Prediction method of bursting pressure of carbon fiber fully wrapped gas cylinder based on machine vision and deep learning

A method using machine vision and deep learning to predict bursting pressure of carbon fiber fully wrapped gas cylinders by establishing a virtual test model, performing response analysis, constructing a neural network, and using grating projection to obtain radial expansion data, achieving accurate and efficient bursting pressure prediction.

CN118332851BActive Publication Date: 2024-12-06ZHEJIANG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410415657.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-08
Publication Date
2024-12-06
Estimated Expiration
2044-04-08

AI Technical Summary

Technical Problem

Current methods for predicting the bursting pressure of carbon fiber fully wrapped gas cylinders are inaccurate and inefficient, and existing inspection techniques fail to effectively detect internal defects, leading to safety risks and high costs due to the need for excessive safety factors.

Method used

A method using machine vision and deep learning to predict bursting pressure by establishing a virtual test model, performing response analysis, constructing a neural network, and using grating projection to obtain radial expansion data, thereby predicting bursting pressure with high accuracy and efficiency.

Benefits of technology

The method provides fast, efficient, and non-contact detection of gas cylinders, ensuring reliable bursting pressure prediction and reducing the gap between simulation and actual results, facilitating large-scale industrial use and performance evaluation of composite structures.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118332851B_ABST
    Figure CN118332851B_ABST
Patent Text Reader

Abstract

A method for predicting the bursting pressure of a fully wrapped carbon fiber gas cylinder based on machine vision and deep learning, comprising the following steps: Step 1, establishing a virtual test model, performing response analysis on the gas cylinder, outputting the radial expansion under the design pressure (simulated value), and then batch calculating through a finite element program to predict the bursting pressure of the gas cylinder; Step 2, constructing and training a deep learning neural network; Step 3, obtaining the radial expansion under the design pressure (real value) based on grating projection technology; Step 4, predicting the bursting pressure using the trained neural network. The method for predicting the bursting pressure based on machine vision and deep learning of the present invention has the characteristics of fast, efficient and non-contact detection, and provides a feasible method for the industrialization of detection of large composite materials structures; compared with the existing simulation results, the bursting pressure obtained by this method has a higher degree of consistency with the actual results, and this method can provide a direct and effective reference for the evaluation of related performance of composite gas cylinders.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for predicting bursting pressure of a carbon fiber fully-wrap gas cylinder. Background Art

[0002] Carbon fiber reinforced composite materials have the advantages of high specific strength and specific stiffness, corrosion resistance, good fatigue resistance, and designability of material mechanical properties. They are ideal materials for lightweight and high-strength structural design. In the field of new energy, carbon fiber reinforced composite materials are widely used in wind turbine blades, hydrogen-powered vehicles, hydrogen-powered aircraft, hydrogen transmission pipelines, etc. At present, carbon fiber fully wrapped composite hydrogen storage cylinders face different problems in their applications in various fields, such as: explosion failure, fatigue damage, lining buckling failure, impact damage, etc., and there are certain safety risks. In order to ensure the safe and stable operation of composite gas cylinders, it is very necessary to study their explosion mechanical properties.

[0003] Burst pressure is an important design parameter for carbon fiber fully wrapped gas cylinders. Although the predicted value of the burst pressure of the gas cylinder can be obtained through mechanical analysis of the gas cylinder and simulation of the failure steps, the predicted value obtained by this method is quite different from the actual value obtained from the gas cylinder burst test.

[0004] Carbon fiber fully wrapped gas cylinders are highly dependent on process methods and process steps due to their integrated manufacturing and molding characteristics. In the manufacturing process, it is inevitable or impossible to completely eliminate the internal microscopic defects, such as inclusions, pores, fiber wrinkles and warping, delamination, etc. These defects will seriously weaken the strength, stiffness and fatigue life of the composite material, and lead to a large dispersion of composite material product performance, so the measured burst pressures of different gas cylinders vary greatly.

[0005] Type testing is a necessary step in the manufacturing process of gas cylinders to ensure the integrity and sealing of the cylinder body. Gas cylinders generally need to be inspected regularly after leaving the factory, but due to the dispersed microscopic defects in the composite material layer of the carbon fiber fully wrapped gas cylinder, there is currently no effective regular inspection method. The only way to ensure the safety of gas cylinders in service is to increase the safety factor, resulting in high gas cylinder costs.

[0006] Commonly used strain gauge measurement technology can detect the surface strain under the internal pressure of the gas cylinder, but the strain gauge can only measure the strain at some points and cannot perform full-field measurement. In addition, the grinding and patching steps will damage the surface of the gas cylinder. Ultrasonic testing is more effective for metal materials, but the delamination and pore defects inside the composite material will significantly attenuate the ultrasonic signal. Industrial CT can detect internal defects in gas cylinders, but industrial CT equipment is expensive, detection is time-consuming, and there are radioactive hazards. It is not suitable as a detection tool that can be embedded in industrial process flows. In addition, although industrial CT can detect defect morphology, there is still a huge gap between the distribution of defect morphology and the quantitative evaluation of structural performance that needs to be bridged.

[0007] Therefore, how to conduct efficient and reliable testing and evaluation of gas cylinders in type testing and service steps is a difficult problem in the field of carbon fiber fully wrapped composite gas cylinders. Summary of the invention

[0008] In order to overcome the above problems, the present invention provides a method for predicting the bursting pressure of a carbon fiber fully wrapped gas cylinder based on machine vision and deep learning.

[0009] The technical solution adopted by the present invention is: a method for predicting the bursting pressure of a carbon fiber fully wrapped gas cylinder based on machine vision and deep learning, comprising the following steps:

[0010] Step 1: Establish a virtual test model, perform response analysis on the gas cylinder, output the radial expansion (simulated value) under the design pressure, and then perform batch calculations using a finite element program to predict the bursting pressure of the gas cylinder;

[0011] Specifically include:

[0012] (1.1) Aiming at the distribution pattern of weak interface bonding defects that mainly exist in carbon fiber fully wrapped gas cylinders, a large number of virtual test models with randomly distributed defects that meet statistical indicators are automatically generated. The weak interface bonding defects cover the reduction of interface bonding strength and the complete delamination of the interface. In the virtual test model, defects can be automatically and randomly assigned to each unit according to the statistical indicators of the severity of the defects, such as Figure 2 shown.

[0013] (1.1.1) Interface model. The weak bonding defect of the interface is usually described as the relaxation of the interface bonding. When interlayer slip occurs, although the interlayer lateral stress is still continuous under equilibrium conditions, the displacement between the layers is no longer continuous. When the interface is completely debonded, it becomes a delamination defect, which can slide relative to each other or change from closed to open when loaded. In the interface description, it is assumed that the interface traction force T n(t) for ||u n(t) ||, the interface condition is expressed as:

[0014]

[0015] In the formula, Γ represents the interface layer, ||*|| Γk,k+1 It represents the discontinuity of the physical quantity at the interface Γ between two adjacent layers k and k+1, which can be expressed as ||*|| Γk,k+1 =(*) (k) -(*) (k+1) ;n j is a unit vector orthogonal to the interface; K n and K t They are the interface normal stiffness and interface tangential stiffness, respectively, which can be considered as the characterization of the interface bonding strength, with the unit of GPa / mm.

[0016] (1.1.2) Micromechanical modeling. The equivalent stiffness coefficient of the laminated structure considering different bonding degrees is obtained as follows:

[0017]

[0018] In the formula,

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025] In the formula,

[0026]

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] In the formula, C ijkl (1) , C ijkl (2) and C ijkl (I) are the material stiffness coefficients of the fiber reinforced layers 1, 2 and the interface layer I respectively.

[0033] (1.1.3) Defect probability distribution model. When considering the random perturbation of defects in fiber reinforced composite structures, the defect structure characterized by parameters can be regarded as a mathematical random sequence. In the description of weak bonding defect distribution, the distribution of exponent x is used as a control variable to describe the severity of weak bonding defects. Assume that x is a random variable and obeys the expectation x u , the variance is S 2 The normal distribution model description is denoted as X(x)~N[x u ,S 2 ], then the corresponding K nObey the lognormal distribution ln(K n )~N(x u ,S 2 ), whose expression is:

[0034]

[0035] (1.1.4) Spatial random sampling. The statistical distribution law describing the severity of defects is given above, but the specification of the spatial random distribution of defects in the composite reinforcement layer of the gas cylinder also requires a clear description of its random distribution in the spatial discrete field.

[0036] Discretize each region of the structure into small units i (e i ∈Y, i=1,2,3,...), the discrete unit scale corresponds to the representative volume unit that can characterize the defect. Y is the sum of the units of the composite structure discretization. In the finite element model, the sampling space corresponds to the e of the discrete unit. i Assume that the total number is n, and its position information is associated with unit number i. Take a unit of size n from n d The random sampling of y represents the occurrence of defects in any unit. Then the distribution probability of defects p i It can be expressed as:

[0037]

[0038] Wherein, subscript d represents the defect unit, and subscript j represents the sub-division of the fiber reinforced composite material layer.

[0039] (1.2) Two different pressure levels are established according to the design pressure, and the two different pressure differences are equal to the design pressure. The response analysis of the gas cylinder with randomly distributed defects described in steps (1.1.1) to (1.1.4) is performed to obtain the radial expansion (simulated value) under the design pressure.

[0040] (1.3) Considering the dispersion and spatial randomness of the defect morphology, the finite element program is automatically executed for batch calculation to predict the bursting pressure of the gas cylinder containing randomly distributed defects described in steps (1.1.1) to (1.1.4).

[0041] Step 2: Build and train a deep learning neural network. The unique random defect distribution of each gas cylinder determines the radial expansion of different cross sections along the axial direction and the final bursting pressure during its service step. However, it is difficult to establish an explicit functional relationship between the three. Deep learning can establish the relationship between radial expansion and bursting pressure, including:

[0042] (2.1) Based on step 1, the radial expansion (simulated value) and burst pressure under a large number of design pressures are obtained The bursting pressure loss function is constructed as:

[0043]

[0044] Construct a neural network such as Figure 3 shown.

[0045] (2.2) Train the neural network to minimize the value of the loss function.

[0046] (2.2.1) Input parameters such as cylinder inner diameter, thickness, number of composite material layers, ply angle, and radial expansion (simulated value).

[0047] (2.2.2) Through the neural network, the relationship between the key information such as the inner diameter, thickness, number of composite material layers, laying angle and radial expansion (simulated value) of the gas cylinder and the bursting pressure was preliminarily established.

[0048] (2.2.3) Output the predicted bursting pressure.

[0049] (2.2.4) Then, the bursting pressure output by the neural network and the bursting pressure obtained by finite element calculation in step 1 are substituted into the loss function to calculate the error.

[0050] (2.2.5) Then, the bias and weight of each layer are continuously adjusted through the network's back propagation until the error is minimized and the iteration is completed.

[0051] Step 3, obtaining the radial expansion (true value) under the design pressure based on the grating projection technology, specifically includes:

[0052] (3.1) Based on the grating projection machine vision system, the surface depth information of the gas cylinder under different pressures is obtained in the form of a point cloud of millions of levels.

[0053] (3.2) Combined with the cylindrical fitting algorithm, the radial expansion (true value) of different cross sections of the gas cylinder under the above pressure difference is obtained, as follows: Figure 4 shown.

[0054] (3.2.1) Perform N fittings on the point cloud data under the internal pressure P0 state to obtain N sets of cylinder radii R i (i=1,2,3…N) and find the average value, then the radius of the cylinder in the P0 state is R0:

[0055]

[0056] Similarly, the radius R1 of the cylinder under the internal pressure P1 is:

[0057]

[0058] (3.2.2) Subtracting the radius values ​​in the two states (R0<R1) can obtain the radial expansion of the selected cross section. for:

[0059]

[0060] Step 4, using the trained neural network to predict the burst pressure, as follows:

[0061] (4.1) For a given type of gas cylinder, at any step of filling and depressurization, a machine vision system based on grating projection is used to collect and analyze images and output information such as 3D point cloud and radial expansion (true value).

[0062] (4.2) Input the inner diameter, thickness, number of composite material layers, ply angle and radial expansion (true value) of the gas cylinder, and obtain the bursting pressure prediction value through the trained neural network to achieve efficient and reliable online bursting pressure prediction and early warning of the gas cylinder, such as Figure 5 shown.

[0063] The beneficial effects of the present invention are as follows: compared with the currently available detection methods, the method for predicting bursting pressure based on machine vision and deep learning of the present invention has the characteristics of fast, efficient and non-contact detection, and provides a feasible method for the industrialization of detection of large-scale composite structures; compared with the existing simulation results, the bursting pressure obtained by this method has a higher degree of consistency with the actual results, and based on this method, a direct and effective reference basis can be provided for the relevant performance evaluation of composite gas cylinders. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The overall flow chart of the method for predicting bursting pressure of a carbon fiber fully wrapped gas cylinder based on machine vision and deep learning disclosed in the present invention;

[0065] Figure 2 It is a response analysis and output schematic diagram of the virtual test model disclosed in the present invention;

[0066] Figure 3 A schematic diagram of a neural network for training a bursting pressure loss function disclosed in the present invention;

[0067] Figure 4 It is a schematic diagram of measuring the radial expansion amount based on grating projection disclosed in the present invention;

[0068] Figure 5 This is a schematic diagram of predicting bursting pressure by combining machine vision with deep learning disclosed in the present invention. DETAILED DESCRIPTION

[0069] like Figure 1The overall flow chart of the method for predicting bursting pressure of a carbon fiber fully wrapped gas cylinder based on machine vision and deep learning provided by the present invention is shown;

[0070] Step 1: Establish a virtual test model with defect distribution, perform response analysis on the gas cylinder, output the radial expansion under the design pressure (simulated value), and then obtain the predicted value of the gas cylinder bursting pressure through batch calculation using the finite element program.

[0071] Step 2: Construct loss function And build a neural network and train the neural network to minimize the value of the loss function;

[0072] Step 3: Based on the machine vision system of grating projection, the surface depth information of the gas cylinder represented by a point cloud of the order of millions is obtained, and the radial expansion (true value) of different cross sections of the gas cylinder under the design pressure is obtained by combining the cylindrical fitting algorithm;

[0073] Step 4: Input the inner diameter, thickness, number of composite material layers, ply angle and radial expansion (true value) of the gas cylinder into the trained neural network to obtain the predicted burst pressure value p brust .

[0074] The step 1 specifically includes the following steps:

[0075] Step (11), establish a finite element model in ABAQUS / CAE, mesh the composite layer using solid units, constrain the end bosses, use symmetric boundary conditions for the axial section, and apply internal pressure. The interface weak bonding defects are automatically randomly distributed to each unit according to statistical indicators, and a large number of virtual test models with randomly distributed defects that meet the statistical indicators are automatically generated.

[0076] Step (12): Perform response analysis on the gas cylinder to obtain the radial expansion under the design pressure.

[0077] Step (13): Use the finite element program to calculate and obtain a more accurate bursting pressure value. (Analog value).

[0078] The step 2 specifically includes the following steps:

[0079] Step (21): Based on a large number of virtual test models, the radial expansion under the design pressure is obtained. The burst pressure prediction value is obtained by using the Newton-Raphson method.

[0080] Step (22), the neural network output p brust With the Newton-Raphson method we get Subtract the calculated error, i.e. the loss function

[0081] Step (23), build the framework of the neural network, the input layer has multiple neurons, input D i ,t,n,θ n , The output layer has one neuron, which outputs the predicted value of bursting pressure. The number of layers and neurons in the middle layer depends on the situation and is adjusted according to the iteration results.

[0082] Step (24), back-propagate the error value obtained in step (22), and adjust the weights and biases autonomously within the neural network until the error is minimized and the iteration is completed. If the error is still large after multiple iterations, the number of hidden layers and the number of neurons in each layer can be adjusted until the best result is obtained and the neural network training is completed.

[0083] The step three specifically includes the following steps:

[0084] Step (31), build a binocular grating projection measurement system, the system consists of two CCD cameras, a grating projector and a computer, connect them through a network cable and a switch, and adjust the position and parameters of the grating projector to obtain high-quality images and point cloud data.

[0085] Step (32), connect the gas cylinder to the pressurized system. Fill the gas cylinder with water until the thread of the bottle mouth, then tighten the joint to ensure that the bottle mouth is sealed. Fill the gas cylinder with water until the thread of the bottle mouth, then tighten the joint to ensure that the bottle mouth is sealed.

[0086] Step (33), start the pressurization system, apply internal pressure, monitor pressure changes, and adjust the pressure rate.

[0087] Step (34), scanning and photographing the gas cylinder to obtain point cloud data under different internal pressure states.

[0088] Step (35), the obtained large amount of point cloud data is subjected to noise reduction by removing points that are far away from the cluster, and the point cloud is thinned by uniform voxel downsampling, and finally the region of interest is extracted.

[0089] Step (36), using a cylindrical fitting function to fit the cylinder in the point cloud data, the fitted cylinder can be approximately regarded as the gas cylinder entity area corresponding to the selected area.

[0090] Step (37), performing N fittings on the point cloud data under different internal pressure states, obtaining N groups of cylinder radii Ri (i=1, 2, 3...N), and obtaining the radial expansion (true value) of different cross sections of the gas cylinder.

[0091] The step 4 specifically includes the following steps:

[0092] Step (41), select the appropriate radial expansion amount (true value) and design parameter D i ,t,n,θ n Enter neural networks.

[0093] Step (42), obtaining a bursting pressure prediction value through the trained neural network.

[0094] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for predicting the bursting pressure of a carbon fiber fully wrapped gas cylinder based on machine vision and deep learning, comprising the following steps: Step 1: Establish a virtual test model, perform response analysis on the gas cylinder, output the simulated value of radial expansion under the design pressure, and then perform batch calculations using a finite element program to predict the bursting pressure of the gas cylinder; Specifically include: (1.1) Aiming at the distribution pattern of weak interface bonding defects in carbon fiber fully wrapped gas cylinders, a large number of virtual test models with randomly distributed defects that meet statistical indicators are automatically generated. The weak interface bonding defects include the reduction of interface bonding strength and the complete delamination of the interface. In the virtual test model, defects can be automatically and randomly assigned to each unit according to the statistical indicators of the severity of the defects. (1.1.1) Interface model; weak interface bonding defects are described as relaxation of interface bonding. When interlayer slip occurs, although the interlayer lateral stress is still continuous under equilibrium conditions, the displacement between layers is no longer continuous. When the interface is completely debonded, it becomes a delamination defect, which can slide relative to each other or change from closed to open when loaded. In the interface description, it is assumed that the interface traction force T n(t) for ||u n(t) ||, the interface condition is expressed as: In the formula, Γ represents the interface layer, ||*|| Γk,k+1 It represents the discontinuity of the physical quantity at the interface Γ between two adjacent layers k and k+1, which can be expressed as ||*|| Γk,k+1 =(*) (k) -(*) (k+1) ;n j is a unit vector orthogonal to the interface; K n and K t They are the interface normal stiffness and interface tangential stiffness, respectively, which can be considered as the characterization of the interface bonding strength, with the unit of GPa / mm; (1.1.2) Micromechanical modeling; the equivalent stiffness coefficient of the laminated structure considering different bonding degrees is obtained as follows: In the formula, In the formula, In the formula, C ijkl (1) , C ijkl (2) and C ijkl (I) are the material stiffness coefficients of the fiber reinforced layers 1, 2 and the interface layer I, respectively; (1.1.3) Defect probability distribution model; When considering the random perturbation of defects in fiber reinforced composite structures, the defect structure characterized by parameters can be regarded as a mathematical random sequence; In the description of weak bonding defect distribution, the distribution of exponent x is used as a control variable to describe the severity of weak bonding defects; Assume that x is a random variable and obeys the expectation x u , the variance is S 2 The normal distribution model description is denoted as X(x)~N[x u ,S 2 ], then the corresponding K n Obey the lognormal distribution ln(K n )~N(x u ,S 2 ), whose expression is: (1.1.4) Spatial random sampling: The statistical distribution law describing the severity of defects is given above, but the random distribution of defects in the reinforced layer of the composite material of the gas cylinder needs to be specified in the description of its random distribution in the spatial discrete field; Discretize each region of the structure into small units i , e i ∈Y, i=1,2,3,...N, the discrete unit scale corresponds to the representative volume unit that can characterize the defect; Y is the sum of the units of the composite structure discretization; in the finite element model, the sampling space corresponds to the e of the discrete unit i Assume that the total number is n, and its position information is associated with unit number i; select a unit of size n from n d The random sampling of y represents the occurrence of defects in any unit; then the distribution probability of defects p i It can be expressed as: Where, subscript d represents the defect unit, and subscript j represents the sub-division of the fiber reinforced composite layer; (1.2) establishing two different pressure levels according to the design pressure, wherein the two different pressure differences are equal to the design pressure; performing response analysis on the gas cylinder containing randomly distributed defects described in steps (1.1.1) to (1.1.4) to obtain a simulated value of radial expansion under the design pressure; (1.3) Considering the dispersion and spatial randomness of the defect morphology, the finite element program is automatically executed for batch calculation to predict the bursting pressure of the gas cylinder containing randomly distributed defects described in steps (1.1.1) to (1.1.4); Step 2: Build and train a deep learning neural network. The unique random defect distribution of each gas cylinder determines the radial expansion of different cross sections along the axial direction and the final bursting pressure during its service step. However, it is difficult to establish an explicit functional relationship between the three. Deep learning can establish the relationship between radial expansion and bursting pressure, including: (2.1) Based on step 1, a large number of radial expansion simulation values ​​and bursting pressures under design pressure are obtained The bursting pressure loss function is constructed as: Build a neural network; (2.2) Train the neural network to minimize the loss function; (2.2.1) Input the cylinder inner diameter, thickness, number of composite material layers, ply angle, and radial expansion simulation value parameters; (2.2.2) Through the neural network, the relationship between the key information of the cylinder inner diameter, thickness, number of composite material layers, ply angle and radial expansion simulation value and the bursting pressure was preliminarily established; (2.2.3) Output the predicted bursting pressure value; (2.2.4) Then substitute the bursting pressure output by the neural network and the bursting pressure obtained by finite element calculation in step 1 into the loss function to calculate the error; (2.2.5) Then, the bias and weight of each layer are continuously adjusted through the network's back propagation until the error is minimized and the iteration is completed; Step 3, obtaining the true value of radial expansion under design pressure based on grating projection technology, specifically includes: (3.1) Based on the machine vision system of grating projection, the surface depth information of the gas cylinder under different pressures is obtained in the form of point clouds in the millions; (3.2) Combined with the cylindrical fitting algorithm, the true value of the radial expansion of different cross sections of the gas cylinder under the above pressure difference is obtained; (3.2.1) Perform N fittings on the point cloud data under the internal pressure P0 state to obtain N sets of cylinder radii R i , i = 1, 2, 3 ... N, and find the average value, then the radius of the cylinder in the P0 state is R0: Similarly, the radius R1 of the cylinder under the internal pressure P1 is: (3.2.2) Subtracting the radius values ​​of the two states R0<R1 can obtain the radial expansion of the selected cross section. for: Step 4, using the trained neural network to predict the burst pressure, as follows: (4.1) For a given type of gas cylinder, at any step of filling and depressurization, a machine vision system based on grating projection is used to collect and analyze images and output three-dimensional point clouds and true value information of radial expansion; (4.2) Input the true values ​​of the inner diameter, thickness, number of composite material layers, ply angle and radial expansion of the gas cylinder, and obtain the predicted value of the bursting pressure through the trained neural network, so as to realize efficient and reliable online prediction and warning of the bursting pressure of the gas cylinder.