Composite material impregnation optimization method based on fusion network

Abstract

The present invention proposes a composite material impregnation optimization method based on a fusion network, in which a glue layer is first applied to the fiber surface, and then the impregnation process is carried out. However, as the process flow increases, the coupling parameters of the overall process also increase, presenting a deep coupling state. In order to solve the long-term dependency problem of traditional RNN networks in multi-factor long-time series application scenarios, this application proposes a T-BiRNN network architecture, which uses the principle based on random multi-coupling characteristics to analyze the important influencing process parameters of processes such as fiber spreading, spraying and impregnation. And during the propagation process, the network adds the weight parameters of the spray-applied glue layer, and performs adjustment learning to observe the influence coefficient of the spray-applied glue layer on the final result. This application can significantly improve the efficiency and quality of the composite material impregnation process.

Description

Technical Field

[0001] The present invention relates to the field of composite materials, in particular to a composite material impregnation optimization method based on a fusion network. Background Art

[0002] Reinforced manufacturing of composite materials is a novel and highly sought-after research direction. It is also an advanced manufacturing technology that embeds reinforcing materials into matrix materials to form composite materials with excellent performance. Its main purpose is to overcome the performance limitations of traditional single materials. Currently, traditional methods include manual spraying, compression impregnation, and vacuum impregnation, but these methods often have defects. For example, in traditional impregnation methods, the resin penetration of composite materials is often uneven, resulting in inconsistent performance. At the same time, pores, voids, or incompletely impregnated areas may appear, affecting the strength and durability of the material. Therefore, there is an urgent need for a more complete impregnation process and algorithm model to optimize the application of the new process. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention provides a method for monitoring on-site personnel in a substation monitoring scenario, which can realize the intelligentization and precision of composite material impregnation.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] The composite material impregnation optimization method based on the fusion network includes the following steps:

[0006] 1) Impregnation process optimization:

[0007] During the fiber spreading process, a thermoplastic sizing layer is sprayed on the surface of the fiber to cover the surface with a layer of polar molecules, which increases the bonding performance between the resin material and the fiber. After the sizing layer is sprayed, it passes through the melting zone and is impregnated with the final resin material to improve the performance of the reinforced wire material.

[0008] 2) Parameter setting and data processing;

[0009] The collected characteristic data include the tension, speed, spacing, tensile force, temperature during the fiber spreading process, the pressure of the glue during the fiber spreading process, the spraying distance, angle, gas flow, the thickness of the spray layer, and the final impregnation time, impregnation temperature, resin concentration, and impregnation rate. That is, a total of 14 characteristic data are recorded, x1, x2, ..., x 14 Corresponding to the characteristic parameters that need to be collected, the target values that need to be mapped include two aspects: the degree of impregnation and the number of internal defects. After the characteristic parameters and target values are characterized, all values are standardized;

[0010] 3) T-BiRNN network architecture setting;

[0011] T-BiRNN network architecture,The theme architecture is divided into two architectures, which are divided into a single feature selection network on the left and a process area feature selection network on the right.,Both networks were trained separately during this application process and merged,output during prediction;

[0012] 4) Analysis of the influence of sizing parameters;

[0013] In this step, the features are shuffled, which disrupts the order of the original data and presents it in a disordered state. Then, the features are segmented according to the segmentation formula, that is, they are segmented into several groups of sub-vectors of different sizes. There is no overlapping area between each sub-vector. The individual spray coating layer is used as the main input value, and the mapping extraction formula is used to map the fiber spreading and impregnation features to form a Skip-gram structure.

[0014] 5) Long-term dependence avoidance;

[0015] Based on the intermediate output values of the sub-vectors after the segmentation in step 4), a method is proposed to avoid the long-term dependency problem. Specifically, the high-order feature information extracted from the sub-vectors is intercepted in advance and integrated into n groups of Vec vectors. The decision formula is used to perform an integrated analysis on these n groups of Vec vectors.

[0016] 6) Regional time series analysis;

[0017] In order to evaluate the impact of the newly added process on fiber spreading and impregnation, the original 14 features are spliced according to the region to form three feature vectors, namely fiber spreading, sizing, and impregnation. The dimension of each vector is 1*5. If the length is not enough, the <pad>Fill in the blanks and directly analyze the three vectors using RNN. When the network training is completed, that is, when the predicted values of the degree of impregnation and the number of internal defects output by the network have a good fit with the actual values, the adjustment formula is used to determine the importance of the sizing process to the overall process.

[0018] 7) Deployment and application of algorithm models;

[0019] This step deploys the application. First, perform a regional analysis based on step 6) to check whether the adhesive layer has a sufficient impact on the current process. Next, make adjustments to the specific parameters. Analyze which original process parameters will have a high impact on the individual features of the adhesive layer, not in step 3). Finally, analyze whether high-order truncation can be used through step 5) to reduce the length of the entire vector, reduce long-term dependencies, increase calculation speed, and output the predicted value.

[0020] As a further improvement of the present invention, the segmentation formula in step 4) is expressed as:

[0021] The segmentation formula is expressed as follows:

[0022]

[0023]

[0024] m,n,k=Min(Σ(OutPut_1(m,n,k)-True(m,n,k)))

[0025] Among them, n is the number of vectors after segmentation, m is the number of features contained in each vector after segmentation, and k represents the kth feature combination method under the classification method of n and m;

[0026] ∑output is the sum of the predicted values of each group of classified vectors. OutPut_1(m, n, k) means that when the number of segmentation vectors is n, each vector contains m features after segmentation, and each specific feature category is determined to be the kth case in each vector. After a lot of training, the output value of OutPut_1 of this group of segmentation methods includes the degree of immersion and the number of internal defects. Min(∑(OutPut_1(m, n, k)-True(m, n, k))) means that the data is divided into n groups, each group contains m features, and the classification type of each group is the kth type. The minimum error between the predicted value and the true value of multiple groups of data. It should be noted that after being divided into n groups, if the number of features in the last group is not enough, directly use <pad>Just fill it;

[0027] The mapping extraction formula in step 4) is expressed as:

[0028] The mapping extraction formula is expressed as:

[0029] x t =f max (∑(x total -x Iiao ))

[0030] x t ∈(glue pressure, spraying distance, angle, gas flow, spray layer thickness)

[0031] Above, x total For all eigenvalues in the subvector, x Iiao Represented as the feature category of the spray glue layer contained in the sub-vector, x t Represents the characteristic parameters of the spray glue layer in the sub-vector, f max Indicates a strong association marker.

[0032] As a further improvement of the present invention, the determination formula in step 5) is expressed as:

[0033] The judgment formula is as follows:

[0034] OutPut_2=∑RNN(Vec1+Vec2+...)

[0035] Arc(∑RNN(Vec1+Vec2+...))>Arc total -α

[0036] Among them, OutPut_2 represents the predicted value after the high-order features extracted from each sub-vector intercepted in advance are trained again in a one-way RNN network, RNN(Vec1+Vec2+...) represents the global one-way RNN training of the high-order feature vector intercepted in advance, Arc total It indicates the accuracy α when all features are considered in the left network and the BiRNN network is used for training. It is expressed as the accuracy correction hyperparameter. By adjusting its value, the target accuracy requirement can be achieved.

[0037] As a further improvement of the present invention, the adjustment formula in step 6) is expressed as:

[0038] The adjustment formula is expressed as follows:

[0039] Res=RNN(Vec 展纤 , β*Vec 施胶 Vec 浸渍 )

[0040]

[0041] Among them, Vec 展纤 , β*Vec 施胶 , Vec 浸渍 They are represented as vectors composed of three stages, and β is a hyperparameter that adjusts Vec by different sizes. 施胶 The value of is used to change its weight in the entire network. RNN means that the above three vectors are trained in a cyclic network. Res means that its output value can be selected as the degree of impregnation or the defects contained inside. Ero means that Vec with different weights is used. 施胶 When its change value, Res β1 ,Res β2 This indicates the output prediction value when taking β1 and β2.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. This application provides a composite material impregnation optimization method based on a fusion network, which proposes a new composite material optimization method that can significantly improve the degree of impregnation and reduce the number of internal defects.

[0044] 2. This application provides a composite material impregnation optimization method based on a fusion network. It proposes a new composite material optimization method and a single feature impact analysis network that can analyze which original process parameters will have a high impact value on a single feature of the adhesive layer, and then guide the addition of new processes and parameter settings.

[0045] 3. This application provides a fusion network-based composite material impregnation optimization method that intercepts high-eigenvectors and performs enhanced analysis. This approach can reduce the length of the entire sequence, thereby avoiding the long-term dependencies caused by the new process and the additional parameters adopted. This improves calculation speed and accuracy.

[0046] 4. The composite material impregnation optimization method based on a fusion network provided in this application proposes a regional timing analysis method for the first time, which conducts an enhanced analysis of the new process proposed in this application to check whether it has a greater impact on the application of a certain material and accelerate the research and development of process parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Flowchart of a composite material impregnation optimization method based on a fusion network provided in an embodiment of the present application.

[0048] Figure 2 Schematic diagram of process optimization of the composite material impregnation optimization method based on fusion network provided in an embodiment of the present application.

[0049] Figure 3 Schematic diagram of the network architecture of the composite material impregnation optimization method based on fusion network provided in an embodiment of the present application. DETAILED DESCRIPTION

[0050] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0051] like Figure 1 Shown is a flow chart of the composite material impregnation optimization method based on fusion network provided in this application.

[0052] Step S1: Impregnation process optimization

[0053] like Figure 2 Shown is a schematic diagram of the optimization process of the composite material impregnation optimization method based on the fusion network provided in this application.

[0054] In the traditional impregnation process, especially the impregnation of continuous fiber reinforced composite materials, there are often problems such as insufficient impregnation, poor resin-fiber bonding, and many internal defects after molding. Therefore, this application proposes a new impregnation process, such as Figure 2 As shown in the figure, a thermoplastic sizing layer is sprayed on the fiber surface during the fiber spreading process, covering the surface with a layer of polar molecules, which improves the bonding performance between the resin material and the fiber. After the sizing layer is sprayed, it passes through the melting zone and is impregnated with the final resin material to improve the performance of the reinforcing wire material.

[0055] Step S2: parameter setting and data processing.

[0056] In step S1, it is proposed to use a spray-applied glue layer to perform impregnation strengthening treatment to increase the degree of impregnation and reduce internal defects of the formed parts. However, this way of increasing the process flow will further expand the parameters of the already complex process and the internal mutual influence and coupling, which is not conducive to specific applications. Therefore, in this application, it is proposed to use a T-BiRNN network architecture to perform algorithm modeling and enhanced analysis on this new process. The characteristic data that need to be collected include the tension, speed, spacing, tensile force, temperature in the fiber spreading process itself, the pressure of the glue applied during the fiber spreading process, the spraying distance, angle, gas flow, the thickness of the spray layer, and the final impregnation time, impregnation temperature, resin concentration, and impregnation rate. That is, a total of 14 characteristic data are entered, x1, x2, ..., x 14 These correspond to the characteristic parameters that need to be collected as described above. The target values that need to be mapped for this application include two aspects: the degree of impregnation and the number of internal defects. After characterizing the characteristic parameters and target values, all values are standardized to reduce network processing difficulty and speed up fitting.

[0057] Step S3: T-BiRNN network architecture setting

[0058] like Figure 3 Shown is a schematic diagram of the network architecture of the composite material impregnation optimization method based on fusion network provided in this application.

[0059] In step S2, the characteristic parameters of this new process are characterized. It can be seen that after adding the new process flow, the number of overall parameters increases significantly, and the mutual coupling dimension increases further. Therefore, when using traditional RNN networks to analyze this long-term multi-factor coupling problem, the gradient disappearance problem will inevitably occur, and it is difficult to analyze the contribution of characteristic parameters. Therefore, this application proposes to use an improved BiRNN network, namely T-BiRNN, for process flow analysis.

[0060] like Figure 3 The following figure shows the T-BiRNN network architecture used in this application. The main architecture is divided into two architectures: a single feature selection network on the left and a process area feature selection network on the right. Both networks were trained separately during the application process and combined for output during prediction.

[0061] Step S4: Analysis of the impact of sizing parameters

[0062] like Figure 3 As shown on the left, all features are fed into the network individually, initially forming 14 one-dimensional vectors, each 1x1. A shuffle operation is then performed on the data, disrupting the original data sequence and rendering it unordered. The data is then partitioned according to the partitioning formula, splitting it into several sub-vectors of varying sizes, with no overlap between them.

[0063] The segmentation formula is expressed as follows:

[0064]

[0065]

[0066] m,n,k=Min(Σ(OutPut_1(m,n,k)-True(m,n,k))) (3)

[0067] Among them, n is the number of vectors after segmentation, m is the number of features contained in each vector after segmentation, and k represents the kth feature combination method under the classification method of n, m. ∑output is the sum of the predicted values of each group of classified vectors. OutPut_1(m, n, k) means that when the number of segmentation vectors is n, each vector contains m features after segmentation, and each specific feature category is determined to be the kth case in each vector, after a lot of training, the output value of OutPut_1 of this group of segmentation methods includes the degree of immersion and the number of internal defects. Min(∑(OutPut_1(m, n, k)-True(m, n, k))) means that when the data is divided into n groups, each group contains m features, and the classification type of each group is the kth type, the minimum error value between the predicted value and the true value of multiple groups of data. It should be noted that after being divided into n groups, if the number of features in the last group is not enough, directly use <pad>Just fill it.

[0068] In formula 1, the main thing to determine is the number of groups, that is, the number of groups into which the 14 eigenvalues are divided, the number of features contained in each group, and if the last group is not enough, the <pad>The filler term is used to fill in the gaps. Formula 2 primarily determines the optimal feature combination within each group after the number of groups is determined, minimizing the error between the predicted value and the output value across all classification methods. This network training method determines the optimal classification method for the sub-vectors and saves the optimal allocation. This means that under this allocation, the network captures the significant temporal correlations within each sub-vector, resulting in a more accurate overall prediction.

[0069] Relationship extraction is performed on the optimal classification method, that is, the correlation between individual features is first extracted. This application focuses on the impact of the spray coating layer on the overall process. Therefore, the spray coating layer is used as the main input value, and a mapping extraction formula is used to map the fiber spreading and impregnation features to form a Skip-gram-like structure.

[0070] The mapping extraction formula can be expressed as:

[0071] x t =f max (∑(x total -x Iiao )) (4)

[0072] x t ∈(sizing pressure, spraying distance, angle, gas flow, spraying layer thickness) (5)

[0073] Above, x total For all eigenvalues in the subvector, x Iiao Represented as the feature category of the spray glue layer contained in the sub-vector, x t Represents the characteristic parameters of the spray glue layer in the sub-vector, f max Indicates a strong association marker.

[0074] The above formula 4 removes the features related to the sizing layer in each sub-vector, and the remaining parameters are the parameters of fiber spreading and impregnation. t The parameters of the sizing layer are strongly correlated with the fiber spreading and impregnation parameters in the subvectors. This mapping relationship, when applied in practice, indicates which existing process parameters are most significantly affected by the individual sizing layer features proposed in this application, thus providing guidance for the addition of new processes and parameter settings.

[0075] Step S5: Avoid long-term dependencies

[0076] In addition, this application can propose a method to avoid the problem of long-term dependency based on the intermediate output value of the sub-vector after the above division. Figure 3 As shown in FIG, the specific performance is to intercept the high-order feature information extracted from the sub-vector in advance, integrate it into n groups of Vec vectors, and use the judgment formula to perform integrated analysis on these n groups of Vec vectors.

[0077] The judgment formula is as follows:

[0078] OutPut_2=∑RNN(Vec1+Vec2+...) (6)

[0079] Arc(∑RNN(Vec1+Vec2+...))>Arc total -α (7)

[0080] Among them, OutPut_2 represents the predicted value after the high-order features extracted from each sub-vector intercepted in advance are trained again in a one-way RNN network, and RNN(Vec1+Vec2+...) represents the global one-way RNN training of the high-order feature vector intercepted in advance. total This indicates the accuracy of the BiRNN network when all features are considered. α is a precision correction hyperparameter, and adjusting its value can achieve the target accuracy.

[0081] Using Formulas 6 and 7 above, we can calculate the accuracy of the predicted output when combining the extracted high-order vectors into a single vector group and retraining the RNN network. For example, when taking two vector groups, Vec1 and Vec2, where Vec1 and Vec2 represent the new vector formed by combining the high-order eigenvectors of the first segmented sub-vectors, retraining the RNN network with these two high-order eigenvectors reveals that their accuracy meets certain requirements. This approach reduces the length of the entire sequence, thereby avoiding the long-term dependencies caused by the new process and the additional parameters. This improves computational speed and accuracy.

[0082] Step S6: Regional timing analysis.

[0083] This application proposes a regional timing analysis method for the first time, which strengthens the analysis of the new process proposed in this application, such as Figure 3 As shown on the right, the characteristic parameters that affect the final target value in this application include the tension, speed, spacing, tensile force, temperature during the fiber spreading process, the pressure of glue application during the fiber spreading process, spraying distance, angle, gas flow, thickness of the spray layer, and the final impregnation time, impregnation temperature, resin concentration, and impregnation rate. In order to evaluate the impact of the newly added overall process flow on fiber spreading and impregnation in this application, the original 14 features are spliced according to the region to form three characteristic vectors, namely fiber spreading, gluing, and impregnation. The dimension of each vector is 1*5, and if the length is not enough, it is used. <pad>Fill in. The three vectors mentioned above are directly analyzed using RNN. After the network training is completed, that is, when the predicted values of the degree of impregnation and the number of internal defects output by the network have a good fit with the actual values, the following adjustment formula is used to determine the importance of the sizing process to the overall process.

[0084] The adjustment formula is expressed as follows:

[0085] Res=RNN(Vec 展纤 , β*Vec 施胶 , Vec 浸渍 ) (8)

[0086]

[0087] Among them, Vec 展纤 , β*Vec 施胶 , Vec 浸渍 They are represented as vectors composed of three stages, and β is a hyperparameter that adjusts Vec by different sizes. 施胶 The value of is used to change its weight in the entire network. RNN means that the above three vectors are trained in a cyclic network. Res means that its output value can be selected as the degree of impregnation or the defects contained inside. Ero means that Vec with different weights is used. 施胶 When its change value, Res β1 ,Res β2 This indicates the output prediction value when taking β1 and β2.

[0088] The above formula 8 can be used to obtain the predicted output value of the process with different weights. Formula 9 is mainly used to analyze the importance of the added sizing layer process. By calculating the difference in their final performance when different weights are used, if the difference between β1 and β2 is small, but Res β1 ,Res β2 If the floating surface fluctuation is large, it indicates that the spray coating process proposed this time has great application prospects for this material.

[0089] Step S7: Deploy and apply the algorithm model.

[0090] After the above algorithm training settings are completed, it can be deployed and applied. First, a regional analysis is performed according to step S6 to check whether the glue layer can have a sufficient impact on this process. Secondly, adjustments are made in the specific parameters. Step S3 is used to analyze which original process parameters have a high impact value on the single feature of the glue layer. Finally, step S5 is used to analyze whether high-order truncation can be used to reduce the length of the entire vector, reduce long-term dependencies, increase calculation speed, and output the predicted value.

[0091] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.< / pad> < / pad> < / pad> < / pad> < / pad>

Claims

1. A composite material impregnation optimization method based on a fusion network, characterized by: The steps include: 1) Impregnation process optimization: During the fiber spreading process, a thermoplastic sizing layer is sprayed on the surface of the fiber to cover the surface with a layer of polar molecules, which increases the bonding performance between the resin material and the fiber. After the sizing layer is sprayed, it passes through the melting zone and is impregnated with the final resin material to improve the performance of the reinforced wire material. 2) Parameter setting and data processing; By collecting characteristic data including the tension, speed, spacing, tensile force, temperature during the fiber spreading process, the pressure of glue application during the fiber spreading process, spraying distance, angle, gas flow, thickness of the spray layer, and the final impregnation time, impregnation temperature, resin concentration, impregnation rate, a total of 14 characteristic data are recorded. Corresponding to the characteristic parameters that need to be collected, the target values that need to be mapped include two aspects: the degree of impregnation and the number of internal defects. After the characteristic parameters and target values are characterized, all values are standardized; 3) T-BiRNN network architecture setting; The T-BiRNN network architecture is divided into two main architectures: a single feature selection network on the left and a process area feature selection network on the right. Both networks are trained separately and combined for output during prediction. 4) Analysis of the impact of sizing parameters; In this step, the features are shuffled, which disrupts the order of the original data and presents a disordered state. Then, the features are segmented according to the segmentation formula, that is, they are segmented into several groups of sub-vectors of different sizes, where there is no overlapping area between each sub-vector. The individual spray coating layer is used as the main input value, and the mapping extraction formula is used to map the fiber spreading and impregnation features to form a Skip-gram structure. 5) Long-term dependence avoidance; Based on the intermediate output values of the sub-vectors after the segmentation in step 4), a method is proposed to avoid the long-term dependency problem. Specifically, the high-order feature information extracted from the sub-vectors is intercepted in advance and integrated into n groups of Vec vectors. The decision formula is used to perform a comprehensive analysis on these n groups of Vec vectors. 6) Regional time series analysis; In order to evaluate the impact of the newly added process on fiber spreading and impregnation, the original 14 features are spliced according to the region to form three feature vectors, namely fiber spreading, sizing, and impregnation. The dimension of each vector is 1*5. If the length is not enough, the <pad> Fill in the blanks and directly analyze the three eigenvectors using RNN. When the network training is completed, that is, when the predicted values of the degree of impregnation and the number of internal defects output by the network have a good fit with the actual values, the adjustment formula is used to determine the importance of the sizing process to the overall process.< / pad> 7) Algorithm model deployment and application; This step involves deploying and applying the application. First, a regional analysis is performed based on step 6) to determine whether the adhesive layer has a sufficient impact on the process. Next, adjustments are made to specific parameters. Step 4) is used to analyze which individual features of the adhesive layer have a high impact on existing process parameters. Finally, step 5) analyzes whether high-order truncation can be used to reduce the length of the entire vector, reducing long-term dependencies and improving calculation speed, ultimately outputting a predicted value.

2. The composite material impregnation optimization method based on fusion network according to claim 1, characterized in that: The segmentation formula in step 4) is expressed as: The segmentation formula is expressed as follows: ； ； ； Among them, n is the number of vectors after segmentation, m is the number of features contained in each vector after segmentation, and k represents the kth feature combination method under the classification method of n and m; is the sum of the predicted values of each group of classified vectors, It means that when the number of segmentation vectors is n, each vector contains m features after segmentation, and each feature category is determined as the kth case in each vector, after a large amount of training, the output value of OutPut_1 of this group of segmentation methods includes the degree of impregnation and the number of internal defects. It means that the data is divided into n groups, each group contains m features, and the classification type of each group is k, the minimum error between the predicted value and the true value of the multiple groups of data. It should be noted that after being divided into n groups, if the number of features in the last group is not enough, directly use <pad> Just fill it;< / pad> The mapping extraction formula in step 4) is expressed as: The mapping extraction formula is expressed as: ； ； The above, are all eigenvalues in the subvector, Represented as the feature category of the spray glue layer contained in the sub-vector, Represents the characteristic parameters of the spray glue layer in the sub-vector, Indicates a strong association marker.

3. The composite material impregnation optimization method based on fusion network according to claim 1, characterized in that: The determination formula in step 5) is expressed as: The judgment formula is as follows: ； ； in, It represents the predicted value after the high-order features extracted from each sub-vector intercepted in advance are trained again on the unidirectional RNN network. Indicates that the high-order feature vectors intercepted in advance will be globally trained in a one-way RNN. This shows the accuracy when all features are considered in the left network and trained using the BiRNN network. It is expressed as a precision correction hyperparameter, and the target accuracy requirement can be achieved by adjusting its value.

4. The composite material impregnation optimization method based on fusion network according to claim 1, characterized in that: The adjustment formula in step 6) is expressed as: The adjustment formula is expressed as follows: ； ； in, Represented as vectors composed of three stages, Tuning hyperparameters with different sizes The value of , changes its weight in the entire network, Indicates that the vectors composed of the three stages are trained on a cyclic network. Indicates that its output value can be selected as the degree of impregnation or the defects contained inside, This means using different weights When its changing value, This indicates that when The output prediction value when .

Images