Neural network-based wood-plastic composite material static bending intensity prediction method

Abstract

The invention provides a wood-plastic composite material static bending intensity prediction method based on a neural network. The wood-plastic composite material static bending intensity prediction method comprises the following steps: S1, data collection; s2, data preprocessing; s3, feature extraction; extracting features from the preprocessed data, and performing dimension reduction processing on the features to obtain extracted features; s4, constructing a neural network model based on the extracted features, wherein the neural network model adopts a multi-layer perceptron (MLP), a convolutional neural network (CNN) or a long-short term memory (LSTM) network architecture; s5, training a neural network model by using the preprocessed data to obtain a trained model; s6, evaluating the trained model to obtain an evaluation result; s7, optimizing the neural network model according to an evaluation result to obtain an optimized model; and S8, inputting data of the wood-plastic composite material to be predicted into the optimized model, and outputting a prediction result. According to the method, the reliability of a prediction result is improved, and a more accurate basis is provided for product design and quality control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wood-plastic composites, in particular, and more particularly to a wood-plastic composite static bending strength prediction method based on neural networks. BACKGROUND

[0002] As a new type of environmentally friendly material, wood-plastic composites have been widely used in many fields such as construction, furniture, and packaging due to their advantages of both wood and plastic. In practical applications, static bending strength, as a key indicator of the mechanical properties of wood-plastic composites, plays a decisive role in product quality and service life.

[0003] Traditional wood-plastic composite static bending strength prediction methods have obvious limitations. Empirical formulas are mainly based on the linear relationship of simple material properties and process parameters, which oversimplify the complex interactions within the material and various influencing factors in the actual processing process, resulting in poor prediction accuracy and difficulty in meeting the needs of modern industry for accurate product quality control. Experimental testing methods can obtain more accurate data, but require the preparation of a large number of samples, which not only consumes a large amount of raw materials, time and manpower, but also relies on professional testing equipment, resulting in a significant increase in product development and production costs. In addition, experimental testing is a post-test method that cannot effectively predict and optimize material properties before production, making it difficult to adapt to the pace of modern industrial rapid production and quality control.

[0004] With the development of computer technology, the finite element analysis (FEA) method has been introduced into the field of material property prediction. However, for wood-plastic composites, which are complex in internal structure and diverse in material properties, establishing an accurate finite element model is a major challenge. Its accuracy is highly dependent on the accuracy of the model, and the modeling process requires complex assumptions and simplifications of the material's microstructure and physical processes, as well as high computational costs, requiring specialized modeling and analysis personnel, which limits the widespread promotion of this method in the industry. In recent years, neural network technology has shown great potential in the field of materials science, capable of learning complex nonlinear relationships in data without the need for accurate physical models. However, its application in wood-plastic composite static bending strength prediction is still in the exploratory stage, facing challenges such as how to collect and organize high-quality data, process data from different sources, select appropriate neural network architectures and training methods, and ensure the accuracy and generalization ability of the model. Currently, there is a lack of a systematic and comprehensive neural network-based static bending strength prediction method and system in the field of wood-plastic composites, and there is an urgent need to develop related technologies to address the shortcomings of traditional methods and existing technologies and promote the further development of the wood-plastic composite industry. SUMMARY

[0005] Therefore, the present application aims to provide a wood-plastic composite static strength prediction method based on a neural network to solve the technical problem that the existing finite element analysis method cannot be used for wood-plastic composite static strength prediction.

[0006] The technical means adopted by the present application are as follows: A wood-plastic composite static strength prediction method based on a neural network comprises the following steps: S1, data collection and cleaning; collecting sample data of wood-plastic composites, the sample data sources including laboratory test data, factory production data and public data sets; using a data cleaning algorithm to preliminarily screen and remove erroneous data from the sample data, the erroneous data including data beyond the set range; S2, preprocessing the sample data; fusing the laboratory data using a sensor calibration data fusion algorithm; integrating the factory production data according to the production process logical relationship; processing the public data using an adaptive outlier detection algorithm; obtaining the preprocessed data; S3, feature extraction; extracting virtual features of wood-plastic composites from the preprocessed data using a generative adversarial network, obtaining the extracted features; S4, constructing a neural network model based on Transformer-CNN fusion based on the extracted features; S5, using the preprocessed data to adjust the learning rate using an adaptive learning rate adjustment strategy, combining the learning rate warm-up and cosine annealing mechanism to train the neural network model, obtaining the trained model; S6, using a combination of leave-one-out and Monte Carlo cross-validation evaluation methods to evaluate the trained model, obtaining the evaluation results; S7, inputting the data of the wood-plastic composite to be predicted into the optimized model and using a knowledge graph to assist in prediction, and outputting the prediction results.

[0007] Further, it also includes: S8, dynamic data updating and model evolution: establishing a real-time data monitoring system, collecting new wood-plastic composite data in real time, when the amount of new data reaches the rated threshold, updating the model using an incremental learning algorithm, the formula is as follows: , wherein, is the updated model parameter, is the original model parameter, is the learning rate, is the loss function gradient of the new data , periodically evaluate the model performance, when the accuracy of the model on the validation set decreases by more than 5% for 3 consecutive evaluations, start the model evolution process, and re-adjust the network structure and hyperparameters.

[0008] Further, it also includes: S9, microstructure-property correlation prediction: Obtain the microstructure image of wood-plastic composite material by atom probe tomography and transmission electron microscope, establish a quantitative correlation model between microstructure characteristics and static bending strength, and analyze the microstructure image by deep learning method, as follows: , Wherein, is the interface bonding strength index, x i , y i is the feature vector of the material on both sides of the interface, and n is the number of sampling points on the interface. The microstructure characteristic parameters are taken as additional input and integrated into the neural network model.

[0009] Further, in S1, metadata is recorded using blockchain technology, and metadata is stored on nodes using the distributed ledger feature of blockchain. The data is encrypted by a hash function H(x), as follows: , The data from different sources is quality rated, and the weight of the training data is adjusted according to the data quality, as follows: , Wherein, w i is the weight of the i-th data, q i is the data quality score, and e is the total number of data.

[0010] Further, in S4, the CNN module uses the dilated convolution technique, as follows: , Wherein, y[i] is the convolution output, is the input data, w[j] is the convolution kernel weight, r is the dilated rate, and h is the convolution kernel size; Capsule network structure is introduced, and the capsule network uses a dynamic routing mechanism, as follows: , Wherein, c ij is the routing coefficient, b ij is the initial coupling coefficient, and K is the number of high-level capsules.

[0011] Further, in S5, the adversarial training technique is used to add adversarial samples during the training process, as follows: , Wherein, is the adversarial sample, x is the original sample, is the perturbation strength, J( x,y) is a loss function, is a model parameter, is a gradient of the loss function on the sample x; The model is trained in combination with the model integration technology, and a weighted fusion method is adopted, and the formula is as follows: , wherein, is a final prediction value, is a prediction value, is a weight, and ; The transfer learning technology is adopted, the parameters of the pre-trained model are migrated to the wood-plastic composite material static strength prediction model, the migration parameters are fine-tuned according to the material data characteristics, and the formula is as follows: , wherein, is a fine-tuned parameter, is a pre-trained model parameter, is a fine-tuning learning rate, is a loss function gradient of the wood-plastic composite material data.

[0012] Further, in S6, the model complexity index and the model explainability index are introduced, the feature importance calculation method based on Shapley value is adopted, and the formula is as follows: , wherein, is a Shapley value of feature i, S is a feature subset, N is a set of all features, and v(S) is a value function of the feature subset S.

[0013] Further, in S3, the wood fiber and plastic features are analyzed in multiple scales, and the wavelet packet transform is adopted, and the formula is as follows: , , wherein, h n and g n are wavelet packet filter coefficients, j is a decomposition scale, and k is a translation parameter; The molecular dynamics simulation technology is used to obtain the structure and interaction information of the wood-plastic composite material at the molecular level, and the molecular dynamics simulation results are converted into a feature vector, and the formula is as follows: , wherein, f i is a different feature value obtained by molecular dynamics simulation, and the feature vector is fused with the traditional features.

[0014] Further, in S6, the Monte Carlo cross-validation randomly divides the training set and the test set multiple times, calculates the average evaluation index, and adjusts the model hyperparameters using the Bayesian optimization algorithm, as follows: , wherein, is the posterior probability of the parameter under data D, is the likelihood function, P( ) is the prior probability of the parameter , and P(D) is the marginal probability of data D. The optimization algorithm based on swarm intelligence is combined with the Bayesian optimization algorithm. In the particle swarm optimization algorithm, the particle velocity update formula is as follows: , wherein, is the velocity of particle i at the (t+1) th iteration, w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, is the historical optimal position of particle i, is the global optimal position, is the iteration position of particle i, the global search is performed through the particle swarm optimization algorithm to locate the hyperparameter region, and the Bayesian optimization algorithm is used to search in the region.

[0015] Further, in S2, to address the noise problem in the data, a deep learning-based denoising autoencoder is used for noise reduction processing. The denoising autoencoder learns to reconstruct the original data from the noisy data by adding noise to the input data, as follows: , wherein L is the reconstruction loss, x is the original data, is the added noise, f is the encoder function, and g is the decoder function. The denoising autoencoder is used to remove noise interference in the data.

[0016] Compared with the prior art, the present application has the following advantages: In terms of prediction accuracy, the method can automatically capture the complex interactions and nonlinear relationships within the material by means of the powerful learning ability of the neural network, and can deeply analyze a large amount of wood-plastic composite material data under different formulations and process conditions. Compared with traditional empirical formulas and simple linear models, it can more accurately predict the static strength, providing a reliable basis for product design and quality control, and greatly reducing product quality problems caused by inaccurate strength prediction.

[0017] From the efficiency point of view, this invention avoids a large number of tedious experimental tests. Reduces the consumption of raw materials and the dependence on test equipment, greatly shortens the time of product development and quality detection. Once the model training is completed, the static strength prediction under the new material combination and process condition can quickly give the result, which helps to accelerate the product development cycle, improve the production efficiency, and then reduce the production cost, enhance the competitiveness of enterprises in the market.

[0018] In terms of model adaptability and generalization ability, through careful design of feature extraction and enhancement steps, the micro and macro characteristics of the material are fully tapped. The model can better adapt to different types of wood-plastic composites, whether it is common material combination or special formula material, it can achieve more accurate prediction. At the same time, the system's data collection and processing module can conveniently obtain data from various channels, providing rich data support for the continuous optimization of the model, further improving the performance of the model.

[0019] In addition, the model evaluation and optimization steps of the present application can continuously adjust and improve the model according to the actual situation, ensuring that the model always maintains good performance. The user interaction module is simple and intuitive, and users with different permissions can easily use the system for prediction, evaluation and optimization operations. The system also provides visual evaluation results, making it easy for users to intuitively understand the model performance and provide strong support for decision-making, comprehensively promoting the quality control and performance optimization of wood-plastic composites from research and development to production, and promoting the industry towards intelligent and efficient direction. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0021] Figure 1 The schematic block diagram of the wood-plastic composite static strength prediction method based on neural network proposed by the present application; Figure 2 The average absolute error (MAE) comparison column chart of the prediction results of different models and the actual value; Figure 3 The loss function change line chart in the model training process of the present application; Figure 4 The importance score radar chart of the features on the prediction results of the present model. DETAILED DESCRIPTION

[0022] In the following, the technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative effort should belong to the scope of protection of the present application.

[0023] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device including a series of steps or units does not necessarily have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product, or device.

[0024] As Figures 1-4 shown, the present application provides a wood-plastic composite material static bending strength prediction method based on a neural network, comprising the following steps: S1, data collection and cleaning: collect wood-plastic composite material data from extensive data sources, covering 50 professional laboratories, 30 production factories and 10 authoritative professional databases. The collected data contains more than 10 kinds of wood fibers (such as wood fibers of different origins, including northeast pine wood fibers, southern cedar wood fibers, etc.), more than 8 kinds of plastic raw materials (such as high-density polyethylene HDPE, polypropylene PP, etc.) and more than 20 kinds of processing process parameters (such as extrusion temperature, screw speed, mold pressure, molding time, etc.), ensuring that the total amount of sample data exceeds 3000 groups. A specially designed data cleaning algorithm is used to perform initial screening on the data. The specific steps of this algorithm are as follows: first, the Euclidean distance algorithm is used to calculate the similarity between data records, regarding each data record as a point in a multi-dimensional space, and the dimensions correspond to different data characteristics (such as wood fiber types, plastic proportions, processing temperatures, etc.). For any two data records A and B, the Euclidean distance calculation formula is where n is the number of data characteristics, a i and b iare the values of data records A and B on the ith feature, respectively. Similarity is obtained by calculating the distance, and if the similarity of two records is such that the repetition rate is higher than 80%, one of them is identified and removed. At the same time, in combination with the rule base established by material experts, the data is subjected to secondary screening. For example, according to the rules that the sum of the proportions of wood fibers and plastics cannot exceed 100%, and the processing temperature needs to be within the range that the material can withstand, data that does not meet the material characteristics and process logic is removed. Taking a certain group of data as an example, the proportion of wood fiber is recorded as 80%, and the proportion of plastic is recorded as 40%, the sum of which exceeds 100%, so this group of data is determined to be abnormal and is removed.

[0025] S2, Multimodal data fusion preprocessing: For data from different sources, a variety of fusion techniques are used. For laboratory data, since it is usually collected by multiple sensors, sensor calibration data fusion algorithm is used to fuse these sensor data using Kalman filter algorithm, the formula is as follows:

[0026] wherein, Through this algorithm, measurement errors can be effectively reduced, and data accuracy can be improved. In actual operation, for multiple sensor data measuring the density of wood-plastic composite materials, the standard deviation of the data is reduced after Kalman filter fusion, effectively improving the data quality.

[0027] For factory production data, production process correlation fusion method is used according to the logical relationship of production process. Time series model is established to analyze the correlation between data at different stages of production process, and related data is integrated. For example, in the injection molding process, the data of injection pressure, injection speed and mold temperature changing with time are correlated and fused according to the order of production process, so that the data can better reflect the actual situation of the production process.

[0028] For public data, semantic matching fusion technology is used. With the help of natural language processing tools such as Word2Vec model, the text description in public data is converted into vector form, and the similarity between vectors is calculated to realize the matching and fusion of data. Taking the description of wood-plastic composite materials in a certain public data set as an example, the semantic similarity of the data with other data is calculated by the Word2Vec model, and the data with high relevance is selected for fusion, enriching the diversity of the data.

[0029] An adaptive outlier detection algorithm is used to process the data, the formula is as follows: .

[0030] wherein, T is the threshold value for judging outliers, M is the median of the data, MAD is the mean absolute deviation, and k is the adaptive adjustment coefficient (dynamically adjusted according to the degree of data fluctuation, ranging from 1.5 to 3.5). In practical applications, for a set of processing temperature data, the algorithm detects abnormal temperature values that exceed the threshold and marks and processes them to ensure data reliability.

[0031] For missing data, a filling method based on K-nearest neighbors combined with deep learning is used. A deep neural network is constructed, and the formula is as follows: .

[0032] wherein, is the predicted missing value, are the 2k+1 data points around the missing value, and f is the neural network mapping function. Taking a set of missing wood fiber length data as an example, by selecting k=5, using the surrounding 5 data points as input, after neural network training and prediction, the missing values are successfully filled, and the consistency of the filled values with the actual situation is high.

[0033] S3, deep feature mining and enhancement: using feature generation technology based on generative adversarial network (GAN), virtual features of wood-plastic composite materials are generated. Generator G and discriminator D are trained in opposition, and the formula is as follows: , wherein, x is the real data, z is the random noise, is the real data distribution, p z is the noise distribution. Through adversarial training, virtual features such as the virtual distribution of wood fibers in the plastic matrix and the virtual interaction strength between fibers and plastic are generated. These virtual features further enrich the feature dimension of the data and provide more information for the model.

[0034] Multi-scale analysis is performed on the wood fiber and plastic features, and wavelet packet transform formulas (decomposition) and (reconstruction) are used, wherein h n and g n are wavelet packet filter coefficients, j is the decomposition scale, and k is the translation parameter. In actual operation, the decomposition scale j is set from 1 to 3, and wavelet packet transform is performed on the microstructure image of wood fiber to extract its texture features from different scales. For example, at scale j=1, larger-scale texture information such as the overall distribution trend of wood fiber can be extracted; at scale j=3, finer texture features such as the microtexture details on the surface of wood fiber can be captured. Combined with the material microstructure image, image semantic segmentation technology is used to label the distribution areas of wood fiber and plastic. By calculating the related parameters of the labeled areas, topological features such as distribution uniformity and agglomeration degree are extracted. Taking the distribution uniformity as an example, the calculation formula is as follows: , where, is the area of the wood fiber or plastic uniform distribution part, is the area of the entire image. Through this formula, the uniformity of each component in the material can be quantified, providing important feature basis for subsequent intensity prediction.

[0035] S4, new neural network architecture construction: construct a neural network model based on the fusion of Transformer-CNN. The Transformer module uses multi-head attention mechanism, and the formula is as follows: , where, , is the query, key, and value matrix, d k is the dimension of the key matrix, h is the number of heads, , , , is the weight matrix, which is used to capture the long sequence feature dependence. The CNN module uses the dilated convolution technology to expand the receptive field, and the formula is as follows: , where y[i] is the convolution output, is the input data, w[j] is the convolution kernel weight, r is the dilated rate, h is the convolution kernel size, which is used to extract local features. Combined with the gating mechanism, the Sigmoid gating function controls the flow of information. In the model training process, the gating mechanism can automatically adjust the transmission of information according to the importance of the input features, and enhance the model's ability to capture key features. For example, when the model learns the key feature of the interface bonding strength between wood fibers and plastics, the gating mechanism will increase the transmission weight of this feature information, so that the model pays more attention to this important information.

[0036] S5, reinforcement training and dynamic optimization: adopt adaptive learning rate adjustment strategy, combined with learning rate preheating and cosine annealing mechanism, formula as follows: , where, is the learning rate of the t-th step, is the initial learning rate, is the preheating step number, T is the total training step number. Use the adversarial training technology to add adversarial samples in the training process to enhance the robustness of the model. Through the generation of adversarial samples, the formula is as follows: , where, is the adversarial sample, x is the original sample, is the perturbation strength, J(x,y) is the loss function, is the model parameter, is the gradient of the loss function with respect to the sample x. In combination with the model ensemble technique, multiple models with different initializations are trained, and a weighted fusion method is used, as follows: wherein is the final prediction value, is the prediction value of the i-th model, is the weight, and improve the prediction accuracy and stability. Figure 3 is the "loss function change line chart in the model training process of the invention", the horizontal coordinate is the training round (1-50 rounds), and the vertical coordinate is the loss function value. From the figure, it can be observed that as the training round increases, the loss function value overall presents a downward trend. In the early training stage (1-10 rounds), the loss function value decreases rapidly, from about 8 to about 6.5; in the middle training stage (10-30 rounds), the decrease slows down, and gradually stabilizes between 5 and 6; in the later training stage (30-50 rounds), the loss function value tends to be stable, and finally stabilizes at about 5, indicating that the model gradually converges through training, and the prediction performance is continuously optimized and tends to be stable.

[0037] S6, precise evaluation and intelligent feedback optimization: the evaluation method of leave-one-out and Monte Carlo cross-validation is combined. The leave-one-out method leaves one sample as the test set each time, and the remaining samples as the training set, as follows: , wherein Error is the leave-one-out evaluation error, n is the total number of samples, is the loss function. Monte Carlo cross-validation randomly divides the training set and test set multiple times, and calculates the average evaluation index. According to the evaluation result, the Bayesian optimization algorithm is used to adjust the model hyperparameters, and a probability model of the objective function is constructed, as follows: , wherein, is the posterior probability of the parameter under the data D, is the likelihood function, P(D) is the marginal probability of the data D, and the optimal hyperparameter combination is found to optimize the model performance. Figure 2 ​​​The average absolute error (MAE) of the prediction results of different models and the actual values is compared in a column chart. The horizontal coordinate is the sample number (1-5) of the wood-plastic composite material, and the vertical coordinate is the average absolute error (unit: MPa). In the figure, three different styles of columns represent the traditional regression model (black solid column), the basic neural network model (white hollow column) and the method of the application (diagonal filled column). From the figure, it can be seen that the average absolute error of the method of the application is lower than that of the traditional regression model and the basic neural network model in all samples, especially in sample 2 and sample 4, the error reduction is more obvious, which fully reflects the advantage of the method of the application in prediction accuracy.

[0038] S7, knowledge graph assisted prediction: construct a knowledge graph of wood-plastic composite materials, integrate knowledge of material components, processes, properties, etc. Use knowledge graph embedding technology to map entities and relationships in the knowledge graph to a low-dimensional vector space, as follows: , where e i , e j are entity vectors, r ij is a relationship vector, is a scoring function. When the model is predicting, combine the knowledge graph information and dynamically adjust the weight of the input features through the attention mechanism, as follows: , where Q is the query vector of the prediction model, K and V are knowledge graph embedding vectors. For example, when the model is predicting the modulus of rupture of a certain wood-plastic composite material, through the attention mechanism, the model can automatically adjust the weight of the input features according to the relationship between the material components and properties in the knowledge graph, pay more attention to the features closely related to the modulus of rupture, and thus improve the accuracy and interpretability of the prediction. Figure 4 The importance score radar chart of the features of the application to the prediction results of the model is shown in the figure, which shows the importance scores (value range 0-1) of wood fiber content, additive content, plastic type, fiber length and processing temperature. Among them, the importance score of wood fiber content is the highest, about 0.6; the processing temperature is second; the importance scores of fiber length and plastic type are relatively low; the importance score of additive content is the lowest, about 0.2. The figure clearly reflects the contribution of each feature in the model prediction process, provides an intuitive basis for understanding the model decision logic and optimizing feature selection, and also verifies the effective capture of key features by the model with the assistance of the knowledge graph.

[0039] In the application, it also includes: S8. Dynamic Data Update and Model Evolution: Establish a real-time data monitoring system to collect new wood-plastic composite material data in real time. When the amount of new data reaches a predetermined threshold (e.g., 500 sets), an incremental learning algorithm is used to update the model. The incremental learning algorithm learns from the new data based on the original model, and the formula is as follows: , in, For the updated model parameters, These are the parameters of the original model. For learning rate, For new data The gradient of the loss function. In the model, these parameters exist and have their corresponding storage and computation logic. Original model parameters. The parameters are stored in the model's parameter storage unit and retrieved for computation during incremental learning. The learning rate η is a pre-set hyperparameter that determines the step size for each parameter update. For the gradient of the loss function... During calculation, new data First, the data is input into the model. After forward propagation, the predicted value is calculated. Then, the predicted value is compared with the actual value y. new Calculate the loss function (e.g., mean squared error loss function). ,in (The predicted value is used as the basis for calculation), and then the loss function with respect to the model parameters is calculated using the backpropagation algorithm. gradient Finally, the model parameters are updated according to the above incremental learning formula to obtain... .

[0040] Regularly evaluate model performance. When the model's accuracy on the validation set drops by more than 5% in three consecutive evaluations, initiate the model evolution process. During model evolution, readjust the network structure and hyperparameters. For example, increase or decrease the number of hidden layers, adjust the kernel size of convolutional layers, etc., so that the model can adapt to new data distributions and changes in material properties, and maintain good performance.

[0041] This invention also includes: S9. Microstructure-Performance Correlation Prediction: Microstructure images of wood-plastic composites are acquired using Atomic Probe Tomography (APT) and Transmission Electron Microscopy (TEM). Combining materials science theory, a quantitative correlation model between microstructure characteristics and static bending strength is established. Deep learning methods are used to analyze the microstructure images and extract characteristic parameters such as the bonding strength between wood fibers and plastic interfaces and the uniformity of phase distribution. The formulas are as follows: , in, As an interface-based strength index, x i, y i is the characteristic vector of the material on both sides of the interface, and n is the number of sampling points on the interface. Through the formula, the interface bonding strength can be quantified, and more accurate micro-basis is provided for the prediction of the static bending strength. The microstructure characteristic parameters are taken as additional inputs and integrated into the neural network model. In actual operation, the microstructure characteristic data obtained by APT and TEM is preprocessed and input into the neural network model together with other macroscopic characteristic data. For example, the microstructure characteristics such as interface bonding strength and phase distribution uniformity are combined with macroscopic characteristics such as wood fiber content and plastic type, so that the model can comprehensively consider the material properties from the micro and macro levels, and improve the accuracy of the static bending strength prediction.

[0042] In the present application, in the data collection and cleaning step S1, the metadata such as data source, collection time and collection personnel are recorded by using the blockchain technology, so as to ensure the traceability and credibility of the data. The metadata is stored on multiple nodes by using the distributed ledger feature of the blockchain, and the data is encrypted by using the hash function H(x), and the formula is as follows: , to prevent data tampering. The data of different sources are quality rated, and the weight of the training data is adjusted according to the data quality, and the formula is as follows: , wherein w i is the weight of the i th data, q i is the data quality score (range 0-1), and e is the total number of data, so as to improve the prediction accuracy and robustness of the model.

[0043] In the present application, in the novel neural network architecture construction step S4, the capsule network (Capsule Network) structure is introduced. The capsule network routes the output of the low-level capsules to the high-level capsules through a dynamic routing mechanism, and the formula is as follows: , wherein c ij is the routing coefficient, b ij is the initial coupling coefficient, and K is the number of high-level capsules. The capsule network combines the capsule network with the Transformer-CNN fusion model, improves the expression ability of the model to the complex characteristics of the wood-plastic composite material, and enhances the generalization performance of the model.

[0044] In the present application, in the reinforcement training and dynamic optimization step S5, the transfer learning technology is adopted. The model is pre-trained in the related material field (such as fiber reinforced composite material), and then the parameters of the pre-trained model are migrated to the wood-plastic composite material static bending strength prediction model. According to the data characteristics of the wood-plastic composite material, the migrated parameters are fine-tuned, and the formula is as follows: , where, is the fine-tuned parameter, is the pre-trained model parameter, is the fine-tuning learning rate, is the loss function gradient of wood-plastic composite material data. Transfer learning is used to accelerate model training and improve the prediction performance of the model in small sample cases.

[0045] In the precise evaluation and intelligent feedback optimization step S6, in addition to the traditional evaluation indicators, model complexity indicators (such as network parameter quantity, computational complexity, etc.) and model interpretability indicators (such as feature importance ranking, interpretability score, etc.) are introduced. The feature importance calculation method based on Shapley value is adopted, and the formula is as follows: , where, is the Shapley value of feature i, S is the feature subset, N is the set of all features, and v(S) is the value function of feature subset S. Comprehensive consideration of these indicators can comprehensively evaluate the model performance and provide more abundant basis for model selection and optimization.

[0046] In the deep feature mining and enhancement step S3, molecular dynamics simulation technology is used to obtain the structure and interaction information of wood-plastic composite materials at the molecular level, such as intermolecular forces, chain segment motion, etc. The molecular dynamics simulation results are converted into feature vectors, and the formula is as follows: , where, i is the different characteristic value (such as bond length, bond angle, dihedral angle, etc.) obtained by molecular dynamics simulation. These feature vectors are fused with traditional features to further enrich the model input and improve the accuracy of wood-plastic composite material static bending strength prediction.

[0047] In the precise evaluation and intelligent feedback optimization step S6, a combination of swarm intelligence-based optimization algorithm (such as particle swarm optimization algorithm PSO or ant colony optimization algorithm ACO) and Bayesian optimization algorithm is adopted. In the particle swarm optimization algorithm, the particle velocity update formula is as follows: , where, is the velocity of particle i at the t+1 iteration, w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, is the historical optimal position of particle i, is the global optimal position, Xi,t is the position of particle i at the tth iteration. The global search is carried out through the particle swarm optimization algorithm, the better hyperparameter region is quickly located, then the Bayesian optimization algorithm is used for fine search in the region, the optimization efficiency and search accuracy are improved, and the optimal model parameters are found faster.

[0048] In the present application, in the multimodal data fusion preprocessing step S2, for the missing data, a deep neural network is constructed to predict the missing values, and the formula is as follows: , Xi,t is the position of particle i at the tth iteration. The global search is carried out through the particle swarm optimization algorithm, the better hyperparameter region is quickly located, then the Bayesian optimization algorithm is used for fine search in the region, the optimization efficiency and search accuracy are improved, and the optimal model parameters are found faster. is the 2k+1 data points around the missing value, and f is the neural network mapping function; for the noise problem in the data, a denoising autoencoder (DAE) based on deep learning is used for denoising processing. The denoising autoencoder adds noise to the input data, then learns to reconstruct the original data from the noisy data, and the formula is as follows: , Wherein, L is the reconstruction loss, x is the original data, is the added noise, f is the encoder function, and g is the decoder function. The denoising autoencoder removes the noise interference in the data, improves the data quality, and provides more reliable data for subsequent model training.

[0049] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting the static bending strength of wood-plastic composites based on neural networks, characterized in that, Includes the following steps: S1. Data collection and cleaning; collect sample data of wood-plastic composite materials, the sample data sources include laboratory test data, factory production data and public datasets; use data cleaning algorithms to perform preliminary screening of the sample data to remove erroneous data, including data that is outside the set range; S2. Preprocess the sample data; Fusing laboratory data using sensor calibration data fusion algorithms; Factory production data is integrated based on the logical relationships of production processes; publicly available data is processed using an adaptive outlier detection algorithm. Obtain the preprocessed data; S3. Feature extraction: Virtual features of wood-plastic composite materials are extracted from the preprocessed data using generative adversarial networks to obtain the extracted features. S4. Construct a neural network model based on the extracted features and the fusion of Transformer-CNN. S5. Using the preprocessed data, an adaptive learning rate adjustment strategy is adopted, and the neural network model is trained by combining the learning rate warm-up and cosine annealing mechanism to obtain the trained model. S6. The trained model is evaluated using a combination of leave-one-out method and Monte Carlo cross-validation to obtain the evaluation results. S7. Input the data of the wood-plastic composite material to be predicted into the optimized model, use knowledge graph to assist in the prediction, and output the prediction results.

2. The method for predicting the static bending strength of wood-plastic composite materials based on neural networks according to claim 1, characterized in that, Also includes: S8. Dynamic Data Update and Model Evolution: Establish a real-time data monitoring system to collect new wood-plastic composite material data in real time. When the amount of new data reaches a certain threshold, an incremental learning algorithm is used to update the model, as shown in the following formula: ， in, For the updated model parameters, These are the parameters of the original model. For learning rate, For new data The loss function gradient is used to periodically evaluate model performance. When the model's accuracy on the validation set drops by more than 5% in three consecutive evaluations, the model evolution process is initiated to readjust the network structure and hyperparameters.

3. The method for predicting the static bending strength of wood-plastic composite materials based on neural networks according to claim 1, characterized in that, Also includes: S9. Microstructure-Performance Correlation Prediction: Microstructure images of wood-plastic composites were acquired using atomic probe tomography and transmission electron microscopy. A quantitative correlation model between microstructure characteristics and static bending strength was established. Deep learning methods were used to analyze the microstructure images. The formula is as follows: ， in, As an interface-based strength index, x i y i Let n be the feature vectors of the materials on both sides of the interface, and n be the number of sampling points on the interface. The microstructure feature parameters are used as additional inputs and incorporated into the neural network model.

4. The method for predicting the static bending strength of wood-plastic composite materials based on neural networks according to claim 1, characterized in that, In S1, blockchain technology is used to record metadata. Leveraging the distributed ledger characteristics of blockchain, the metadata is stored on nodes, and the data is encrypted using a hash function H(x), as shown in the following formula: ， The quality of data from different sources is rated, and the weights of the training data are adjusted according to the data quality, using the following formula: ， Among them, w i Let q be the weight of the i-th data point. i The data quality score is given, where e represents the total number of data points.

5. The method for predicting the static bending strength of wood-plastic composite materials based on neural networks according to claim 1, characterized in that, In S4, the CNN module uses dilated convolution, as shown in the following formula: ， Where y[i] is the convolution output, For the input data, w[j] is the kernel weight, r is the dilatation rate, and h is the kernel size; Introducing the capsule network structure, the capsule network utilizes a dynamic routing mechanism, as shown in the following formula: ， Among them, c ij b is the routing coefficient. ij is the initial coupling coefficient, and K is the number of high-level capsules.

6. The method for predicting the static bending strength of wood-plastic composite materials based on neural networks according to claim 1, characterized in that, In S5, adversarial training techniques are used to incorporate adversarial examples during the training process, as shown in the following formula: ， in, As an adversarial example, x is the original sample. For the disturbance intensity, J( (x, y) is the loss function. For model parameters, The gradient of the loss function with respect to sample x; The model is trained using model ensemble techniques, and a weighted fusion method is employed, as shown in the following formula: ， in, This is the final predicted value. For predicted values, As weight, and ; Using transfer learning technology, after pre-training the model, the parameters of the pre-trained model are transferred to the static bending strength prediction model of wood-plastic composites. Based on the characteristics of the material data, the transfer parameters are fine-tuned, as shown in the following formula: ， in, These are the parameters after fine-tuning. For pre-trained model parameters, To fine-tune the learning rate, The gradient of the loss function for wood-plastic composite data.

7. The method for predicting the static bending strength of wood-plastic composite materials based on neural networks according to claim 1, characterized in that, In S6, model complexity and model interpretability metrics are introduced, and a feature importance calculation method based on Shapley values ​​is adopted, as shown in the following formula: ， in, Let be the Shapley value of feature i, S be the feature subset, N be the set of all features, and v(S) be the value function of feature subset S.

8. The method for predicting the static bending strength of wood-plastic composite materials based on neural networks according to claim 1, characterized in that, In S3, multi-scale analysis of wood fiber and plastic characteristics is performed using wavelet packet transform, as shown in the following formula: ， ， Among them, h n and g n Here, represents the wavelet packet filter coefficients, j represents the decomposition scale, and k represents the translation parameter; Molecular dynamics simulations are used to obtain structural and interaction information at the molecular level in wood-plastic composites. The results of the molecular dynamics simulations are then transformed into eigenvectors, as shown in the following formula: ， Among them, f i For different eigenvalues ​​obtained from molecular dynamics simulations, the eigenvectors are fused with traditional features.

9. The method for predicting the static bending strength of wood-plastic composite materials based on neural networks according to claim 1, characterized in that, In S6, Monte Carlo cross-validation involves multiple random partitioning of the training and test sets, calculating the average evaluation metric, and using a Bayesian optimization algorithm to adjust the model hyperparameters, as shown in the following formula: ， in, For parameters under data D The posterior probability, Let P( be the likelihood function) ) is a parameter The prior probability of data D is given by P(D), and the marginal probability of data D is given by P(D). The particle swarm optimization algorithm combines swarm intelligence-based optimization with Bayesian optimization. The particle velocity update formula in the particle swarm optimization algorithm is as follows: ， in, Let be the velocity of particle i in the (t+1)th iteration, w be the inertia weight, c1 and c2 be the learning factors, and r1 and r2 be random numbers. This represents the historical best position of particle i. The optimal position globally. Given the iteration position of particle i, a global search is performed using the particle swarm optimization algorithm to locate the hyperparameter region, and then a Bayesian optimization algorithm is used to search within that region.

10. The method for predicting the static bending strength of wood-plastic composite materials based on neural networks according to claim 1, characterized in that, In S2, to address the noise issue in the data, a deep learning-based denoising autoencoder is used for noise reduction. The denoising autoencoder adds noise to the input data and then learns to reconstruct the original data from the noisy data, as shown in the following formula: ， Where L is the reconstruction loss and x is the original data. The noise is added, f is the encoder function, and g is the decoder function. The noise interference in the data is removed by using a denoising autoencoder.

Images