Method for predicting mechanical properties of particle board based on simulation and deep learning

Abstract

The application discloses a method for predicting and analyzing the mechanical properties of a shaving board based on simulation and deep learning, and utilizes finite element simulation and a theoretical construction unit thickness shaving board anisotropic small piece layer splicing structure equivalent single-layer board mechanical model to obtain a shaving board geometric appearance and stress-strain distribution dataset; further, a shaving board vertical direction stacking structure equivalent laminate mechanical model is constructed to obtain a shaving board structure parameter and mechanical property dataset. The dataset is used to train an artificial intelligence model, and based on the real-time paving structure of the shaving piece obtained in the shaving board processing process, the trained artificial intelligence model is used to realize the rapid intelligent prediction and quality control of the mechanical properties of the shaving board. The method has the advantages of simple operation, strong universality, high detection efficiency, high processing precision, low technical cost, strong quality control, etc., and has important significance for the analysis of the structure-activity relationship of the shaving board small piece layer paving structure, the prediction of the mechanical properties, and the production online quality control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of flakeboard mechanical property prediction analysis, and particularly relates to a flakeboard mechanical property prediction analysis method based on simulation and deep learning. BACKGROUND

[0002] The flakeboard is a kind of board prepared by flake pieces of wood, bamboo and the like. Because it has good mechanical properties, firmness and durability, green environmental protection, low cost and the like, it has a wide application prospect in the construction, furniture, packaging and transportation industries. Clarifying the mechanism of the mechanical properties of the flakeboard is an important basic problem for improving the performance of the board and performing nondestructive testing of the board. However, the random structure of the board formed by a large number of anisotropic small piece layers spliced and stacked is complex, and the structure-activity relationship between the structure and the mechanical properties is not clear.

[0003] The current preparation optimization of anisotropic small piece layer paving structure such as flake pieces faces many challenges. Due to the large discrete nature of the performance and geometric parameters of the flake pieces, the complex splicing and stacking structure of the flake pieces, and the discontinuous load transfer path between the flake pieces, the structure-activity relationship between the complex structure of the flakeboard and the mechanical properties is not clear, the design and optimization of the flake piece paving and gluing process lack theoretical guidance, and it is difficult to exert the inherent mechanical properties of the flake pieces. For example, the random assembly structure of a large number of small piece layers leads to stress concentration in the horizontal direction of the discontinuous anisotropic flake piece splicing structure, and complex mechanical behaviors such as interlayer stress difference and interlayer failure in the vertical direction of the stacking structure. How to study the structure-activity relationship between the anisotropic small piece layer paving structure (including the anisotropic piece layer splicing structure in the horizontal direction and the anisotropic piece layer stacking structure in the vertical direction) and the macro mechanical properties of the board is still an important problem with challenges.

[0004] The current methods to regulate the performance of particleboard include regulating the mechanical properties of particleboard by surface densification (Patent No. CN202110669498.0), gradient paving structure (Patent No. CN202122101122.8), and optimizing the selection range of particle flakes (Patent No. CN202211559697.7); and improving the moisture resistance, moisture resistance and antibacterial properties of particleboard by surface chemical modification (Patent No. CN202110408694.2). The current methods for analyzing the mechanical properties of particleboard include reliability and importance analysis of composite laminated plates (Patent Nos. CN202011144951.8 and CN202011144952.2), finite element simulation of variable stiffness composite laminated plates (Patent No. CN202010662044.6), micro finite element modeling method of fiber reinforced composites based on CT scan images (Patent No. CN202011201949.X), testing the compression performance of thermoplastic composites by a mechanical testing machine (Patent No. CN202210851152.7), composite material mechanical property automatic test system and method (Patent No. CN202211214484.0), and unidirectional composite material tensile mechanical property experimental device and method (Patent No. CN201910073695.9). The above research progress shows that the experimental techniques for the preparation and performance regulation of wood-based panels have developed well, but the methods for testing, simulating and analyzing the mechanical properties of wood-based panels are mainly based on the laminated plate model. Theoretical analysis, numerical simulation and machine learning research on the structure characteristics of particleboard, which is composed of anisotropic small particle flakes, are still lacking, resulting in unclear structure-property relationship between the structure and mechanical properties of particleboard, and lack of theoretical guidance for particleboard preparation.

[0005] On the other hand, the randomness of the flake chip laying structure on the production line is large, and the stability of the flake board mechanical property control is insufficient. Real-time performance prediction and quality control of flake board mechanical properties during production are of great significance. The current detection and analysis methods of flake board mechanical properties include surface defect detection based on neural network and color camera (Patent No. CN202310238055.5), image defect classification detection (Patent No. CN202110133237.7), adaptive board thickness surface defect detection (Patent No. CN202111680878.0), visual Transformer surface defect detection (Patent No. CN202111680499.1), non-destructive testing of bamboo-wood composite container floor based on machine learning algorithm using image processing (Patent No. CN202010699496.1), flake board surface defect detection based on deep learning algorithm (paper: https: / / doi.org / 10.3390 / s22207733), flake board mechanical property prediction based on principal component regression-random forest (paper: https: / / www.doi.org / 10.15376 / biores.16.2.2448-2471). Among them, one type of method for detecting surface and internal structural defects of flake board has the advantages of obtaining flake board structure information, fast detection speed, and continuous automatic detection, but this type of method cannot obtain stress distribution information based on structure. Through mechanical experiments, the mechanical properties of the board can be obtained, but this method usually needs to test the board after processing, and requires a lot of time and economic cost, and it is difficult to realize real-time online rapid detection and analysis of the mechanical properties of the board. The above research progress shows that the current technology is still difficult to realize "obtaining stress distribution based on geometric image, obtaining mechanical properties based on structure parameters" of flake board mechanical property analysis and prediction with fast, high precision and low cost.

[0006] With the continuous progress of science and technology, there are many methods for predicting the mechanical properties of materials through artificial intelligence and other means. Such as the mechanical property prediction method of carbon fiber reinforced composite materials based on cross-scale simulation (Patent No. CN201811116128.9), the material performance prediction method based on deep learning (Patent No. CN202010415084.0), etc. The mechanical property prediction method of carbon fiber reinforced composite materials based on cross-scale simulation realizes variable parameter modeling, reduces the difficulty of grid division, and improves the calculation efficiency, but it is not suitable for the mechanical properties of artificial boards with large discrete anisotropic small piece layer assembly and stacking structure. The material performance prediction method based on deep learning simplifies the operation and improves the operation speed, but the structure is not simplified, and the RVE volume of the whole mechanical properties of the composite material is large, which makes it difficult to capture the key mechanism.

[0007] In summary, it is urgent to develop an intelligent prediction and analysis method for the mechanical properties of particle board, which can simplify the complex structure of particle board, comprehensively analyze the influence of horizontal splicing structure and vertical stacking structure, has fast detection speed, high processing precision and low technical cost. SUMMARY

[0008] Therefore, the purpose of the present application is to provide a particle board mechanical property prediction and analysis method based on simulation and deep learning, which has a simplified structure, strong universality, fast detection speed and strong quality control.

[0009] In order to achieve the above technical purpose, the technical scheme adopted by the present application is:

[0010] A particle board mechanical property prediction and analysis method based on simulation and deep learning, comprising:

[0011] S01, according to the particle sheet paving structure data in the particle board preparation process, a finite element model of unit thickness particle board structure is constructed;

[0012] S02, based on the finite element model, an equivalent single-layer board model of the particle board is constructed according to the finite element simulation method and the equivalent modulus theory, and then an equivalent laminated board model is established according to the classical laminated plate theory;

[0013] S03, based on the constructed finite element model, equivalent single-layer board model and / or equivalent laminated board model, a data set of particle board structure and performance is constructed, and then the obtained data set is used to train an artificial intelligence model, which is trained to convergence;

[0014] S04, collect the parameters in the particle board paving process, input them into the trained artificial intelligence model, and predict and analyze the mechanical properties of the particle board by the artificial intelligence model, and output the prediction and analysis results.

[0015] As a possible implementation, further, in the present scheme S03, the artificial intelligence model includes a deep learning model and a machine learning model; both the deep learning model and the machine learning model are trained using the data set to meet the preset requirements;

[0016] The deep learning model is used to predict one or more of the stress distribution of the splicing structure during particle board paving, the equivalent single-layer board equivalent paving angle, and the equivalent material engineering constant; the machine learning model is used to predict the mechanical properties of the particle board.

[0017] As a relatively preferred implementation option, preferably, the present scheme S04 comprises:

[0018] S041. Collect geometric morphology data of particleboard per unit thickness during particleboard installation. Using this data as input parameters, use a trained deep learning model to predict the stress distribution of the particleboard splicing structure, as well as one or more of the equivalent ply angle and equivalent material engineering constant of the equivalent single-layer board, to obtain the equivalent single-layer board data.

[0019] S042. Based on the equivalent single-layer board data, the equivalent single-layer boards are vertically stacked to form an equivalent laminate for particleboard, so as to deduce the structural parameters of the equivalent laminate. Then, these parameters are used as input parameters, and a trained machine learning model is used to predict the mechanical properties of particleboard.

[0020] As a preferred implementation option, preferably, in this solution S01, the chip shavings paving structure data is an optical image of the chip shavings paving structure, which is acquired by an industrial camera instrument to provide structural information such as chip length, width, distribution density, and fiber angle.

[0021] As a preferred implementation option, in this scheme S01, the finite element model of the unit thickness particleboard structure is based on the optical diagram of the particleboard fabrication process. The finite element model depicts one or more of the following structural features: particleboard length, width, distribution density, and fiber angle. Furthermore, the anisotropic material engineering constant and layup angle of the particleboard are set according to the fiber angle of the particleboard in each region. A general static analysis is performed on the established unit thickness splicing structure model to calculate the stress-strain distribution and / or constraint reaction force of the unit thickness particleboard structure under tensile and shear displacement conditions.

[0022] As a preferred implementation option, in this scheme S02, the equivalent single-layer board model of the particleboard is the simulation result output by the finite element model of the particleboard structure with unit thickness. By solving the principal stress direction, equivalent stress, equivalent strain, and equivalent anisotropic material parameters of the two-dimensional shell element model, the particleboard splicing structure with unit thickness is simplified to an equivalent single-layer board with continuous anisotropy and consistent ply angle, thereby obtaining the equivalent material engineering constant and equivalent ply angle of the equivalent single-layer board.

[0023] As a preferred implementation option, in this scheme S02, the equivalent laminate model is established based on the classical laminate theory by stacking equivalent single-layer boards of different heights and unit thicknesses of particleboard horizontally spliced ​​in the vertical direction of the particleboard to form the equivalent laminate model. Then, based on the equivalent material engineering constant and equivalent ply angle of each equivalent single-layer board, the equivalent laminate of the particleboard is calculated. The mechanical properties of the particleboard are obtained by simulation based on the classical laminate theory and the equivalent laminate model.

[0024] As a preferred implementation option, preferably, in the S03 scheme, the data set of the structure and performance of the particle board includes the following two types:

[0025] (1) The data set takes the unit thickness particle sheet splicing geometry as the input quantity and the stress-strain distribution as the output quantity;

[0026] (2) The data set takes the particle board layer structure parameters as the input quantity and the mechanical properties of the particle board as the output quantity.

[0027] As a preferred implementation option, preferably, the deep learning model in the present scheme uses the aforementioned experimental, simulation and theoretical methods to construct the data set of the stress distribution under the unit displacement action of the unit thickness particle sheet splicing structure geometry; it takes the unit thickness particle sheet splicing geometry as the input quantity and the stress distribution as the output quantity, and establishes the correlation between the geometric image and the stress distribution through the training of the hidden layer and the nonlinear activation function of the deep learning model. That is, the present scheme takes the current unit thickness particle sheet laying geometry on the bamboo particle board production line collected by the industrial camera as the input parameter, and uses the trained deep learning model to quickly predict the stress distribution of the splicing structure.

[0028] As a preferred implementation option, preferably, the machine learning model in the present scheme uses the aforementioned experimental, simulation and theoretical methods to construct the particle board structure parameter and macro mechanical property data set, which takes the structure parameter as the independent variable and the macro mechanical property as the target variable, trains the weight and bias of the machine learning model, and establishes the correlation between the structure parameter and the mechanical property. The machine learning model records the equivalent layer angle and equivalent material engineering constant of the "equivalent single-layer board" at different vertical positions in turn according to the laying sequence of the particle sheet on the bamboo particle board production line, and stacks the "equivalent single-layer board" vertically into a "equivalent laminate" of the particle board. The current "equivalent laminate" structure parameters (number of layers, layer thickness, layer angle ratio, layer sequence, etc.) are used as input parameters, and the trained machine learning model is used to quickly predict the macro mechanical properties of the board.

[0029] In the present scheme, the equivalent laminate model simplifies the complex structure of the particle board and embodies the characteristics of the random splicing structure of the particle sheet in the horizontal direction and the layer-by-layer stacking structure of the particle board in the vertical direction.

[0030] Preferably, the machine learning model is one or more of support vector machines, decision trees, random forests, and Bayesian learners.

[0031] As a preferred implementation option, preferably, the deep learning model in the present scheme is a fully convolutional neural network deep learning model; and the machine learning model is an integrated machine learning model based on decision trees.

[0032] Compared with the prior art, the present application has the beneficial effects that: the present application analyzes and predicts the mechanical properties of the shaving board by combining the finite element method with artificial intelligence. According to the optical image of the shaving piece paving structure in the shaving board preparation process, a unit thickness shaving structure model is constructed, finite element analysis is performed, and finally an "equivalent single-layer board" model suitable for analyzing the performance influence of horizontal splicing of the shaving piece splicing structure and an "equivalent laminated board" model suitable for analyzing the performance influence of vertical stacking of the stacking structure are constructed, thereby providing a theoretical basis for the "structure-performance integrated design" of such a type of artificial board; based on the above method, a data set of the shaving board structure and performance, the shaving piece splicing appearance and stress distribution is established, a machine learning and deep learning model is constructed and trained, and the stress distribution of the unit thickness splicing structure is predicted based on the paving structure of the shaving piece through the deep learning model. Thus, the equivalent unit thickness "equivalent single-layer board" model is stacked into the "equivalent laminated board" model, the "equivalent laminated board" model is quickly predicted by the machine learning model to obtain the mechanical properties of the shaving board, and finally the shaving board quality control information is output, thereby providing a method for online intelligent detection of product quality. In addition, the method has the advantages of simple structure, strong universality, fast detection speed, and strong quality control. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief descriptions will be given to the drawings needed to be used in the embodiments or prior art descriptions. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort.

[0034] Figure 1 is a simple flowchart in one embodiment of the present application; in the figure, (a) is the stress distribution predicted based on the bamboo shaving piece paving geometry by using the deep learning model, and (b) is the mechanical properties predicted based on the bamboo shaving board equivalent laminated board structure parameters by using the machine learning model;

[0035] Figure 2 is a technical roadmap in one embodiment of the present application. The structure-property relationship and intelligent prediction of the mechanical properties of the bamboo shaving board are studied by the methods of finite element numerical simulation, theoretical model construction, and artificial intelligence prediction.

[0036] Figure 3 is a main process schematic diagram of the preparation of the bamboo shaving board in one embodiment of the present application; in the figure, (a) is the drying of the bamboo shaving piece, (b) is the gluing, (c) is the forming, (d) is the paving, (e) is the hot pressing, and (f) is the post-treatment;

[0037] Figure 4Figure 1 is a schematic diagram of the equivalent single-layer board model of the horizontal splicing structure of bamboo shaving pieces in an embodiment of the present application; in the figure, (a) the splicing structure of bamboo shaving pieces is modeled, (b) the fiber orientation of the bamboo shaving pieces and the anisotropic material parameters are set, (c) the stress distribution of the structure under uniaxial tensile displacement and shear displacement, (d) the equivalent stress analysis of the structure by the equivalent modulus theory, and (e) the equivalent single-layer board of the unit thickness bamboo shaving piece splicing structure;

[0038] Figure 5 Figure 2 is a schematic diagram of the equivalent laminated board model of the vertical stacking structure of bamboo shaving pieces in an embodiment of the present application; in the figure, (a) the optical picture of the bamboo shaving piece paving structure, (b) the intelligent identification method of the bamboo shaving pieces, (c) the model of the bamboo shaving piece splicing structure constructed according to the geometric parameter law of the bamboo shaving pieces, (d) the model of the bamboo shaving piece splicing structure and the grid division, (e) the stress and strain distribution on the anisotropic bamboo shaving piece splicing structure, (f) the vertical direction stacking structure of the bamboo shaving pieces and the corresponding (g) equivalent single-layer board stacking structure, and (h) the equivalent laminated board structure of the bamboo shaving board;

[0039] Figure 6 Figure 3 is a schematic diagram of the model analysis of the equivalent laminated board of the vertical stacking structure of bamboo shaving pieces in an embodiment of the present application; in the figure, (a) the equivalent laminated board of the bamboo shaving board is constructed by stacking the equivalent single-layer boards, (b) the analysis of the mechanical properties of the equivalent laminated board of the bamboo shaving board by the classical laminated board theory, (c) the finite element layer setting of the equivalent laminated board of the bamboo shaving board, and (d) the mechanical simulation;

[0040] Figure 7 Figure 4 is a structural diagram of the convolutional neural network model composed of an encoder and a decoder in the embodiment of the present application;

[0041] Figure 8 Figure 5 is a diagram showing the changes of the model indicators during the model training in the embodiment of the present application; in the figure, (a) is the mean absolute error MAE, (b) is the mean square error MSE, and (c) is the correlation coefficient R 2 . DETAILED DESCRIPTION

[0042] The present application will be further described below in conjunction with the drawings and embodiments. It is particularly pointed out that the following embodiments are only used to illustrate the present application, but do not limit the scope of the present application. Similarly, the following embodiments are only part of the embodiments of the present application, not all the embodiments, and all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0043] In conjunction with Figures 1 to 6 one of the embodiments of the present application, a method for predicting and analyzing the mechanical properties of shaving board based on simulation and deep learning includes:

[0044] S01, constructing a finite element model of a unit thickness particle board structure according to the particle sheet laying structure data in the particle board preparation process;

[0045] S02, constructing an equivalent single-layer board model of the particle board based on the finite element model, the finite element simulation method and the equivalent modulus theory, and then establishing an equivalent laminated board model according to the classical laminated board theory;

[0046] S03, constructing a data set of the particle board structure and performance based on the constructed finite element model, equivalent single-layer board model and / or equivalent laminated board model, and then training an artificial intelligence model using the obtained data set until it converges;

[0047] S04, collecting parameters in the particle board laying process and inputting them into the trained artificial intelligence model to predict and analyze the mechanical properties of the particle board by the artificial intelligence model and output the prediction and analysis results.

[0048] In the S03 of the embodiment, the artificial intelligence model includes a deep learning model and a machine learning model; both the deep learning model and the machine learning model are trained using the data set to meet the preset requirements.

[0049] Correspondingly, the deep learning model is used to predict one or more of the stress distribution of the splicing structure during the particle board laying, the equivalent laying angle of the equivalent single-layer board, and the equivalent material engineering constant; and the machine learning model is used to predict the mechanical properties of the particle board.

[0050] As a preferred implementation option, the S04 of the present scheme preferably comprises:

[0051] S041, collecting the geometric appearance data of the unit thickness particle sheet during the particle board laying process, using it as an input parameter, and using the trained deep learning model to predict one or more of the stress distribution of the particle board splicing structure, the equivalent laying angle of the equivalent single-layer board, and the equivalent material engineering constant to obtain the equivalent single-layer board data;

[0052] S042, based on the equivalent single-layer board data, stacking the equivalent single-layer board vertically to obtain the equivalent laminated board of the particle board, deriving the equivalent laminated board structure parameters, and then using the trained machine learning model to predict the mechanical properties of the particle board by using the equivalent laminated board structure parameters as input parameters.

[0053] Briefly, the method of the present scheme comprises experimental preparation and testing, finite element numerical simulation, theoretical model construction, and artificial intelligence prediction.

[0054] Among them, the combination of Figure 3As shown, the experimental preparation includes obtaining size parameter information (length, width, distribution density, fiber angle, etc.) of bamboo shaving pieces, exploring suitable gluing schemes and directional laying methods for bamboo shaving boards, constructing board blanks with different bamboo shaving piece laying angles, and exploring hot pressing conditions and post-processing conditions for the preparation of bamboo shaving boards. In this scheme, the mechanical properties of the bamboo shaving board can be quantified by bending stiffness and strength. Through the above work, the experimental results of the structure parameters and mechanical properties of the bamboo shaving board are obtained, which provide experimental data for subsequent theoretical model construction, finite element numerical simulation, machine learning and deep learning model training.

[0055] For finite element numerical simulation, in the present scheme S01, the finite element model of the unit thickness shaving board structure is drawn according to the optical diagram of the shaving piece laying structure in the shaving board preparation process. One or more structures of the shaving piece length, width, distribution density, and fiber angle are depicted by finite element modeling. The anisotropic material engineering constants and laying angles of the shaving pieces are set according to the fiber angle of the shaving pieces in each region. The general static analysis is performed on the established unit thickness bottom splicing structure model to calculate the stress and strain distribution and / or constraint reaction force of the unit thickness shaving board structure under tensile and shear displacement conditions.

[0056] As a possible implementation example, in combination with Figure 4 , Figure 5 As shown in the present scheme S01, the shaving piece laying structure data is an optical diagram of the shaving piece laying structure, which is obtained by an industrial camera instrument to provide statistical distribution of one or more structure parameters of the shaving piece length, width, distribution density, and fiber angle. According to the parameter statistical law, a two-dimensional shell element model of unit thickness anisotropic small piece layer splicing is constructed by using the ABAQUS pre-processing module based on MATLAB and Python secondary development, and the anisotropic material engineering constants and laying angles of the bamboo shaving pieces are set according to the fiber angle of the bamboo shaving pieces in each region. The stress and strain distribution and constraint reaction force of the two-dimensional model under uniaxial tension and shear displacement conditions are simulated by using finite element method.

[0057] For the construction of the theoretical model construction, as a preferred implementation option, preferably, in the present scheme S02, the equivalent single-layer board model of the shaving board is the simulation result output by the finite element model of the unit thickness shaving board structure. The principal stress direction, equivalent stress, equivalent strain, and equivalent anisotropic material parameters of the two-dimensional shell element model are solved, and then the unit thickness shaving piece splicing structure is equivalent and simplified to a continuous anisotropic equivalent single-layer board with consistent laying angles, to obtain the equivalent material engineering constants and equivalent laying angles of the equivalent single-layer board.

[0058] In addition, in the scheme S02, the equivalent laminate model is established according to the classical laminate theory, the equivalent single-layer plate of the unit-thickness flake veneer horizontally spliced structure with different heights is stacked in turn to form the equivalent laminate model of the flake veneer, and the equivalent material engineering constant and the equivalent layer angle of each equivalent single-layer plate are calculated to obtain the equivalent laminate of the flake veneer, and the mechanical properties of the flake veneer are simulated based on the classical laminate theory and the equivalent laminate model.

[0059] As a specific example, as shown in Figure 5 , Figure 6 , the scheme includes solving the principal stress direction, equivalent stress, equivalent strain, and equivalent anisotropic material parameters of the two-dimensional shell element model based on the finite element simulation results, and then simplifying the unit-thickness bamboo flake spliced structure into a continuous anisotropic "equivalent single-layer plate" with consistent layer angles, obtaining the equivalent material engineering constant and the equivalent layer angle of the "equivalent single-layer plate". According to the paving sequence in the preparation process of the bamboo flake veneer, the "equivalent single-layer plate" models of the bamboo flake spliced structure at different paving times are stacked in turn to form the "equivalent laminate" model of the bamboo flake veneer. The ABAQUS software is used to construct the "equivalent laminate" model of the bamboo flake veneer, and the equivalent material engineering constant and the equivalent layer angle of each layer are set in turn; the stress-strain distribution and constraint reaction of the bamboo flake veneer under uniaxial tension and three-point bending loading are simulated; the stiffness matrix of the "equivalent laminate" of the bamboo flake veneer is calculated using the classical laminate theory, and the mechanical properties of the bamboo flake veneer are tested by uniaxial tension and three-point bending experiments.

[0060] As a preferred implementation option, preferably, in the scheme S03, the data set of the structure and performance of the flake veneer includes the following two types:

[0061] (1) The data set with the unit-thickness flake veneer spliced geometry structure as the input and the stress-strain distribution as the output;

[0062] (2) The data set with the flake veneer layer structure parameters as the input and the mechanical properties of the flake veneer as the output.

[0063] In combination with Figure 1As shown, in the aspect of artificial intelligence prediction, the deep learning model described in the scheme utilizes the aforementioned experimental, simulation and theoretical methods to construct a dataset of stress distribution under the action of unit displacement and the geometric morphology of the unit thickness flake veneer splicing structure; it takes the geometric morphology of the unit thickness flake veneer splicing as the input quantity and the stress distribution as the output quantity, and establishes the correlation between the geometric image and the stress distribution through the training of the hidden layer and the nonlinear activation function of the deep learning model. That is, the scheme takes the current unit thickness flake veneer paving geometric morphology on the bamboo shaving board production line collected by the industrial camera as the input parameter, and uses the trained deep learning model to quickly predict the stress distribution of the splicing structure. Based on the stress distribution prediction result, the equivalent layer angle and equivalent material engineering constant of the "equivalent single-layer board" are determined.

[0064] As a preferred implementation option, preferably, the machine learning model described in the scheme utilizes the aforementioned experimental, simulation and theoretical methods to construct a flake board structure parameter and macro mechanical property dataset, which takes the structure parameter as the independent variable and the macro mechanical property as the target variable, trains the weight and bias of the machine learning model, and establishes the correlation between the structure parameter and the mechanical property. The machine learning model records the equivalent layer angle and equivalent material engineering constant of the "equivalent single-layer board" at different positions in the vertical direction in turn according to the flake veneer paving sequence on the bamboo shaving board production line, and stacks the "equivalent single-layer board" vertically into a flake board "equivalent laminate". The current "equivalent laminate" structure parameters (paving layer number, paving layer thickness, paving layer angle ratio, paving sequence, etc.) are used as input parameters, and the trained machine learning model is used to quickly predict the macro mechanical properties of the board. Based on the stress distribution prediction result, the equivalent layer angle and equivalent material engineering constant of the "equivalent single-layer board" are determined, the "equivalent single-layer board" is stacked into an "equivalent laminate" according to the paving sequence, and the laminate structure parameters are extracted; the machine learning model is applied to predict the mechanical properties of the bamboo shaving board according to the laminate structure parameters. The mechanical properties of the bamboo shaving board are predicted and quality control is realized quickly, accurately and at low cost.

[0065] In the scheme, the equivalent laminate model simplifies the complex structure of the flake board and embodies the characteristics of the horizontal random splicing structure of the flake veneer and the vertical stacking structure of the flake board.

[0066] Preferably, the machine learning model is one or more of support vector machines, decision trees, random forests, and Bayesian learners.

[0067] As a preferred implementation option, preferably, the deep learning model described in the scheme is a fully convolutional neural network deep learning model; and the machine learning model is an integrated machine learning model based on decision trees.

[0068] In order to further illustrate the embodiments of the scheme, an example is given as follows:

[0069] Obtain the size parameter information of bamboo shaving board. Including measuring hundreds of bamboo shaving pieces, obtaining the length and width range of 40-90mm, the width range of 10-28mm, measuring a certain number of shaving pieces in the shaving board, and the distribution density is 0.68-0.75g / cm 3 , the measured shaving pieces are glued, the glue amount is 12%, and each stirring direction is stirred for 15 minutes. The shaving pieces are randomly laid into the mold to form a board blank, and the board blank is placed for 3-5min, and then hot pressed by a hot press at 140℃ for 15min, and the hot pressing pressure is 20Mpa. Finally, the bamboo shaving board with random angle laying is made as shown in Figure 3 , and the tensile properties of the pressed bamboo shaving board are tested to obtain the material parameters in the finite element simulation.

[0070] Finite element model numerical simulation. The finite element model calculation solution is carried out in ABAQUS2020 2.0 software. According to the structure parameters of the bamboo shaving pieces measured during the preparation of the bamboo shaving board experiment, the PYTHON secondary development is used to establish a 200 150x150mm two-dimensional shell element random angle anisotropic small piece layer model as shown in Figure 5 c. According to the setting of the shaving piece area, the fiber angle of the shaving piece is set, and the material engineering constant is as shown in Figure 4 b. The material engineering constant and the fiber angle range are shown in Table 1. According to the consistent setting of the material engineering constant, the splicing angle of each shaving piece is randomly arranged around the x-axis, which can represent the discrete type of the material to the greatest extent, and the mechanical property difference problem. The grid division adopts four-node double-bending thin shell, reduces integration, hourglass control and finite membrane strain. The load and boundary condition is set as left side full fixed, right side 1mm displacement load(x-axis loading); The lower side is fully fixed, and the upper side is 1mm displacement load(y-axis loading). After the operation is completed, the Mises stress value of the node is exported from ABAQUS, and the finite element post-processing program based on PYTHON secondary development is used to extract the shaving board node coordinates from ABAQUS.

[0071] Theoretical analysis. By applying single direction displacement strain in ABAQUS, the macroscopic response and material performance of the material are calculated. Loading is carried out in 1, 2 normal direction and 12 tangent direction, and the stress response of the representative plane is calculated. The extracted node coordinates and stress values are processed by MATLAB interpolation to calculate the average stress value, and the macroscopic average stress of the representative volume element is obtained by volume averaging the stress, so as to solve the following material constitutive stiffness matrix component coefficients.

[0072]

[0073] After the constitutive stiffness matrix is solved, the constitutive stiffness matrix analysis coefficient is input to the ABAQUS material property module for equivalent stress operation, and then the equivalent single-layer plate material performance is obtained. The equivalent single-layer plates are input to the laminated plate model in layers to obtain the equivalent laminated plate model material performance.

[0074] Table 1 Flaking board engineering constants and flaking sheet random angle range

[0075]

[0076] A deep learning model is established. The two-dimensional image data derived from the ABAQUS software is used for model learning and training. The two-dimensional image data is sliced into 64x64 pixel size. Data from different plane models is used to train and evaluate the model. The processed two-dimensional data is divided into a training set and a validation set. The training set and the validation set respectively account for about 93% and 7% of the size of the entire data set. In this embodiment, the training data set size is 595, and the validation set data size is 45. The corresponding material two-dimensional image, x-direction and y-direction displacement are loaded with corresponding tensile load values and two-dimensional stress distribution images to reshape into a one-dimensional vector with a length of 8194 (64x64+2+64x64). Therefore, a model is represented as a two-dimensional matrix with a dimension of 64x8194.

[0077] The convolutional neural network model structure adopted is a structure composed of an encoder and a decoder as shown in Figure 7 The encoder is used to extract image features, and the decoder allows the model to finally output an image size restored to be consistent with the input image size, both being 64x64x1, and the three dimensions of the image being length, width and channel respectively. Each convolutional layer has a convolution kernel with a size of 3x3 and a step of 1x1, and the third dimension of the convolution kernel is the same as the channel number of the input image. The feature map output after passing through the encoder in the SCSNet model has a size of 8x8x128, and the shape of the feature map is reshaped into a one-dimensional vector with a length of 1x1x8192 in the reshaping layer E8 to meet the input requirements of the fully connected layer. After passing through the fully connected layers E9 and E10, the model learns to obtain the potential features of the anisotropic small piece layer microstructure. Before the result of the encoder is input to the decoder, different tensile load values in the x and y directions are connected to the vector output by the fully connected layer E10 in the E11 layer. The decoder is opposite to the process of the encoder, and through the methods of deconvolution and de-pooling, the low-resolution feature map generated by the encoder is mapped and restored to a high-resolution image. The prediction accuracy of the model is evaluated by the mean absolute error MAE, the mean square error MSE, the correlation coefficient R 2 and the like, and the formulas are as follows.

[0078] Mean absolute error:

[0079] Mean square error:

[0080] Correlation coefficient:

[0081] In the formula, n represents the number of data samples, i represents the i-th data sample, y i represents the true value of the i-th sample, represents the predicted value, represents the average value of the true values of all samples. The data in the validation set is input into the stress distribution prediction model according to the geometric morphology graph, and it can be seen from the above indexes that the MSE and MAE curves rapidly decrease within the first fifty training iterations (epochs), and then enter the convergence stage and gradually tend to be stable Figure 8 . After 100 training iterations (epochs), the loss function is basically constant. The correlation coefficient finally stabilizes at 0.92, and the stress prediction distribution can meet the prediction requirements.

[0082] The above only describes some embodiments of the present application, and does not limit the protection scope of the present application, and any equivalent device or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for predicting and analyzing the mechanical properties of chipboard based on simulation and deep learning, characterized by, The method comprises the steps of: S01, constructing a finite element model of a unit thickness particle board structure according to particle sheet paving structure data in the particle board preparation process; S02, based on the finite element model, constructing an equivalent single-layer board model of the particle board according to the finite element simulation method and the equivalent modulus theory, and then establishing an equivalent laminated board model according to the classical laminated plate theory; wherein the equivalent single-layer board model of the particle board is: based on the simulation results output by the finite element model of the unit thickness particle board structure, the principal stress direction, equivalent stress, equivalent strain, and equivalent anisotropic material parameters of the two-dimensional shell element model are solved, and then the unit thickness particle sheet splicing structure is equivalent and simplified into a continuous anisotropic equivalent single-layer board with consistent layer angle; the establishment of the equivalent laminated board model according to the classical laminated plate theory comprises: according to the paving sequence in the vertical direction of the particle board, the equivalent single-layer boards of the unit thickness particle sheet splicing structures with different heights are stacked in turn to form the equivalent laminated board model of the particle board; S03, based on the constructed finite element model, equivalent single-layer board model and equivalent laminated board model, a dataset of particle board structure and performance is constructed, and then the obtained dataset is used to train an artificial intelligence model until it meets the requirements; the artificial intelligence model comprises a deep learning model and a machine learning model; S04, collecting parameters in the particle board paving process and inputting them into the trained artificial intelligence model to predict and analyze the mechanical properties of the particle board by the artificial intelligence model and output the prediction analysis results; Wherein, S04 includes: S041, collecting unit thickness particle sheet splicing geometric appearance data in the particle board paving process, using it as an input parameter, and predicting the stress distribution of the particle board splicing structure by using the trained deep learning model; based on the stress distribution prediction result, the equivalent layer angle and equivalent material engineering constant of the equivalent single-layer board are determined to obtain the equivalent single-layer board data; S042, based on the equivalent single-layer board data, the equivalent single-layer board is vertically stacked into the equivalent laminated board of the particle board to deduce the equivalent laminated board structure parameters, which are used as input parameters to predict the mechanical properties of the particle board by using the trained machine learning model.

2. The method of claim 1, wherein the method is based on analog and deep learning. In S03, the deep learning model and the machine learning model are both trained using the dataset until they meet the preset requirements; Wherein, the deep learning model is used to predict the stress distribution of the splicing structure during the paving of the particle board; the machine learning model is used to predict the mechanical properties of the particle board.

3. The simulation and deep learning-based analysis method for predicting the mechanical properties of chipboard according to claim 1 or 2, characterized in that, In S01, the particle sheet paving structure data is an optical image of the particle sheet paving structure, which is obtained by an industrial camera instrument to provide one or more structure information of the particle sheet length, width, distribution density, and fiber angle.

4. The method of claim 1, wherein the method is based on analog and deep learning. In S01, the finite element model of the unit thickness particle board structure is used to depict one or more structures of the particle sheet length, width, distribution density, and fiber angle, and the anisotropic material engineering constant and layer angle of the particle sheet are set according to the fiber angle of the particle sheet in each region to calculate the stress and strain distribution and / or constraint reaction force of the unit thickness particle board structure under the conditions of tension and shear displacement.

5. The method of claim 2, wherein the method is based on analog and deep learning. In S03, the dataset of particle board structure and performance includes the following two types: (1) a dataset taking the geometric morphology of unit-thickness shaving piece splicing as an input quantity and a stress distribution on a structure as an output quantity; (2) a dataset taking a shaving board layer structure parameter as an input quantity and a shaving board mechanical property as an output quantity.

6. The method of claim 2, wherein the method is based on analog and deep learning. The deep learning model takes the geometric morphology of unit-thickness shaving piece splicing as an input quantity and a stress distribution as an output quantity, and establishes the correlation between the geometric image and the stress distribution by training the hidden layer and the nonlinear activation function of the deep learning model.

7. The method of claim 2, wherein the method is based on analog and deep learning. The machine learning model takes the shaving board layer structure parameter as an independent variable and the shaving board mechanical property as a target variable, trains the weight and bias of the machine learning model, and establishes the correlation between the structure parameter and the mechanical property. The machine learning model is one or more of a support vector machine, a decision tree, a random forest, and a Bayesian learner.

8. The method of claim 2, wherein the method is characterized by, The deep learning model is a fully convolutional neural network deep learning model; and the machine learning model is an integrated machine learning model based on a decision tree.

Images