimensional feature extraction through a lightweight convolutional neural network with attention mechanism, and weight allocation through attention mechanism.

[0039] Specifically, the image modal data includes infrared thermal imaging frames and high-speed camera frames, each with hundreds of thousands of pixels. Principal component analysis (PCA) is used for processing. The principle of PCA is to transform the original high-dimensional variables into a few uncorrelated low-dimensional new variables, i.e., principal components, through orthogonal transformation. Finding the principal components involves solving for the eigenvalues ​​and eigenvectors of the original data covariance matrix. Larger eigenvalues ​​indicate richer variance information contained in that direction. Projecting the original image data matrix onto the eigenvectors corresponding to the first few largest eigenvalues ​​achieves dimensionality reduction while retaining most of the main variation information. Let the image modal data matrix be X, and its covariance matrix be C. Solving for the eigenvalues ​​and eigenvectors of C, and taking the eigenvectors corresponding to the first d largest eigenvalues ​​to form the projection matrix W, then the dimensionality-reduced low-dimensional image features are... The value of d is determined by the cumulative variance contribution rate, for example, retaining more than 95% of the variance.

[0040] The electrothermal modal data consists of insulation resistance, withstand voltage, thermal conductivity, surface temperature, and internal temperature. While the number of parameters is relatively small, the dimensionality is low, and the physical units vary. Dimensionality reduction is achieved using an autoencoder, which comprises an encoder and a decoder. The encoder progressively maps the input vector to a low-dimensional latent space, while the decoder attempts to reconstruct the original vector from the latent space representation. The training objective is to minimize the reconstruction error. After training, the encoding of the latent space becomes the low-dimensional electrothermal features, capturing the most fundamental correlation patterns and temporal evolution trends among the various electrothermal parameters. The image modality dimensionality reduction results and the electrothermal modality dimensionality reduction results together constitute the low-dimensional feature set.

[0041] Although the low-dimensional feature set has reduced dimensionality, it still contains rapid oscillation components caused by short-term sensor fluctuations and environmental interference. These instantaneous changes are unrelated to the slow, inherent degradation process of material properties. A slow feature analysis algorithm based on a sliding time window is used to filter out the slowest changing components from the time-series features, because slow-changing components correspond to irreversible performance degradation caused by continuous high temperatures, such as carbonization of the insulation layer and collapse of the thermal insulation material structure.

[0042] The core optimization objective of the slow feature analysis algorithm is to minimize the mean square of the time derivatives of the feature maps output at adjacent time steps. Let the time series data be... Each of them These are feature vectors from a low-dimensional feature set. The goal is to find a set of feature mapping functions. This ensures that the mapped sequence changes as slowly as possible over time, while the output feature components are uncorrelated to each other to avoid redundancy.

[0043] The formal constraint optimization for this problem is as follows. First, the mean of each output component is required to approach zero to eliminate constant bias; the constraint condition is as follows: Secondly, the variance of each output component is required to be a unit value to avoid trivial solutions; the constraint is as follows. Finally, it is required that the different output components are uncorrelated, with the constraint being that when... hour .

[0044] Under the above constraints, it is transformed into an unconstrained optimization problem to be solved approximately, with the objective being the expected value of the squared difference between the outputs of adjacent time steps, i.e. For a discrete time series, the expected value is approximated by the sample mean; the loss function of the slow feature analysis algorithm... for: in, For time step index, The total number of time steps. and The first Time and the The low-dimensional feature vector at time step 1. This is the feature mapping function.

[0045] When eigenvalues ​​at adjacent time points are very close, their difference approaches zero, and the loss function is small. This indicates that the feature changes gradually over time, representing a property drift caused by the slow accumulation of thermal damage in the material. Conversely, if a eigencomponent changes drastically between adjacent time steps, its difference is large, and the loss function imposes a significant penalty on it. By minimizing this loss over the entire time series, the slow feature analysis algorithm automatically filters out those eigencomponents that change slowly over time, suppressing the impact of high-frequency noise and transient interference on subsequent performance evaluation. The final output reflects the spatiotemporal synchronous characteristics that reflect the material's performance degradation.

[0046] The spatiotemporal synchronization features obtained through slow feature analysis require further cross-dimensional interaction modeling for spatiotemporal correlation. First, the spatiotemporal synchronization features are reorganized according to their inherent spatial and temporal structures, forming a high-dimensional tensor containing temporal, spatial, and feature channel dimensions. This feature tensor is then input into a lightweight convolutional neural network equipped with an attention mechanism.

[0047] This lightweight convolutional neural network consists of alternating stacks of depthwise separable convolutional layers and point-level convolutional layers. Its design principle lies in decomposing standard convolution into two steps. The depthwise separable convolutional layers independently perform two-dimensional spatial convolution on each feature channel of the input tensor, extracting only the local spatial features within each channel, resulting in a computationally much lower cost than standard convolution. The point-level convolutional layers, or 1×1 convolutions, linearly combine multiple channels output from the depthwise separable convolutional layers, achieving information mixing and dimensionality transformation between channels. This alternating stacking of these two types of layers enables efficient cross-dimensional interaction between the spatial and channel domains with minimal parameters and computational cost, outputting initial fused features.

[0048] Finally, by using a parallel mechanism of channel attention and spatial attention, weights are adaptively assigned to the initial fused features, enabling the model to focus on the feature dimensions and spatial regions that are most critical to performance evaluation.

[0049] The channel attention branch is used to learn which feature channels are more important to the current task. Global average pooling and global max pooling are performed on the initial fused features in the spatial dimension, resulting in two one-dimensional vectors. These two vectors are input into a multilayer perceptron with shared weights, and after activation, they are summed to obtain the channel attention weight map. This weight map captures the dependencies between channels, such as increasing the weight of channels with decreasing insulation resistance and decreasing the weight of channels with irrelevant textures.

[0050] The spatial attention branch is used to highlight key regions in the initial fused features based on their spatial location. The initial fused features are subjected to average pooling and max pooling along the channel dimension, resulting in two two-dimensional matrices. These matrices are concatenated along the channels and then passed through a convolutional layer to reduce the dimensionality to a single-channel spatial weight map. Finally, an activation function is applied to generate the spatial attention weight map. This weighted map can automatically mark key spatial points such as the high-temperature area on the unexposed side, the edge of carbonization propagation, and the location of arc breakdown.

[0051] The initial fusion features are multiplied element-wise with the channel attention weight map and the spatial attention weight map to obtain the multimodal fusion features. The calculation method is as follows: in, Indicates the initial fusion features, This represents the channel attention weight map. Represents the spatial attention weight map. This indicates an element-wise multiplication operation. The formula shows that the activation value of each channel and each spatial location in the initial fused feature is jointly modulated by the channel weight map and the spatial weight map. This dual attention mechanism enables the model to adaptively adjust the feature response when facing different test samples, significantly enhancing its ability to extract key characteristics of cable fire blanket performance degradation, suppressing background interference and irrelevant changes, and ultimately generating high-quality multimodal fused features.

[0052] Step S103: Construct a multimodal fusion deep learning model, input the multimodal fusion features into the model, and use an adaptive learning rate optimization algorithm to iteratively train and optimize the model to obtain the optimal test model.

[0053] Specifically, the performance degradation of cable fire blankets during combustion is a dynamic temporal process, and its multimodal fusion feature sequence simultaneously contains both short-term dependent features and long-term trend features. For example, the instantaneous fluctuations of the flame and localized cracking of the material are short-term dependent features, while the continuous decrease in insulation resistance and the monotonic increase in thermal conductivity are long-term trends. To simultaneously capture feature patterns at these two time scales, a temporal dual-stream hybrid network is constructed as a multimodal fusion deep learning model.

[0054] This temporal dual-stream hybrid network comprises two feature extraction streams configured in parallel. The first stream is a recurrent neural network stream composed of gated recurrent units (RNNs). A gated RRN is an improved variant of a recurrent neural network that controls the flow of information through two gating mechanisms: an update gate and a reset gate. This effectively addresses the vanishing gradient problem that standard RRNs often encounter when processing long sequences. The update gate determines how much of the hidden state from the previous time step needs to be passed to the current time step, while the reset gate determines how the hidden state from the previous time step is combined with the current input. Gated RRNs are suitable for capturing short-term dependencies and local dynamic changes between adjacent time steps in a sequence, and can sensitively detect feature mutations caused by localized combustion anomalies.

[0055] The second stream is a temporal convolutional network stream composed of temporal convolutional networks. These temporal convolutional networks employ structural designs such as causal convolution, dilated convolution, and residual connections. Causal convolution ensures that at time... The output depends only on Inputs at or before a given time do not leak future information; dilated convolutions exponentially expand the receptive field by applying convolutional kernels to the input sequence in a skip-like manner, allowing shallower network layers to cover long time spans; residual connections alleviate the gradient decay problem in deep network training through direct cross-layer connections. Temporal convolutional networks are particularly adept at capturing long-term temporal patterns, effectively modeling the monotonic degradation trend of insulation resistance and the slow rise of temperature gradients during combustion processes lasting tens of minutes.

[0056] The gated recurrent unit stream and the temporal convolutional network stream are set up in parallel, sharing the input multimodal fusion feature sequence. The multimodal fusion features are fed into these two streams simultaneously, and the two streams independently extract feature representations at different time scales from their respective network structures.

[0057] The final hidden state sequence output by the gated recurrent unit (ROU) stream is concatenated with the temporal feature map output by the temporal convolutional network (TCNN) stream. The hidden state sequence output by the GROU stream is a two-dimensional matrix, with rows corresponding to each time step and columns corresponding to the feature dimensions of the stream's output. The hidden state vector at each time step encodes the sequence information and contextual dependencies up to that moment. The temporal feature map output by the TCNN stream is also a two-dimensional matrix, with the number of rows equal to the sequence length and the number of columns equal to the number of output channels of the stream. Its feature vector at each time step has a larger temporal receptive field, encoding pattern information within a longer time window near that moment.

[0058] These two matrices are concatenated along the feature column dimension, that is, the two sets of feature vectors at each time step are linked end to end to form a new, wider feature matrix. This concatenation result is the hybrid time-series feature map. The hybrid time-series feature map integrates short-term local dependency features and long-term global trend features, and is a comprehensive representation of the multimodal performance evolution of the cable fire blanket throughout the entire combustion process.

[0059] The hybrid temporal feature map is input into a fully connected layer. The fully connected layer learns a weight matrix and a bias vector, linearly combining and mapping the high-dimensional feature representations of the hybrid temporal feature map, outputting three independent low-dimensional feature vectors corresponding to flame retardant performance, thermal insulation performance, and electrical insulation performance, respectively. Let these three feature vectors output by the fully connected layer be... , , Each of them contains high-level semantic features required for the corresponding performance category discrimination.

[0060] Then , , Each input to the Softmax layer is converted into a probability distribution. The Softmax function is defined as follows: for the ... Class performance Each level has an output probability of: This function first uses an exponential function. Each unnormalized value is mapped to the positive number domain, and then normalized by dividing by the sum of all level exponents. The monotonically increasing property of the exponential function ensures that larger terms in the original values ​​still have larger probabilities after normalization, while the sum of all output probabilities is strictly 1, and each probability value is between 0 and 1, satisfying the mathematical definition of a probability distribution.

[0061] After passing through the Softmax layer, the model outputs a three-dimensional joint probability distribution vector, which gives the predicted probabilities of the flame retardant level, heat insulation level and electrical insulation level of the fire blanket of the cable under test, respectively, completing the end-to-end mapping from multimodal fusion features to multidimensional performance evaluation results.

[0062] To train the model parameters, a loss function needs to be defined to quantify the difference between the joint probability distribution of the model output and the true label, and this loss is minimized to guide the model to update its parameters in the correct direction. This invention employs the multi-task cross-entropy loss function.

[0063] Assuming the real label uses one-hot encoding, for the first... Class performance, When the actual level is number Level All other elements are zero. The predicted probability distribution output by the model is as follows: The single-task cross-entropy loss is defined as: Since only one element in the one-hot encoded vector is 1, let the th element be... The index of the actual performance level in the class is Therefore, the above summation actually only leaves... The corresponding item, namely .

[0064] The multi-task cross-entropy loss function is the sum of the cross-entropy losses of three independent performance tasks. in, Indicates the total number of performance categories. Indicates performance category, Indicates the level number within the performance category. Indicates the first The total number of performance levels for each class. The one-hot encoded value of the real label. This represents the predicted probability value output by the temporal two-stream hybrid network. The value is 3, which corresponds to flame retardant performance, heat insulation performance and electrical insulation performance, respectively.

[0065] Suppose that for a single sample, the model outputs a joint probability distribution. The true label of the sample is Assuming the predictions of the three performance classes are independent given the input features, the joint likelihood function of the true label of the sample is the product of the probabilities of the three independent tasks: in This indicates that the model predicts the sample in the th... Real-world performance level The probability of.

[0066] The goal of maximum likelihood estimation is to find a set of model parameters that maximizes the joint likelihood product across all training samples. For ease of optimization, taking the log-likelihood transforms the product into a summation: ; Using the properties of one-hot encoding, the above equation can be equivalently written as: Maximizing the log-likelihood is equivalent to minimizing its negative value, resulting in: Regarding the cross-entropy loss function with respect to the unnormalized score Find the derivative using the derivative property of the Softmax function. ,in Let Kronecker function be used when The value is 1 if the condition is met, and 0 otherwise. Therefore, the cross-entropy loss... The partial derivatives are: because exist When the value is 1, the rest are 0, and Therefore, we finally obtain a concise gradient expression. The difference between the predicted probability and the true one-hot encoding. When the model predicts correctly, When the gradient is close to 1 at the true level, it is close to zero, and the parameter update magnitude is very small. When the model makes a prediction error, the predicted probability deviates significantly from the true label, and the gradient drives the parameters to update significantly in the direction of reducing the deviation. This concise gradient form ensures the efficiency and numerical stability of backpropagation training.

[0067] After obtaining the loss function value and the gradients of each parameter, the parameters are updated using an adaptive learning rate optimization algorithm based on Adam. The Adam algorithm combines the advantages of the momentum method and the RMSprop algorithm. By maintaining the first and second moment estimates of the gradient, it adaptively adjusts the learning rate for each parameter. It is particularly suitable for handling non-stationary objective functions and high-dimensional parameter spaces, and can effectively accelerate convergence and enhance the robustness of training.

[0068] Let the number of training iterations be... The network's learnable parameter vector is , No. The gradient of the loss function with respect to the parameters is: The Adam algorithm first calculates the first and second moment estimates of the gradient: in and These are the exponential moving averages of the first and second moments, respectively. The decay rate of the first moment is typically controlled by a value of 0.9. The decay rate of the second moment is controlled, typically set to 0.999. The first moment approximates the first-order origin moment of the gradient, reflecting the direction and strength of the gradient. It acts as a momentum term, accelerating descent in a consistent gradient direction and counteracting fluctuations in oscillating directions. The second moment approximates the second-order origin moment of the gradient, reflecting the square of the gradient magnitude, and is used to measure the mean square magnitude of the gradient.

[0069] because and Initialized as zero vectors, their estimates will be biased towards zero in the early stages of training, requiring bias correction. The corrected estimates are as follows: ; ; The parameter update formula is: ; in The initial learning rate controls the global update step size; It is the square root of the corrected second moment; It is a very small constant, usually taking the value of This is used to prevent division by zero errors caused by a denominator of zero.

[0070] The adaptive adjustment mechanism of the Adam algorithm is reflected in the fact that when the gradient of a certain parameter remains large, its second moment... The corresponding increase makes the effective step size The step size is automatically reduced to prevent skipping the optimal point due to excessively large step sizes. When the gradient is small, the second moment is small, and the effective step size is automatically increased to help the parameters quickly pass through flat regions. At the same time, the momentum mechanism of the first moment allows the parameters to accelerate when the gradient direction is consistent and to reduce oscillations when the gradient direction frequently reverses, thereby achieving fast and stable convergence.

[0071] During the training phase, the multimodal fusion features of a large number of standard combustion test samples labeled with flame retardant, heat insulation, and electrical insulation ratings are input into a temporal dual-stream hybrid network in batches. The model performs forward propagation on each batch of samples, calculating the joint probability distribution output and cross-entropy loss value. Subsequently, the gradient of the loss function with respect to all learnable parameters of the network is calculated using the backpropagation algorithm. Finally, the Adam optimization algorithm described above is used to update all weight parameters and bias parameters based on the gradient.

[0072] As training iterations proceed, the loss value gradually decreases, and the model's predictive performance on the training set continuously improves. To prevent overfitting, the loss or evaluation metric is calculated on an independent validation set after each training cycle. When the validation loss no longer decreases within a preset number of consecutive cycles, an early termination strategy is triggered, and the model parameters at which the validation loss is lowest are saved as the optimal test model. This optimal test model can accurately and stably output the joint probability distribution of the flame retardant rating, heat insulation rating, and electrical insulation rating of the cable fire blanket for new input real-time multimodal fusion features, thereby completing the automated multidimensional performance evaluation of unknown samples.

[0073] Step S104: After the image data and electrothermal data of the fireproof blanket of the cable to be tested are filtered out by noise and extracted by feature extraction, they are input into the optimal test model and the performance test results of the fireproof blanket of the cable to be tested are output.

[0074] Specifically, when a brand-new fire blanket for a test cable undergoes a standard combustion test, the resulting real-time data stream, after undergoing the same noise filtering, time alignment, and feature extraction processes as the training phase, can be mapped to the same feature space as the training data. The optimal test model then performs forward inference on this real-time multimodal fusion feature to output a joint probability distribution, enabling automated performance evaluation of unknown samples. The entire inference process requires no manual intervention, achieving an end-to-end mapping from raw sensor data to the final performance level.

[0075] The optimal test model, obtained by inputting real-time multimodal fusion features, is the converged temporal dual-stream hybrid network. These features are simultaneously fed into a gated recurrent unit stream and a temporal convolutional network stream, extracting short-term temporal dependency features and long-term temporal pattern features respectively. The features output from the two streams are concatenated to obtain a real-time hybrid temporal feature map. This map is then passed through fully connected layers and a softmax layer to output a three-dimensional joint test result vector. This vector contains three sub-vectors, corresponding to the probability distributions of flame retardancy rating, thermal insulation rating, and electrical insulation rating, respectively. For each performance category, the rating with the highest probability is taken as the final judgment result for that performance.

[0076] If any flame retardant rating, heat insulation rating, or electrical insulation rating in the joint test result vector is lower than a preset safety threshold, it indicates a safety risk in that performance aspect of the cable fire blanket, and the system automatically generates a warning signal. To trace the cause and location of performance defects, the spatiotemporal synchronization features corresponding to that moment are subjected to inverse spatiotemporal mapping. Inverse spatiotemporal mapping refers to propagating the high-response feature channels and spatial locations layer by layer back to the original input data space along the reverse path of slow feature analysis and attention weight allocation, locating the defective areas in the original infrared thermal image and high-speed camera image that cause the low performance level, such as areas of severe localized carbonization, areas of abnormally concentrated temperature, or electrical breakdown points. These areas are then visually marked on the original image, enabling precise tracing and location of performance defects.

[0077] 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. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0078] In the description of this specification, the 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 this application. 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. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0079] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

Claims

1. A method for testing fire-retardant, heat-insulating, and electrically insulating properties of a cable fire blanket based on machine learning, characterized by, Includes the following steps: Image data and electrothermal data of the cable fire blanket during combustion were acquired, noise was filtered out, and the data were integrated into a multimodal test dataset in time stamp order. Feature extraction is performed on the multimodal test dataset to generate multimodal fusion features; wherein, the feature extraction includes multi-source data dimensionality reduction, spatiotemporal feature filtering, cross-dimensional feature extraction through a lightweight convolutional neural network with attention mechanism, and weight allocation through attention mechanism; A multimodal fusion deep learning model is constructed. The multimodal fusion features are input into the model, and an adaptive learning rate optimization algorithm is used to iteratively train and optimize the model to obtain the optimal test model. After noise filtering and feature extraction, the image data and electrothermal data of the fireproof blanket for the cable under test are input into the optimal test model, and the performance test results of the fireproof blanket for the cable under test are output.

2. The method of claim 1, wherein the machine learning based cable fire blanket fire-retardant, thermal insulation, and electrical insulation testing method is characterized by, The image data includes infrared thermal imaging data and high-speed camera data during the burning of the cable fire blanket; the electrothermal data includes insulation resistance, withstand voltage, thermal conductivity, surface temperature, and internal temperature during the burning of the cable fire blanket.

3. The method of claim 2, wherein the machine learning based cable fire blanket fire-retardant, thermal insulation, and electrical insulation testing method is characterized by, The step of acquiring image data and electrothermal data during the combustion process of the cable fire blanket, filtering out noise, and integrating them into a multimodal test dataset in time-stamp order includes: Infrared thermal imaging data is acquired using an infrared thermal imager, and high-speed video data is acquired using a high-speed camera. Insulation resistance and withstand voltage values ​​were collected using an insulation resistance tester and a withstand voltage tester, respectively. Thermal conductivity was collected using a heat flow meter. Surface temperature and internal temperature were collected using multiple thermocouples installed on the surface and inside the cable fire blanket, respectively.

4. The method of claim 2, wherein the machine learning based cable fire blanket fire-retardant, thermal insulation, and electrical insulation testing method is characterized by, The step of acquiring image data and electrothermal data during the combustion process of the cable fire blanket, filtering out noise, and integrating them into a multimodal test dataset in time-stamp order further includes: A variational autoencoder model is constructed for the infrared thermal imaging data, high-speed camera data, and electrothermal data consisting of insulation resistance, withstand voltage, thermal conductivity, surface temperature, and internal temperature, respectively. Each variational autoencoder model encodes and decodes its corresponding data to reconstruct it. Isolated data points with reconstruction errors greater than a preset threshold are filtered out as noise components to obtain clean image data and clean electrothermal data. Using the ignition time of the combustion test as the base timestamp, the frame rate and sampling rate of the pure image data and pure electrothermal data are time-aligned using an interpolation algorithm and then stitched together along the time dimension to form the multimodal test dataset.

5. The test method for flame retardant, heat-insulating, and electrical insulation of cable fire blankets based on machine learning according to claim 4, characterized in that, The step of constructing a variational autoencoder model for the infrared thermal imaging data, high-speed camera data, and electrothermal data consisting of insulation resistance, withstand voltage, thermal conductivity, surface temperature, and internal temperature includes: A first variational autoencoder model is constructed for the infrared thermal imaging data. The encoder of the first variational autoencoder model is composed of a multi-layer convolutional neural network. It extracts the spatial temperature distribution features in the infrared thermal image step by step and compresses the features into a low-dimensional latent vector. The decoder of the first variational autoencoder model is composed of a multi-layer deconvolutional neural network. It is used to reconstruct an infrared thermal image with the same size as the input from its low-dimensional latent vector. A second variational autoencoder model is constructed for the high-speed camera data. The encoder of the second variational autoencoder model is composed of a multi-layer convolutional neural network. It extracts texture, edge and deformation features in the visible light image and compresses the features into a low-dimensional latent vector. The decoder of the second variational autoencoder model is composed of a multi-layer deconvolutional neural network. It is used to reconstruct a visible light image with the same size as the input from its low-dimensional latent vector. A third variational autoencoder model is constructed for the electrothermal data. The encoder and decoder of the third variational autoencoder model are both composed of multi-layer fully connected networks. By capturing the crossover and temporal evolution between insulation resistance, withstand voltage, thermal conductivity, surface temperature and internal temperature in the electrothermal data and compressing them into a low-dimensional latent vector, the electrothermal data consistent with the input dimension is reconstructed from its low-dimensional latent vector.

6. The test method for flame retardant, heat-insulating, and electrical insulation of cable fire blankets based on machine learning according to claim 5, characterized in that, Feature extraction is performed on the multimodal test dataset to generate multimodal fusion features; wherein, the feature extraction includes the steps of multi-source data dimensionality reduction, spatiotemporal feature filtering, cross-dimensional feature extraction using a lightweight convolutional neural network with an attention mechanism, and weight allocation using an attention mechanism, including: Principal component analysis was used to reduce the dimensionality of the image modal data in the multimodal test dataset, and an autoencoder was used to reduce the dimensionality of the electrothermal modal data to obtain a low-dimensional feature set. A slow feature analysis algorithm based on a sliding time window is used to process the low-dimensional feature set and extract spatiotemporal synchronous features that characterize the slow changes in material properties over time. A lightweight convolutional neural network based on an attention mechanism is constructed. The lightweight convolutional neural network is composed of multiple depthwise separable convolutional layers and point convolutional layers stacked alternately. The spatiotemporal synchronization features are input into the lightweight convolutional neural network to obtain initial fusion features. A channel weight map is generated through a channel attention mechanism, and a spatial weight map is generated through a spatial attention mechanism. The initial fusion feature is then multiplied element-wise with the channel weight map and the spatial weight map to obtain the multimodal fusion feature.

7. The test method for flame retardant, heat-insulating, and electrical insulation of cable fire blankets based on machine learning according to claim 6, characterized in that, The loss function of the slow feature analysis algorithm for: in, For time step index, The total number of time steps. and The first Time and the The low-dimensional feature vector at time step 1. It is a feature mapping function; The initial fusion features are multiplied element-wise with the channel attention weight map and the spatial attention weight map to obtain the multimodal fusion features. The calculation method is as follows: in, Indicates the initial fusion features. This represents the channel attention weight map. Represents the spatial attention weight map. This indicates an element-wise multiplication operation.

8. The test method for flame retardant, heat-insulating, and electrical insulation of cable fire blankets based on machine learning according to claim 1, characterized in that, The step of constructing a multimodal fusion deep learning model, inputting the multimodal fusion features into the model, and using an adaptive learning rate optimization algorithm to iteratively train and optimize the model's parameters to obtain the optimal test model includes: A temporal dual-stream hybrid network is constructed as the multimodal fusion deep learning model. The temporal dual-stream hybrid network includes a recurrent neural network stream composed of gated recurrent units and a temporal convolutional network stream composed of temporal convolutional networks. The recurrent neural network stream and the temporal convolutional network stream are set in parallel. The multimodal fusion features are simultaneously input into the recurrent neural network stream and the temporal convolutional network stream. The recurrent neural network stream is used to extract temporal dependent features, and the temporal convolutional network stream is used to capture long-term temporal patterns. The hidden state sequence output by the recurrent neural network stream is concatenated with the temporal feature map output by the temporal convolutional network stream to obtain a hybrid temporal feature map. The hybrid temporal feature map is input into the fully connected layer and the Softmax layer, and the joint probability distribution of flame retardant rating, thermal insulation rating and electrical insulation rating is output. An Adam-based adaptive learning rate optimization algorithm is used to minimize the cross-entropy loss between the joint probability distribution and the true label as the training objective. The weight parameters of the temporal two-stream hybrid network are iteratively updated until the model converges, thus obtaining the optimal test model.

9. The test method for flame retardant, heat-insulating, and electrical insulation of cable fire blankets based on machine learning according to claim 8, characterized in that, The cross-entropy loss function for: in, Indicates the total number of performance categories. Indicates performance category, Indicates the level number within the performance category. Indicates the first The total number of performance levels for each class. The one-hot encoded value of the real label. This represents the predicted probability value output by the temporal two-stream hybrid network.

10. The test method for flame retardant, heat-insulating, and electrical insulation of cable fire blankets based on machine learning according to claim 1, characterized in that, The step of inputting the image data and electrothermal data of the fire blanket of the cable under test into the optimal test model after noise filtering and feature extraction, and outputting the performance test results of the fire blanket of the cable under test, includes: The multimodal fusion features of the fireproof blanket of the cable under test are input into the optimal test model, and a joint test result vector containing flame retardant rating, heat insulation rating and electrical insulation rating is output. When any level in the joint test result vector is lower than the preset safety threshold, an early warning signal is generated, and the spatiotemporal synchronization feature corresponding to that moment is reversed and located to the defect area in the original image data for visual marking.

Images