Silicone rubber lossless aging prediction method and system based on multi-modal data fusion
Through multimodal data fusion and deep belief network, combined with terahertz spectroscopy and other data sources, the single data dependence problem of silicone rubber aging assessment is solved, and accurate non-destructive assessment of silicone rubber aging is achieved, improving prediction accuracy and reliability sex.
Patent Information
- Application Number
- CN202411436395.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-15
- Publication Date
- 2024-11-15
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The evaluation of silicone rubber aging in the prior art mostly relies on a single data source, which is difficult to fully reflect the aging state, and the impact of different components and environmental conditions has not been effectively integrated, resulting in limited accuracy and reliability of the prediction model, and some The detection method requires intercepting samples, which leads to non-destructive testing difficulties.
Using multimodal data fusion method, aging prediction model is constructed to achieve the aging degree of silicon rubber aging through terahertz spectral analysis and deep belief network, combining time domain spectral data, vibration response signal and hydrophobic droplet area and contact angle data. and lossless assessment of time.
It improves the accuracy and reliability of silicone rubber aging prediction, avoids equipment failures and losses caused by aging, provides a clearer use and replacement time, and realizes non-destructive testing.
Smart Images

Figure CN118968314A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of silicone rubber detection, and in particular to a method and system for predicting silicone rubber non-destructive aging based on multi-modal data fusion. Background Art
[0002] Silicone rubber refers to a rubber whose main chain is composed of alternating silicon and oxygen atoms, and the silicon atom is usually connected to two organic groups. Ordinary silicone rubber is mainly composed of siloxane chains containing methyl and a small amount of vinyl. The introduction of phenyl can improve the high and low temperature resistance of silicone rubber, and the introduction of trifluoropropyl and cyano can improve the heat and oil resistance of silicone rubber. Silicone rubber has good low temperature resistance and can generally work at -55℃. After the introduction of phenyl, it can reach -73℃. The heat resistance of silicone rubber is also outstanding. It can work for a long time at 180℃, and can withstand several weeks or more at a temperature slightly higher than 200℃ and still be elastic. It can withstand high temperatures above 300℃ instantly. Silicone rubber has good air permeability and the oxygen permeability is the highest among synthetic polymers. In addition, silicone rubber also has the outstanding characteristics of physiological inertness and will not cause blood coagulation, so it is widely used in the medical field.
[0003] As an important industrial material, accurate prediction of the aging performance of silicone rubber is crucial to ensure product quality and extend service life. In the prior art, the evaluation of silicone rubber aging mostly relies on a single data source, such as spectral analysis, physical property testing, etc., which is difficult to fully reflect the aging state of silicone rubber. In addition, the effects of different components and environmental conditions on silicone rubber aging have not been effectively integrated, resulting in limited accuracy and reliability of the prediction model. In addition, some detection methods require cutting samples from silicone rubber products, causing damage to the silicone rubber. Summary of the invention
[0004] In view of the above-mentioned deficiencies in the prior art, the present invention aims to provide a method and system for non-destructive aging prediction of silicone rubber based on multimodal data fusion, so as to solve the problem of inaccurate evaluation caused by evaluating the existing silicone rubber aging assessment prediction through single data.
[0005] In order to solve the above problems, the present invention adopts the following technical solutions:
[0006] On the one hand, the present invention provides a non-destructive aging prediction method for silicone rubber based on multimodal data fusion, comprising:
[0007] The composition and content of silicone rubber are analyzed by terahertz spectroscopy or the proportion of components during the preparation of silicone rubber to obtain composition and content data;
[0008] Collect the time domain spectrum data, vibration response signal time domain data, hydrophobic droplet area and contact angle data of silicone rubber with different compositions, thicknesses and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees over aging time;
[0009] The collected data was used to train and verify the deep belief network, and the relationship between the aging degree and aging time of silicone rubber with different compositions, thicknesses and densities at different temperatures, compression or stretching degrees, ambient humidity and corona treatment degrees was explored to obtain an aging prediction model.
[0010] The composition data, thickness and density of the silicone rubber to be estimated, as well as the temperature, compression or tension degree, ambient humidity and / or corona degree data and spectral data of the silicone rubber to be estimated in the working state, vibration response signal data and hydrophobic droplet area and contact angle data are input into the aging prediction model, and the aging prediction model outputs a relationship curve between the aging degree of the silicone rubber to be estimated and the aging time.
[0011] As an implementation method, the composition and content of silicone rubber are analyzed by terahertz spectroscopy or the composition ratio during the preparation of silicone rubber to obtain composition and content data, including:
[0012] The silicone rubber sample is scanned using terahertz spectroscopy technology to obtain its spectral data;
[0013] The acquired spectral data were preprocessed to eliminate noise and background interference, and the Legendre moment method was used to extract features from the spectral data;
[0014] The extracted features are used to construct a three-dimensional image of the sample absorbance spectrum and converted into a grayscale image;
[0015] The least squares-partial least squares regression model was used to perform regression analysis on the features and obtain the composition and content data of silicone rubber.
[0016] As an implementation method, the deep belief network includes a visible layer and multiple hidden layers, each hidden layer has a hidden unit multi-layer restricted Boltzmann machine; the energy function of the restricted Boltzmann machine for:
[0017]
[0018] in, and are the neuron states of the visible layer and the hidden layer, respectively. and is the bias term, is the connection weight, and the goal of the network is to minimize the energy function;
[0019] By training the restricted Boltzmann machine layer by layer and using the features learned by the restricted Boltzmann machine as the input of the next restricted Boltzmann machine, it is used to learn deep features and the associations between them, thereby realizing feature mining and fusion of the collected data.
[0020] As an implementable method, the training and verification of the deep belief network using the collected data includes:
[0021] Normalize the collected time-domain spectral data, vibration response signal time-domain data, and hydrophobic droplet area data of silicone rubber with different compositions, thicknesses, and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees over aging time to form training sets and validation sets;
[0022] Pre-train the multi-layer restricted Boltzmann machine in the deep belief network and use the contrastive divergence algorithm to optimize the connection weights and bias terms;
[0023] Use the back-propagation algorithm to fine-tune the entire deep belief network, calculate the adjustment amount of the connection weights and bias items of each hidden layer, and achieve fine-tuning of the entire deep belief network;
[0024] When the training meets the prediction accuracy requirements, the data in the validation set is used to input the composition data, thickness and density of the silicone rubber as well as the temperature, compression or stretching degree, ambient humidity and / or corona degree data and spectral data, vibration response signal data and hydrophobic droplet area and contact angle data to determine whether the relationship between the aging degree and aging time output by the model is accurate.
[0025] As an implementable method, the area data of the hydrophobic droplets is collected using a Grab Cut algorithm, including:
[0026] Collect raw images of hydrophobic droplets;
[0027] Select the approximate hydrophobic droplet area in the original image;
[0028] Construct a st image of the approximate hydrophobic droplet area, and regard each pixel in the image as a node in the network graph;
[0029] Define the weights between nodes in the network graph based on the positive correlation between the similarities between nodes;
[0030] Determine relevant pixels of the foreground and the background, and determine independent nodes and edges connecting the independent nodes based on the relevant pixels of the foreground and the background;
[0031] Construct an energy function based on independent nodes, their edges and weights;
[0032] Through multiple iterations, the energy function is converged to the minimum, and the boundary nodes between the image subject and the background are obtained;
[0033] By calculating the area of the main body of the image, the area data of the hydrophobic droplet is obtained, and by calculating the contact angle of the main body of the image, the contact angle data of the hydrophobic droplet is obtained.
[0034] On the other hand, the present invention provides a silicone rubber nondestructive aging prediction system based on multimodal data fusion, including a silicone rubber component analysis module, a data acquisition and processing module, an aging prediction model construction module and a detection module;
[0035] The silicone rubber component analysis module is used to analyze the composition of silicone rubber components through terahertz spectroscopy or the proportion of components during the preparation of silicone rubber to obtain component and content data;
[0036] The data acquisition and processing module is used to collect time-domain spectrum data, vibration response signal time-domain data, and hydrophobic droplet area and contact angle data of silicone rubber with different compositions, thicknesses, and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees over aging time;
[0037] The aging prediction model building module is used to train and verify the deep belief network using the collected data, and to mine the relationship between the aging degree and aging time of silicone rubber with different compositions, thicknesses and densities at different temperatures, compression or stretching degrees, ambient humidity and corona treatment degrees, so as to obtain an aging prediction model;
[0038] The detection module is used to input the composition data, thickness and density of the silicone rubber to be estimated, as well as the temperature, compression or tension degree, ambient humidity and / or corona degree data and spectral data of the silicone rubber to be estimated in the working state, vibration response signal data and hydrophobic droplet area and contact angle data into the aging prediction model, and the aging prediction model outputs a relationship curve between the aging degree of the silicone rubber to be estimated and the aging time.
[0039] As an implementation method, the composition and content of silicone rubber are analyzed by terahertz spectroscopy or the composition ratio during the preparation of silicone rubber to obtain composition and content data, including:
[0040] The silicone rubber sample is scanned using terahertz spectroscopy technology to obtain its spectral data;
[0041] The acquired spectral data were preprocessed to eliminate noise and background interference, and the Legendre moment method was used to extract features from the spectral data;
[0042] The extracted features are used to construct a three-dimensional image of the sample absorbance spectrum and converted into a grayscale image;
[0043] The least squares-partial least squares regression model was used to perform regression analysis on the features and obtain the composition and content data of silicone rubber.
[0044] As an implementation method, the deep belief network includes a visible layer and multiple hidden layers, each hidden layer has a hidden unit multi-layer restricted Boltzmann machine; the energy function of the restricted Boltzmann machine for:
[0045]
[0046] in, and are the neuron states of the visible layer and the hidden layer, respectively. and is the bias term, is the connection weight, and the goal of the network is to minimize the energy function;
[0047] By training the restricted Boltzmann machine layer by layer and using the features learned by the restricted Boltzmann machine as the input of the next restricted Boltzmann machine, it is used to learn deep features and the associations between them, thereby realizing feature mining and fusion of the collected data.
[0048] As an implementable method, the training and verification of the deep belief network using the collected data includes:
[0049] Normalize the collected time-domain spectral data, vibration response signal time-domain data, and hydrophobic droplet area data of silicone rubber with different compositions, thicknesses, and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees over aging time to form training sets and validation sets;
[0050] Pre-train the multi-layer restricted Boltzmann machine in the deep belief network and use the contrastive divergence algorithm to optimize the connection weights and bias terms;
[0051] Use the back-propagation algorithm to fine-tune the entire deep belief network, calculate the adjustment amount of the connection weights and bias items of each hidden layer, and achieve fine-tuning of the entire deep belief network;
[0052] When the training meets the prediction accuracy requirements, the data in the validation set is used to input the composition data, thickness and density of the silicone rubber as well as the temperature, compression or stretching degree, ambient humidity and / or corona degree data and spectral data, vibration response signal data and hydrophobic droplet area and contact angle data to determine whether the relationship between the aging degree and aging time output by the model is accurate.
[0053] As an implementable method, the area data of the hydrophobic droplets is collected using a Grab Cut algorithm, including:
[0054] Collect raw images of hydrophobic droplets;
[0055] Select the approximate hydrophobic droplet area in the original image;
[0056] Construct a st image of the approximate hydrophobic droplet area, and regard each pixel in the image as a node in the network graph;
[0057] Define the weights between nodes in the network graph based on the positive correlation between the similarities between nodes;
[0058] Determine relevant pixels of the foreground and the background, and determine independent nodes and edges connecting the independent nodes based on the relevant pixels of the foreground and the background;
[0059] Construct an energy function based on independent nodes, their edges and weights;
[0060] Through multiple iterations, the energy function is converged to the minimum, and the boundary nodes between the image subject and the background are obtained;
[0061] By calculating the area of the main body of the image, the area data of the hydrophobic droplet is obtained, and by calculating the contact angle of the main body of the image, the contact angle data of the hydrophobic droplet is obtained.
[0062] The beneficial effects of the present invention are as follows: the present invention uses the composition, thickness, density, etc. of silicone rubber as product factors, and uses temperature, compression or stretching degree, environmental humidity, and corona degree as silicone rubber aging factors, and forms multimodal data through time domain spectral data, vibration response signal time domain data, and hydrophobic droplet area and contact angle data under the above product factors and aging factors, and uses a deep belief network to mine and extract the features in the multimodal data and the relationship between them, thereby forming an accurate silicone rubber aging prediction model. The silicone rubber aging prediction model of the present invention can be used to non-destructively evaluate or predict the aging of silicone rubber, and the result is more accurate than the current aging evaluation result through single data, making the use / replacement time of silicone rubber clearer, and avoiding problems and losses caused by equipment failure due to the aging of silicone rubber. BRIEF DESCRIPTION OF THE DRAWINGS
[0063] Figure 1 The present invention is a flow chart of a method for non-destructive aging prediction of silicone rubber based on multimodal data fusion.
[0064] Figure 2 This is a schematic diagram of a silicone rubber non-destructive aging prediction system based on multimodal data fusion according to the present invention. DETAILED DESCRIPTION
[0065] The present invention is further described in detail below in conjunction with specific embodiments.
[0066] It should be noted that these embodiments are only used to illustrate the present invention rather than to limit the present invention. Simple improvements to the method based on the concept of the present invention all fall within the scope of protection claimed by the present invention.
[0067] See also Figure 1 , which is a non-destructive aging prediction method for silicone rubber based on multimodal data fusion, including:
[0068] S100, analyzing the composition and content of silicone rubber by terahertz spectroscopy or the proportion of components during the preparation of silicone rubber to obtain composition and content data.
[0069] This embodiment provides a possible implementation method for collecting component and content data:
[0070] The terahertz spectroscopy technique is used to scan the silicone rubber sample and obtain its spectral data, aiming to capture the absorption characteristics of the sample in the terahertz frequency band, which is sensitive to the composition changes of silicone rubber.
[0071] The acquired spectral data is preprocessed to eliminate noise and background interference, and the Legendre moment method is used to extract features from the spectral data. The Legendre moment can effectively extract geometric features related to the composition from the spectrum, and these features are sensitive to changes in the composition.
[0072] The feature extraction formula is as follows:
[0073]
[0074] in, is the Legendre polynomial, n is the order of the polynomial, and x is the normalized spectral absorbance value.
[0075] The extracted features are used to construct a three-dimensional image of the sample absorbance spectrum and converted into a grayscale image.
[0076] Convert the extracted absorbance spectrum of each sample into an image. First, repeat the absorbance spectrum to construct a three-dimensional image of the spectrum, and then grayscale the three-dimensional image to obtain the corresponding sample grayscale image. According to the construction principle of three-dimensional spectrum, in theory, as long as there are two two-dimensional arrays, a three-dimensional image can be constructed. When constructing a three-dimensional absorbance graph, at least one repetition is required to obtain its three-dimensional graph.
[0077] The least squares-partial least squares regression model was used to perform regression analysis on the features and obtain the composition and content data of silicone rubber.
[0078] LM-PLSR combines the robustness of least squares and the multivariate analysis capability of partial least squares, and is suitable for dealing with the problem of simultaneous determination of multiple components. The basic form of the LM-PLSR model is as follows:
[0079]
[0080] Among them, Y is the observation matrix of component content, X is the feature matrix, B is the regression coefficient matrix, and E is the error matrix.
[0081] Taking a silicone rubber sample as an example, we first perform terahertz spectrum scanning to obtain spectral data. The above steps are used for data preprocessing, feature extraction, three-dimensional image construction, and grayscale conversion. Assuming that we have extracted the features of three main components, we use the LM-PLSR model to model the relationship between these features and the known component content.
[0082] The specific steps are as follows:
[0083] 1. Perform terahertz spectrum scanning on the sample to obtain raw spectrum data.
[0084] 2. The Legendre moment method is applied to extract features from the spectral data and obtain the eigenvectors of the three main components.
[0085] 3. Construct a 3D image and convert it into a grayscale image for further feature analysis.
[0086] 4. Use the LM-PLSR model to perform regression analysis on the feature vector and the known component content to establish a prediction model.
[0087] 5. After verifying the accuracy of the model, input the spectral data of the sample to be tested into the model to predict its component content.
[0088] Through the above steps, the accurate prediction of the component content of silicone rubber samples can be achieved.
[0089] S200, collect time domain spectral data, vibration response signal time domain data, and hydrophobic droplet area and contact angle data of silicone rubber with different compositions, thicknesses, and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees over aging time.
[0090] Time-domain spectroscopy data collection: Silicone rubber samples are measured using time-domain spectroscopy at different temperatures (e.g., room temperature, high temperature, and low temperature), compression or tension (e.g., -10%, 0%, 10%, 20% deformation), humidity, and corona levels. This step is designed to capture the absorption and scattering properties of the material in different physical states, thereby reflecting changes in its internal structure and aging state.
[0091] Data collection formula:
[0092]
[0093] in, is the spectral signal within the aging time t, is the light intensity of wavelength λ at time t, and is the spectral range considered.
[0094] Vibration response signal acquisition: Apply vibration excitation of different frequencies and amplitudes to the silicone rubber sample and record its vibration response signal. By analyzing the time domain characteristics of the signal (such as amplitude, frequency, and attenuation), the elastic modulus and internal damage of the material can be inferred.
[0095] Vibration signal feature extraction:
[0096]
[0097] in, is the frequency spectrum of the vibration signal, is the time domain vibration signal, represents Fourier transform.
[0098] The area data of hydrophobic droplets is collected using the Grab Cut algorithm, including:
[0099] 1. Collect the original image of the hydrophobic droplet;
[0100] 2. Select the approximate hydrophobic droplet area in the original image;
[0101] 3. Construct a rough ST image of the hydrophobic droplet area, and regard each pixel in the image as a node in the network graph;
[0102] 4. Define the weights between nodes in the network graph based on the positive correlation between the similarities between nodes;
[0103] 5. Determine the relevant pixels of the foreground and background, and determine the independent nodes and the edges connecting the independent nodes based on the relevant pixels of the foreground and background;
[0104] 6. Construct an energy function based on independent nodes, their edges and weights;
[0105] 7. Through multiple iterations, the energy function is converged to the minimum and the boundary nodes between the image subject and the background are obtained;
[0106] 8. By calculating the area of the main body of the image, the area data of the hydrophobic droplet is obtained, and by calculating the contact angle of the main body of the image, the contact angle data of the hydrophobic droplet is obtained.
[0107] Take a group of silicone rubber samples as an example, that is, the composition, thickness, and density data of the silicone rubber samples are determined. After collecting the time domain spectrum data, vibration response signal time domain data, and hydrophobic droplet area and contact angle data under different conditions, the time domain spectrum data, vibration response signal time domain data, and hydrophobic droplet area and contact angle data of silicone rubber of a certain composition, thickness, and density at a certain temperature, compression or stretching degree, ambient humidity, and corona treatment degree can be obtained. Repeating the above process frequently can obtain the time domain spectrum data, vibration response signal time domain data, and hydrophobic droplet area and contact angle data of silicone rubber of different compositions, thicknesses, and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees with aging time.
[0108] S300. Use the collected data to train and verify the deep belief network, explore the relationship between the aging degree and aging time of silicone rubber with different compositions, thicknesses and densities at different temperatures, compression or stretching degrees, ambient humidity and corona treatment degrees, and obtain an aging prediction model.
[0109] The deep belief network (DBN) adopted in the present invention makes full use of the information provided by different data sources to effectively fuse and process multimodal data, and can mine the intrinsic connections between different data sources to improve the accuracy and reliability of aging assessment.
[0110] The deep belief network used in the present invention includes a visible layer and multiple hidden layers, each hidden layer has a hidden unit multi-layer restricted Boltzmann machine (RBM); the energy function of the restricted Boltzmann machine for:
[0111]
[0112] in, and are the neuron states of the visible layer and the hidden layer, respectively. and is the bias term, is the connection weight, and the goal of the network is to minimize the energy function.
[0113] The DBN structure is shown below:
[0114]
[0115] By training the restricted Boltzmann machine layer by layer and using the features learned by the restricted Boltzmann machine as the input of the next restricted Boltzmann machine, it is used to learn deep features and the associations between them, thereby realizing feature mining and fusion of the collected data.
[0116] The collected data is used to train and validate the deep belief network including:
[0117] The collected time domain spectral data, vibration response signal time domain data and hydrophobic droplet area data of silicone rubber with different compositions, thicknesses and densities at different temperatures, compression or tension degrees, ambient humidity and corona treatment degrees with aging time were normalized to form training sets and validation sets.
[0118] The multi-layer restricted Boltzmann machine in the deep belief network is pre-trained, and the contrastive divergence algorithm is used to optimize the connection weights and bias terms. During the pre-training process, the parameters are adjusted through multiple iterations so that the model can capture the potential characteristics of the input data.
[0119] The back-propagation algorithm is used to fine-tune the entire deep belief network, and the adjustment amount of the connection weights and bias items of each hidden layer is calculated to achieve fine-tuning of the entire deep belief network.
[0120] During fine-tuning, the network parameters are adjusted using labeled data to improve prediction accuracy. The basic formula for backpropagation algorithm to adjust connection weights and biases is as follows:
[0121]
[0122]
[0123] in, and , is the adjustment amount for the connection weights and biases, is the learning rate, and the loss function is the above energy function.
[0124] When the training meets the prediction accuracy requirements, the data in the validation set is used to input the composition data, thickness and density of the silicone rubber as well as the temperature, compression or stretching degree, ambient humidity and / or corona degree data and spectral data, vibration response signal data and hydrophobic droplet area and contact angle data to determine whether the relationship between the aging degree and aging time output by the model is accurate.
[0125] Implement the following DBN training and model building process:
[0126] A DBN structure with three hidden layers is designed, and the number of nodes in each hidden layer is set to 100, 50, and 25 according to the complexity of data features.
[0127] Each RBM was pre-trained and parameter optimization was performed using the contrastive divergence algorithm, with the number of iterations set to 100.
[0128] After pre-training, the whole network was fine-tuned using the back-propagation algorithm, with the learning rate set to 0.01 and the number of iterations set to 500.
[0129] The model performance was evaluated through cross-validation, and the network structure and hyperparameters were adjusted, such as increasing the number of hidden layer nodes or reducing the number of layers, to improve prediction accuracy.
[0130] The optimized model is used for aging assessment prediction, and the output is the aging degree of silicone rubber.
[0131] Through the above steps, a DBN-based aging assessment prediction model can be constructed, which can effectively mine the intrinsic relationship between multimodal data and provide high-precision prediction results for the aging assessment of silicone rubber.
[0132] S400, inputting the composition data, thickness and density of the silicone rubber to be estimated, as well as the temperature, compression or stretching degree, environmental humidity and / or corona degree data and spectral data of the silicone rubber to be estimated in the working state, vibration response signal data and hydrophobic droplet area and contact angle data into the aging prediction model, and the aging prediction model outputs a relationship curve between the aging degree of the silicone rubber to be estimated and the aging time.
[0133] Taking a set of multimodal data of silicone rubber samples to be tested as an example, the following evaluation and curve generation process is implemented:
[0134] Preprocess the multimodal data of the samples to be tested, including data standardization and feature extraction.
[0135] The extracted feature vectors are formatted according to the model requirements and input into the DBN model.
[0136] The DBN model evaluates the feature vector of the sample to be tested and outputs an aging score.
[0137] According to the aging score, an aging curve is drawn to show the aging trend of the sample over time.
[0138] The accuracy of the prediction results can be verified by comparing with the actual aging data, and the model parameters or training data can be adjusted as needed to improve the prediction accuracy.
[0139] Through the above steps, the aging degree of silicone rubber samples can be accurately evaluated, and an intuitive aging curve can be provided to help users better understand the aging process of the material.
[0140] See also Figure 2 Based on the process of the above-mentioned silicone rubber non-destructive aging prediction method, a silicone rubber non-destructive aging prediction system based on multimodal data fusion is designed, including a silicone rubber component analysis module 100, a data acquisition and processing module 200, an aging prediction model construction module 300 and a detection module 400;
[0141] The silicone rubber component analysis module 100 is used to analyze the composition of silicone rubber components or the proportion of components during the preparation of silicone rubber through terahertz spectroscopy to obtain component and content data;
[0142] The data acquisition and processing module 200 is used to collect time domain spectrum data, vibration response signal time domain data, and hydrophobic droplet area and contact angle data of silicone rubber with different compositions, thicknesses, and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees over aging time;
[0143] The aging prediction model building module 300 is used to train and verify the deep belief network using the collected data, and to mine the relationship between the aging degree and aging time of silicone rubber with different compositions, thicknesses and densities at different temperatures, compression or stretching degrees, environmental humidity and corona treatment degrees, so as to obtain an aging prediction model;
[0144] The detection module 400 is used to input the composition data, thickness and density of the silicone rubber to be estimated, as well as the temperature, compression or tension degree, ambient humidity and / or corona degree data and spectral data, vibration response signal data and hydrophobic droplet area and contact angle data of the silicone rubber to be estimated in the working state into the aging prediction model, and the aging prediction model outputs a relationship curve between the aging degree of the silicone rubber to be estimated and the aging time.
[0145] Among them, the composition and content of silicone rubber are analyzed by terahertz spectroscopy or the proportion of components during the preparation of silicone rubber to obtain composition and content data, including:
[0146] The silicone rubber sample is scanned using terahertz spectroscopy technology to obtain its spectral data;
[0147] The acquired spectral data were preprocessed to eliminate noise and background interference, and the Legendre moment method was used to extract features from the spectral data;
[0148] The extracted features are used to construct a three-dimensional image of the sample absorbance spectrum and converted into a grayscale image;
[0149] The least squares-partial least squares regression model was used to perform regression analysis on the features and obtain the composition and content data of silicone rubber.
[0150] The deep belief network includes a visible layer and multiple hidden layers, each hidden layer has a hidden unit multi-layer restricted Boltzmann machine; the energy function of the restricted Boltzmann machine is for:
[0151]
[0152] in, and are the neuron states of the visible layer and the hidden layer, respectively. and is the bias term, is the connection weight, and the goal of the network is to minimize the energy function;
[0153] By training the restricted Boltzmann machine layer by layer and using the features learned by the restricted Boltzmann machine as the input of the next restricted Boltzmann machine, it is used to learn deep features and the associations between them, thereby realizing feature mining and fusion of the collected data.
[0154] Among them, the collected data is used to train and verify the deep belief network, including:
[0155] Normalize the collected time-domain spectral data, vibration response signal time-domain data, and hydrophobic droplet area data of silicone rubber with different compositions, thicknesses, and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees over aging time to form training sets and validation sets;
[0156] Pre-train the multi-layer restricted Boltzmann machine in the deep belief network and use the contrastive divergence algorithm to optimize the connection weights and bias terms;
[0157] Use the back-propagation algorithm to fine-tune the entire deep belief network, calculate the adjustment amount of the connection weights and bias items of each hidden layer, and achieve fine-tuning of the entire deep belief network;
[0158] When the training meets the prediction accuracy requirements, the data in the validation set is used to input the composition data, thickness and density of the silicone rubber as well as the temperature, compression or stretching degree, ambient humidity and / or corona degree data and spectral data, vibration response signal data and hydrophobic droplet area and contact angle data to determine whether the relationship between the aging degree and aging time output by the model is accurate.
[0159] The area data of the hydrophobic droplets is collected using the Grab Cut algorithm, including:
[0160] Collect raw images of hydrophobic droplets;
[0161] Select the approximate hydrophobic droplet area in the original image;
[0162] Construct a st image of the approximate hydrophobic droplet area, and regard each pixel in the image as a node in the network graph;
[0163] Define the weights between nodes in the network graph based on the positive correlation between the similarities between nodes;
[0164] Determine relevant pixels of the foreground and the background, and determine independent nodes and edges connecting the independent nodes based on the relevant pixels of the foreground and the background;
[0165] Construct an energy function based on independent nodes, their edges and weights;
[0166] Through multiple iterations, the energy function is converged to the minimum, and the boundary nodes between the image subject and the background are obtained;
[0167] By calculating the area of the main body of the image, the area data of the hydrophobic droplet is obtained, and by calculating the contact angle of the main body of the image, the contact angle data of the hydrophobic droplet is obtained.
[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described with reference to the preferred embodiments of the present invention, it should be understood by those skilled in the art that various changes may be made in form and details without departing from the spirit and scope of the present invention as defined in the appended claims.
Claims
1. A non-destructive aging prediction method for silicone rubber based on multimodal data fusion, characterized in that: include: The composition and content of silicone rubber are analyzed by terahertz spectroscopy or the proportion of components during the preparation of silicone rubber to obtain composition and content data; Collect the time domain spectrum data, vibration response signal time domain data, hydrophobic droplet area and contact angle data of silicone rubber with different compositions, thicknesses and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees over aging time; The collected data was used to train and verify the deep belief network, and the relationship between the aging degree and aging time of silicone rubber with different compositions, thicknesses and densities at different temperatures, compression or stretching degrees, ambient humidity and corona treatment degrees was explored to obtain an aging prediction model. The composition data, thickness and density of the silicone rubber to be estimated, as well as the temperature, compression or tension degree, ambient humidity and / or corona degree data and spectral data of the silicone rubber to be estimated in the working state, vibration response signal data and hydrophobic droplet area and contact angle data are input into the aging prediction model, and the aging prediction model outputs a relationship curve between the aging degree of the silicone rubber to be estimated and the aging time.
2. The method for non-destructive aging prediction of silicone rubber based on multimodal data fusion according to claim 1 is characterized in that: The composition and content data of the silicone rubber are obtained by analyzing the composition of the silicone rubber by terahertz spectroscopy or by analyzing the composition ratio during the preparation of the silicone rubber, including: The silicone rubber sample is scanned using terahertz spectroscopy technology to obtain its spectral data; The acquired spectral data were preprocessed to eliminate noise and background interference, and the Legendre moment method was used to extract features from the spectral data; The extracted features are used to construct a three-dimensional image of the sample absorbance spectrum and converted into a grayscale image; The least squares-partial least squares regression model was used to perform regression analysis on the features and obtain the composition and content data of silicone rubber.
3. The method for non-destructive aging prediction of silicone rubber based on multimodal data fusion according to claim 1, characterized in that: The deep belief network includes a visible layer and multiple hidden layers, each hidden layer has a hidden unit multi-layer restricted Boltzmann machine; the energy function of the restricted Boltzmann machine for: in, and are the neuron states of the visible layer and the hidden layer, respectively. and is the bias term, is the connection weight, and the goal of the network is to minimize the energy function; By training the restricted Boltzmann machine layer by layer and using the features learned by the restricted Boltzmann machine as the input of the next restricted Boltzmann machine, it is used to learn deep features and the associations between them, thereby realizing feature mining and fusion of the collected data.
4. The method for non-destructive aging prediction of silicone rubber based on multimodal data fusion according to claim 3 is characterized in that: The training and verification of the deep belief network using the collected data includes: Normalize the collected time-domain spectral data, vibration response signal time-domain data, and hydrophobic droplet area data of silicone rubber with different compositions, thicknesses, and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees over aging time to form training sets and validation sets; Pre-train the multi-layer restricted Boltzmann machine in the deep belief network and use the contrastive divergence algorithm to optimize the connection weights and bias terms; Use the back-propagation algorithm to fine-tune the entire deep belief network, calculate the adjustment amount of the connection weights and bias items of each hidden layer, and achieve fine-tuning of the entire deep belief network; When the training meets the prediction accuracy requirements, the data in the validation set is used to input the composition data, thickness and density of the silicone rubber as well as the temperature, compression or stretching degree, ambient humidity and / or corona degree data and spectral data, vibration response signal data and hydrophobic droplet area and contact angle data to determine whether the relationship between the aging degree and aging time output by the model is accurate.
5. The method for non-destructive aging prediction of silicone rubber based on multimodal data fusion according to claim 1, characterized in that: The area data of the hydrophobic droplets is collected by using a Grab Cut algorithm, including: collecting an original image of the hydrophobic droplets; selecting a rough area of the hydrophobic droplets in the original image; Construct a rough ST image of the hydrophobic droplet area, and regard each pixel in the image as a node in the network graph; define the weights between the nodes in the network graph according to the positive correlation of the similarity between the nodes; Determine relevant pixels of the foreground and the background, and determine independent nodes and edges connecting the independent nodes based on the relevant pixels of the foreground and the background; Construct an energy function based on independent nodes, their edges and weights; Through multiple iterations, the energy function is converged to the minimum, and the boundary nodes between the image subject and the background are obtained; By calculating the area of the main body of the image, the area data of the hydrophobic droplet is obtained, and by calculating the contact angle of the main body of the image, the contact angle data of the hydrophobic droplet is obtained.
6. A non-destructive aging prediction system for silicone rubber based on multimodal data fusion, characterized in that: It includes silicone rubber component analysis module, data acquisition and processing module, aging prediction model building module and detection module; The silicone rubber component analysis module is used to analyze the composition of silicone rubber components or the proportion of components during the preparation of silicone rubber through terahertz spectroscopy to obtain component and content data; The data acquisition and processing module is used to collect time-domain spectrum data, vibration response signal time-domain data, and hydrophobic droplet area and contact angle data of silicone rubber with different compositions, thicknesses, and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees over aging time; The aging prediction model building module is used to train and verify the deep belief network using the collected data, and to mine the relationship between the aging degree and aging time of silicone rubber with different compositions, thicknesses and densities at different temperatures, compression or stretching degrees, ambient humidity and corona treatment degrees, so as to obtain an aging prediction model; The detection module is used to input the composition data, thickness and density of the silicone rubber to be estimated, as well as the temperature, compression or tension degree, ambient humidity and / or corona degree data and spectral data of the silicone rubber to be estimated in the working state, vibration response signal data and hydrophobic droplet area and contact angle data into the aging prediction model, and the aging prediction model outputs a relationship curve between the aging degree of the silicone rubber to be estimated and the aging time.
7. The silicone rubber non-destructive aging prediction system based on multimodal data fusion according to claim 6 is characterized in that: The composition and content data of the silicone rubber are obtained by analyzing the composition of the silicone rubber by terahertz spectroscopy or by analyzing the composition ratio during the preparation of the silicone rubber, including: The silicone rubber sample is scanned using terahertz spectroscopy technology to obtain its spectral data; The acquired spectral data were preprocessed to eliminate noise and background interference, and the Legendre moment method was used to extract features from the spectral data; The extracted features are used to construct a three-dimensional image of the sample absorbance spectrum and converted into a grayscale image; The least squares-partial least squares regression model was used to perform regression analysis on the features and obtain the composition and content data of silicone rubber.
8. The silicone rubber non-destructive aging prediction system based on multimodal data fusion according to claim 6, characterized in that: The deep belief network includes a visible layer and multiple hidden layers, each hidden layer has a hidden unit multi-layer restricted Boltzmann machine; the energy function of the restricted Boltzmann machine for: in, and are the neuron states of the visible layer and the hidden layer, respectively. and is the bias term, is the connection weight, and the goal of the network is to minimize the energy function; By training the restricted Boltzmann machine layer by layer and using the features learned by the restricted Boltzmann machine as the input of the next restricted Boltzmann machine, it is used to learn deep features and the associations between them, thereby realizing feature mining and fusion of the collected data.
9. The silicone rubber non-destructive aging prediction system based on multimodal data fusion according to claim 8, characterized in that: The training and verification of the deep belief network using the collected data includes: Normalize the collected time-domain spectral data, vibration response signal time-domain data, and hydrophobic droplet area data of silicone rubber with different compositions, thicknesses, and densities at different temperatures, compression or stretching degrees, ambient humidity, and corona treatment degrees over aging time to form training sets and validation sets; Pre-train the multi-layer restricted Boltzmann machine in the deep belief network and use the contrastive divergence algorithm to optimize the connection weights and bias terms; Use the back-propagation algorithm to fine-tune the entire deep belief network, calculate the adjustment amount of the connection weights and bias items of each hidden layer, and achieve fine-tuning of the entire deep belief network; When the training meets the prediction accuracy requirements, the data in the validation set is used to input the composition data, thickness and density of the silicone rubber as well as the temperature, compression or stretching degree, ambient humidity and / or corona degree data and spectral data, vibration response signal data and hydrophobic droplet area and contact angle data to determine whether the relationship between the aging degree and aging time output by the model is accurate.
10. The silicone rubber non-destructive aging prediction system based on multimodal data fusion according to claim 6, characterized in that: The area data of the hydrophobic droplets is collected by using a Grab Cut algorithm, including: collecting an original image of the hydrophobic droplets; selecting a rough area of the hydrophobic droplets in the original image; Construct a rough ST image of the hydrophobic droplet area, and regard each pixel in the image as a node in the network graph; define the weights between the nodes in the network graph according to the positive correlation of the similarity between the nodes; Determine relevant pixels of the foreground and the background, and determine independent nodes and edges connecting the independent nodes based on the relevant pixels of the foreground and the background; Construct an energy function based on independent nodes, their edges and weights; Through multiple iterations, the energy function is converged to the minimum, and the boundary nodes between the image subject and the background are obtained; By calculating the area of the main body of the image, the area data of the hydrophobic droplet is obtained, and by calculating the contact angle of the main body of the image, the contact angle data of the hydrophobic droplet is obtained.
Citation Information
Patent Citations
Multi-dimensional parameter assessment method of aging degree of running composite insulator
CN106771765A
Non-destructive aging measurement method for composite insulating materials suitable for field implementation
CN109580690A
Method for applying data fusion in ambient air forecast based on deep belief network (DBN)
CN114077883A
Composite insulator aging detection method and device based on terahertz spectrum
CN114397266A
Method for evaluating aging state of silicone rubber for insulator based on multi-parameter identification
CN117517880A
Cited By
Composite insulator silicone rubber material service life evaluation method based on alkaline hydrolysis aging simulation
CN120142135A
