Insulation paper target attribute prediction method, system and device

Abstract

The invention discloses an insulation paper target attribute prediction method, system and device, and relates to the technical field of insulation materials, and the method comprises the steps: collecting the near infrared spectrum data of various types of insulation paper samples; constructing a domain adversarial neural network model, wherein the domain adversarial neural network model comprises an inter-domain shared feature extractor, a domain discriminator and a regression predictor; training the domain adversarial neural network model by adopting a staged strategy: in a feature learning stage, closing a gradient inversion layer to train the classification capability of a domain discriminator on different types of insulation paper samples; in an adversarial adaptation stage, starting a gradient inversion layer and dynamically adjusting inversion intensity, and forcing an inter-domain shared feature extractor to generate features irrelevant to the type of the insulation paper; inputting near infrared spectrum data of to-be-tested insulation paper into the trained prediction model, and predicting a target attribute value of the insulation paper; the method can improve the prediction performance of unknown samples.

Description

Technical Field

[0001] This invention relates to the field of insulating materials technology, and specifically to a method and system for predicting target properties of insulating paper. Background Technology

[0002] Insulating paper is the most commonly used key insulating material in power equipment, and accurate and non-destructive evaluation of its degree of polymerization and moisture content is of great significance.

[0003] Currently, near-infrared spectroscopy has become an important method for quantitatively assessing the moisture content, degree of polymerization, and other properties of insulating paper. Near-infrared spectroscopy establishes a correlation model between the insulating paper spectrum and the parameter to be measured, enabling non-destructive evaluation of the parameter using the collected sample spectra. However, building this model requires accumulating a database of various paper types to ensure the generalization ability of the technology. However, the types of paper used in transformers are very diverse, and the established database cannot completely cover all types. During field testing, it is unavoidable to encounter paper types covered in the database. In such cases, the predictive ability of the model based on traditional modeling will be significantly reduced. In the field of machine learning, this type of problem is collectively referred to as the domain generalization problem, which aims to improve the generalization ability of the source domain model to cope with the unknown interference of the target domain testing environment. The most significant characteristic of the domain generalization problem is the inability to obtain target domain samples in advance. Therefore, it is impossible to specifically adjust the model for the characteristics of the target domain samples; it can only rely on the model's original generalization ability.

[0004] To address the issue of near-infrared spectroscopy samples exceeding the database range in practical applications, related technological inventions mainly fall into two categories: First, continuously increasing the database samples to improve the model's generalization ability. This method is difficult to implement, especially for samples like insulating paper, which require extensive experimental accumulation, making it largely unsuitable. Second, attempting to obtain field test samples in advance and establish a simplified database for field testing scenarios. This method is commonly used in agriculture and the food industry, but insulating paper is a crucial insulation structure in electrical equipment, making it difficult to collect spectra beforehand, let alone sample it for relevant physicochemical parameter testing.

[0005] In summary, existing technologies mainly rely on expanding the database or pre-establishing a small-scale database containing samples from the scenarios to be tested. This requires extensive experimental operations, is cumbersome, and is difficult to apply effectively in evaluating the properties of insulating paper. Compared to scenarios where samples are readily available, such as agriculture and food, unknown types of insulating paper are often difficult to obtain in field applications, making it challenging to perform non-destructive testing of the properties to be tested. Summary of the Invention

[0006] To address the shortcomings of existing technologies in obtaining unknown types of insulating paper and the difficulty in non-destructively testing the properties to be tested, this invention proposes a method and system device for predicting target properties of insulating paper. By designing a domain adversarial neural network model, the model extracts domain-invariant features through a game-like process between a domain discriminator and a domain regressor, thereby solving the problems existing in the prior art.

[0007] A method for predicting target properties of insulating paper includes the following steps: Near-infrared spectral data of various types of insulating paper samples were collected; the near-infrared spectral data included the spectral data of the insulating paper samples and their corresponding target attribute labels; The spectral data of each type of insulating paper sample is used as input, and the corresponding target attribute label is used as output to train the domain adversarial neural network model. The spectral data of the insulating paper to be tested is input into the trained domain adversarial neural network model, and the multi-scale spectral features of the spectral data are extracted by the inter-domain shared feature extractor; the multi-scale spectral features are input into the regression predictor to generate the corresponding target attributes of the insulating paper. The process involves adversarial training between an inter-domain shared feature extractor and a regression predictor, connected by a gradient inversion layer. This forces the inter-domain shared feature extractor to generate features independent of the insulation paper sample type. The adversarial training includes: in the feature learning phase, disabling the gradient inversion layer to train the inter-domain discriminator's ability to classify different types of insulation paper samples; the regression predictor using backpropagation to guide the inter-domain shared feature extractor to extract features related to the target attribute; in the adversarial adaptation phase, enabling the gradient inversion layer and dynamically adjusting the inversion strength to force the inter-domain shared feature extractor to generate features independent of the insulation paper type; and training is completed when the inter-domain discriminator loses its ability to distinguish insulation paper sample types.

[0008] Furthermore, the shared feature extractor employs a 1D convolutional neural network with an Inception structure and a fully connected layer; the 1D convolutional neural network contains parallel convolutional layers with kernel widths of 3, 5, and 7, respectively, for extracting multi-scale spectral features.

[0009] Furthermore, during the adversarial adaptation phase, the gradient reversal weight coefficient is gradually increased using a linear or exponential function, from an initial value of 0.1 to 1.0.

[0010] Furthermore, the regression predictor optimizes the prediction results using the mean squared error (MSE) loss function.

[0011] Furthermore, the total loss function of the domain adversarial neural network model is: ; in, This is the mean squared error loss for regression prediction. The cross-entropy loss of the neighborhood discriminator. The weighting coefficients are dynamically adjusted. ; in, For loss function, For the domain discriminator to sample x i The prediction results For the sample x i Real domain tags, For mapping functions of inter-domain shared feature extractors, This indicates that the domain discriminator performs domain classification on the input features; in, N For the sample size, X s For the source domain dataset, For the model to sample x i The predicted value, For the sample x i The actual label value.

[0012] The present invention also includes a target property prediction system for insulating paper, comprising: The acquisition module is used to acquire near-infrared spectral data of various types of insulating paper samples; the near-infrared spectral data includes the spectral data of the insulating paper samples and their corresponding target attribute labels; The model building module is used to train the domain adversarial neural network model by taking the spectral data of each type of insulating paper sample as input and the corresponding target attribute label as output. Specifically, an inter-domain shared feature extractor, connected to a domain discriminator and a regression predictor via a gradient inversion layer, is used for adversarial training to force the inter-domain shared feature extractor to generate features unrelated to the type of insulating paper sample. The adversarial training includes: in the feature learning phase, disabling the gradient inversion layer to train the domain discriminator's classification ability for different types of insulating paper samples; the regression predictor using backpropagation to guide the inter-domain shared feature extractor to extract features related to the target attribute; in the adversarial adaptation phase, enabling the gradient inversion layer and dynamically adjusting the inversion strength to force the inter-domain shared feature extractor to generate features unrelated to the type of insulating paper; training is completed when the domain discriminator loses its ability to distinguish the type of insulating paper sample. The prediction module is used to input the spectral data of the insulating paper to be tested into the trained domain adversarial neural network model, extract multi-scale spectral features of the spectral data through the inter-domain shared feature extractor, and input the multi-scale spectral features into the regression predictor to generate the corresponding target attributes of the insulating paper.

[0013] The present invention also includes a computer device for predicting target properties of insulating paper, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method for predicting target properties of insulating paper.

[0014] The present invention also includes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the insulating paper target property prediction method.

[0015] This invention provides a method for predicting the target properties of insulating paper, which has the following beneficial effects: This invention introduces a domain discriminator and gradient inversion mechanism to construct an end-to-end domain adversarial model structure, automatically extracting features that are independent of paper type but strongly correlated with target attributes. The innovative training strategy is divided into a feature learning stage and an adversarial adaptation stage. In the early stage, classification ability is strengthened, and in the later stage, a gradient inversion layer (GRL) enables the extractor to learn "domain-invariant features." Dynamic weight adjustment avoids model oscillation and improves generalization performance. This invention applies domain adversarial neural networks to near-infrared spectroscopy analysis to improve the generalization ability for unknown insulating paper. The proposed staged training mechanism and dynamic loss weight adjustment strategy use the domain discriminator accuracy trend control mechanism (first increasing then decreasing) as the training objective constraint. It can improve the prediction performance for unknown samples without relying on target domain samples. This method is original in the field of near-infrared spectroscopy. The proposed deep learning architecture and training adjustment method are highly compatible with the data characteristics of near-infrared spectroscopy, possessing methodological and theoretical advancements. It can extract features that are independent of sample type but related to the target attribute without expanding the database or pre-establishing a small database containing samples of the test scene. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the domain adversarial neural network structure in an embodiment of the present invention; Figure 2 This is a schematic diagram of the shared feature extractor in an embodiment of the present invention; Figure 3 This is a schematic diagram of the domain discriminator in an embodiment of the present invention; Figure 4 This is a schematic diagram of the domain regressor in an embodiment of the present invention; Figure 5 This is a schematic diagram of the prediction results of the domain regressor on an unknown paper pattern in an embodiment of the present invention; Figure 6 This is a schematic diagram of the training error of the domain discriminator in an embodiment of the present invention; Figure 7This is a flowchart of a model adaptation method applicable to spectral analysis of unknown insulating paper in an embodiment of the present invention. Detailed Implementation

[0017] The technical inventions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] This invention proposes a model adaptation method suitable for spectral analysis of unknown insulating paper. It achieves the extraction of domain-invariant features through a game-like process between a domain discriminator and a domain regressor. The model includes a spectral feature extraction layer, as well as domain discriminator and domain regressor layers. During model training, two losses are recorded: the cross-entropy loss for paper sample classification error and the mean squared error loss for predicting the paper sample's properties. An effective training method is also designed: first, a positive gradient is applied to the domain discriminator during training. When the discriminator reaches a certain accuracy, gradient reversal is used for reverse training, reducing the discriminator's accuracy and enabling the model to extract domain-invariant features.

[0019] like Figure 7 As shown, the method specifically includes the following steps: S1. Establish a spectral database for multiple types of samples.

[0020] This invention aims to enable a feature extractor to capture domain-invariant features that are independent of sample type but related to the properties to be measured through domain adversarial learning. Therefore, it is necessary to establish a sample library of insulating paper with multiple types. This invention is described with a specific embodiment. In this embodiment, a sample library of aging insulating paper of six different types was accumulated, with insulating paper models BZZ075, GDL, BZZ125, 22HCC, JW50, and B31A. A total of 50 samples of each type were collected. In practical applications, it is preferable to have more than four sample types and more than 30 samples of each type. Near-infrared spectrometers were used to collect the sample spectra. The spectrometer used in this study was a Marine Optics NIRQuest+2.2, with a wavelength range of 900-2200 nm, 512 pixels, an integration time of 80 ms, and an average of 12 iterations. Finally, the spectral database of the six collected paper samples was recorded as follows: X all ={ X 1, …, X 6}, of which X i ={ x 1,…, x 50 Each paper sample should also be tested for its corresponding properties. In this embodiment, the degree of polymerization (DP) of insulating paper is used as an example. The degree of polymerization data for each paper sample is as follows: y =[ y 1,…,y n ].

[0021] S2. Construct a domain adversarial neural network model.

[0022] The domain adversarial neural network model designed in this invention is as follows: Figure 1 As shown, it includes an inter-domain shared feature extractor, a domain discriminator, and a regression predictor. The database data Xall is first input into the shared feature extractor, whose structure is as follows: Figure 2 As shown, the network architecture of the shared feature extractor consists of convolutional layers, a three-branch Inception convolutional layer, residual connections, feature tiling layers, and a fully connected layer for feature dimensionality reduction. Specifically, it includes: first, a standard convolutional layer (Conv1D), then using small convolutional kernels (such as...) k =3) Perform local feature extraction and low-level pattern capture, and enhance nonlinear representation capabilities through ReLU activation: The input features are simultaneously fed into three parallel convolutional branches:

[0023] Branch 1: Convolution kernel size is 333, extracting local fine-grained spectral features.

[0024] Branch 2: Convolution kernel size is 555, extracting spectral correlations in the medium range.

[0025] Branch 3: The kernel size is 777, capturing long-range correlations and smoother spectral patterns.

[0026] The outputs of the three branches are concatenated along the channel dimension to obtain a feature representation that incorporates multi-scale information.

[0027] To avoid the gradient vanishing or degradation problems caused by multi-scale convolution, a residual shortcut is added between the Inception output and input. This structure preserves the original features while adding multi-scale enhanced features, improving training stability and convergence speed.

[0028] The multi-channel feature maps extracted by the convolution and Inception modules are flattened into one-dimensional vectors for easier feeding into the fully connected layer. This process is equivalent to converting local feature maps into global feature representations.

[0029] This feature extractor can effectively extract multi-scale features (such as kernel widths of 3, 5, and 7) from spectral data. Combined with domain adversarial learning, it can extract domain-independent features even more effectively. This extractor does not perform feature compression; the output is still 512-dimensional spectral data, but it performs sufficient feature extraction. Subsequently, the data is fed into a domain discriminator and a regression predictor, both of which are simple fully connected networks. In the domain discriminator, the model classifies six types of samples, with the classification loss being the cross-entropy loss. In the early stages of training, the model training objective is to minimize the classification loss, enabling the domain discriminator to possess a certain sample recognition capability. In the later stages of training, gradient inversion is used to maximize the classification loss, gradually causing the domain classifier to lose the ability to identify sample types. In the regression predictor, the mean squared error of the model output is calculated. Ultimately, the total error of the model is the sum of the two, namely .in, This is the mean squared error loss for regression prediction. The cross-entropy loss of the neighborhood discriminator. These are dynamically adjusted weighting coefficients.

[0030] ; in, For loss function, For the domain discriminator to sample x i The prediction results For the sample x i Real domain tags, For mapping functions of inter-domain shared feature extractors, This indicates that the domain discriminator performs domain classification on the input features.

[0031] in, N For the sample size, X s For the source domain dataset, For the model to sample x i The predicted value, For the sample x i The actual label value.

[0032] S3. Model Training Process: The model training process is divided into two stages: the feature learning stage and the adversarial adaptation stage. The stability of training and the domain adaptation performance are balanced by adjusting the loss weights. During the first 25 epochs of training, the model is in the feature learning stage, with the gradient reversal layer (GRL) disabled. The goal of this stage is to improve the domain discriminator's ability to recognize different paper patterns, thereby enhancing the feature extractor's ability to represent intra-domain differences. From the 26th epoch onwards, the gradient reversal mechanism is enabled. By reversing the gradient direction during backpropagation, the feature extractor is forced to learn data features that are difficult for the domain discriminator to distinguish, thus achieving domain-invariant feature extraction. The loss function consists of the paper pattern regression error (mean squared error) and the domain discrimination error (cross-entropy loss), with the weighting coefficients of the two being adjustable hyperparameters. λ In this embodiment, the feature learning stage is set as follows: λ =0.1 to enhance regression ability, the adversarial phase will λ The algorithm was gradually upgraded to version 1.0 to balance prediction accuracy and domain adaptability. A dynamic loss weight adjustment strategy was employed throughout the training process to effectively suppress model oscillations and improve generalization performance.

[0033] S4. Input the near-infrared spectral data of the insulating paper to be tested into the trained prediction model to predict the target attribute value of the insulating paper.

[0034] Model Training Results: During model training, two metrics need to be observed simultaneously. The first is used to measure the model's prediction accuracy for insulating paper properties (such as degree of polymerization): RMSE (Root Mean Square Error): the smaller the better, indicating a smaller gap between the predicted and actual values; MAE (Mean Absolute Error): also the smaller the better; MAPE (Mean Absolute Percentage Error): has strong explanatory power when the amount of data varies greatly; R² (Coefficient of Determination): the closer to 1, the better the model fit. If RMSE < 100, MAPE < 10%, and R² > 0.9 on the target domain test set, the performance can be considered good. In this embodiment, tests were conducted on a sample set outside the database, with the insulating paper type being DLZ, which is cable paper. The test results are as follows: Figure 5 As shown, RMSE=64.3 and MAPE=7.44, indicating good performance. Next is the metric used to measure the sample type recognition accuracy of the domain discriminator. This metric should gradually increase to 1 before gradient inversion is applied, and then gradually decrease to (1 / sample class) after gradient inversion is applied. The training results of the domain discriminator in this embodiment are shown below. Figure 6 As shown, the model basically meets the training objective of accurately rising first and then falling, indicating that the model training effect is good.

[0035] To address the issue of significantly reduced modeling accuracy caused by field sample types exceeding the training database's range, traditional methods cannot obtain the spectra and labels of unknown paper samples beforehand, making it difficult to construct a new database. This invention innovatively proposes a solution: by introducing a domain discriminator and gradient inversion mechanism, an end-to-end domain adversarial model structure is constructed; deep learning is used to automatically extract features that are unrelated to paper sample type but strongly correlated with target attributes (such as aggregation degree); the model structure consists of three parts: a shared feature extractor, a domain discriminator, and a regression predictor. The training strategy is innovative: divided into a feature learning stage (first 25 rounds) and an adversarial adaptation stage (round 26 and beyond); classification ability is strengthened in the early stages, and in the later stages, a gradient inversion layer (GRL) enables the extractor to learn "domain-invariant features"; the weight ratio between cross-entropy loss and regression MSE is set. λInitially, λ=0.1, gradually increasing to 1.0 later; weights are dynamically adjusted to avoid model oscillation and improve generalization performance. Innovative network structure: 1D convolution with an Inception structure is used to extract multi-scale spectral features (kernel widths of 3, 5, and 7); achieving a balance between high-frequency and low-frequency information, adapting to the differences in spectral characteristics of different types of paper samples.

[0036] This invention applies domain adversarial neural networks to near-infrared spectral analysis to improve the generalization ability of unknown insulating paper; the proposed phased training mechanism and dynamic loss weight adjustment strategy include: enhancing the classification ability of the domain discriminator in the early stage of training, and suppressing its accuracy in the later stage through gradient inversion. λ The value is gradually adjusted from 0.1 to 1.0 to coordinate prediction accuracy and domain adaptability; the Inception structural feature extractor for 1D spectral data is designed with different convolutional kernel widths; the control mechanism for the accuracy trend of the domain discriminator (first rising and then falling) is used as a training objective constraint; a method to improve the prediction performance of unknown samples without relying on target domain samples; a dedicated network structure and training method for predicting target attributes such as polymerization degree and moisture content of insulating paper.

[0037] Based on the same inventive concept, this invention also proposes a target property prediction system for insulating paper, comprising: The acquisition module is used to acquire near-infrared spectral data of various types of insulating paper samples; the near-infrared spectral data includes the spectral data of the insulating paper samples and their corresponding target attribute labels.

[0038] The model building module is used to train the domain adversarial neural network model by taking the spectral data of each type of insulating paper sample as input and the corresponding target attribute label as output. Specifically, an inter-domain shared feature extractor, connected to a domain discriminator and a regression predictor via a gradient inversion layer, is used for adversarial training to force the inter-domain shared feature extractor to generate features unrelated to the type of insulating paper sample. The adversarial training includes: in the feature learning phase, disabling the gradient inversion layer to train the domain discriminator's classification ability for different types of insulating paper samples; the regression predictor using backpropagation to guide the inter-domain shared feature extractor to extract features related to the target attribute; in the adversarial adaptation phase, enabling the gradient inversion layer and dynamically adjusting the inversion strength to force the inter-domain shared feature extractor to generate features unrelated to the type of insulating paper; training is completed when the domain discriminator loses its ability to distinguish the type of insulating paper sample.

[0039] The prediction module is used to input the spectral data of the insulating paper to be tested into the trained domain adversarial neural network model, extract multi-scale spectral features of the spectral data through the inter-domain shared feature extractor, and input the multi-scale spectral features into the regression predictor to generate the corresponding target attributes of the insulating paper.

[0040] The present invention also proposes a computer device for predicting the target properties of insulating paper, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method for predicting the target properties of insulating paper.

[0041] The present invention also proposes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the insulating paper target property prediction method.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical invention and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting target properties of insulating paper, characterized in that, Includes the following steps: Near-infrared spectral data of various types of insulating paper samples were collected; the near-infrared spectral data included the spectral data of the insulating paper samples and their corresponding target attribute labels; The spectral data of each type of insulating paper sample is used as input, and the corresponding target attribute label is used as output to train the domain adversarial neural network model. The spectral data of the insulating paper to be tested is input into the trained domain adversarial neural network model, and the multi-scale spectral features of the spectral data are extracted through the inter-domain shared feature extractor. Multi-scale spectral features are input into a regression predictor to generate corresponding target properties of insulating paper; The process involves adversarial training between an inter-domain shared feature extractor and a regression predictor, connected by a gradient inversion layer. This forces the inter-domain shared feature extractor to generate features independent of the insulation paper sample type. The adversarial training includes: in the feature learning phase, disabling the gradient inversion layer to train the inter-domain discriminator's ability to classify different types of insulation paper samples; the regression predictor using backpropagation to guide the inter-domain shared feature extractor to extract features related to the target attribute; in the adversarial adaptation phase, enabling the gradient inversion layer and dynamically adjusting the inversion strength to force the inter-domain shared feature extractor to generate features independent of the insulation paper type; and training is completed when the inter-domain discriminator loses its ability to distinguish insulation paper sample types.

2. The method for predicting target properties of insulating paper according to claim 1, characterized in that, The inter-domain shared feature extractor employs a 1D convolutional neural network with an Inception architecture and a fully connected layer; the 1D convolutional neural network contains parallel convolutional layers with kernel widths of 3, 5, and 7, which are used to extract multi-scale spectral features.

3. The method for predicting target properties of insulating paper according to claim 1, characterized in that, During the adversarial adaptation phase, the gradient reversal weight coefficient is gradually increased using a linear or exponential function, from an initial value of 0.1 to 1.

0.

4. The method for predicting target properties of insulating paper according to claim 1, characterized in that, The regression predictor optimizes the prediction results using the mean squared error (MSE) loss function.

5. The method for predicting target properties of insulating paper according to claim 4, characterized in that, The total loss function of the domain adversarial neural network model is: ； in, This is the mean squared error loss for regression prediction. The cross-entropy loss of the neighborhood discriminator. The weighting coefficients are dynamically adjusted. ； in, For loss function, For the domain discriminator to sample x i The prediction results For the sample x i Real domain tags, For mapping functions of inter-domain shared feature extractors, This indicates that the domain discriminator performs domain classification on the input features; in, N For the sample size, X s For the source domain dataset, For the model to sample x i The predicted value, For the sample x i The actual label value.

6. A target property prediction system for insulating paper, characterized in that, include: The acquisition module is used to acquire near-infrared spectral data of various types of insulating paper samples; the near-infrared spectral data includes the spectral data of the insulating paper samples and their corresponding target attribute labels; The model building module is used to train the domain adversarial neural network model by taking the spectral data of each type of insulating paper sample as input and the corresponding target attribute label as output. Specifically, an inter-domain shared feature extractor, connected to a domain discriminator and a regression predictor via a gradient inversion layer, is used for adversarial training to force the inter-domain shared feature extractor to generate features unrelated to the type of insulating paper sample. The adversarial training includes: in the feature learning phase, disabling the gradient inversion layer to train the domain discriminator's classification ability for different types of insulating paper samples; the regression predictor using backpropagation to guide the inter-domain shared feature extractor to extract features related to the target attribute; in the adversarial adaptation phase, enabling the gradient inversion layer and dynamically adjusting the inversion strength to force the inter-domain shared feature extractor to generate features unrelated to the type of insulating paper; training is completed when the domain discriminator loses its ability to distinguish the type of insulating paper sample. The prediction module is used to input the spectral data of the insulating paper to be tested into the trained domain adversarial neural network model, extract multi-scale spectral features of the spectral data through the inter-domain shared feature extractor, and input the multi-scale spectral features into the regression predictor to generate the corresponding target attributes of the insulating paper.

7. A computer device for predicting the target attributes of insulating paper, characterized in that, include: A memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the insulating paper target property prediction method according to any one of claims 1-5.

8. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which includes program instructions that, when executed by a processor, perform the steps of the insulating paper target property prediction method according to any one of claims 1-5.

Images