Jacket K-type pipe node strength prediction method based on error back propagation

Abstract

The invention provides a jacket K-type pipe node strength prediction method based on error back propagation. The jacket K-type pipe node strength prediction method mainly comprises the following steps: (1) establishing a finite element model of a K-type pipe node; (2) compiling a command stream to realize finite element parameterization batch modeling of the pipe nodes; (3) performing data preparation and preprocessing; (4) establishing a pipe node strength prediction model based on a neural network; and (5) realizing a pipe node strength prediction model by using programming. The method is reasonable in conception, the strength range is directly predicted through the values of the pipe node design parameters, and therefore a large amount of finite element analysis is avoided.

Description

Technical Field

[0001] The invention relates to the technical field of offshore platform tubular node manufacturing, and in particular to a method for predicting the strength of a jacket K-type tubular node based on error back propagation. Background Art

[0002] The construction, development and innovation of offshore platform projects are important foundations for the marine industry. At present, many offshore platforms in my country have been in service for many years. Tubular structures are widely used in offshore platforms, among which tubular nodes are the technical focus and difficulty in the design and construction of platform structures. The stress on the tubular node structure is much greater than that on other parts, and the welds are concentrated. The load transfer between components needs to be carried out through the tubular node. Therefore, it is the part that is most prone to cracks and damage. In its design process, certain strength and stiffness requirements should be met, and multiple parameters need to be considered: for example, the K-type tubular node considers the cross-sectional diameter and wall thickness of the chord and the strut, the spacing between the struts, the angle, etc. Since multiple parameters are involved in the optimization of the tubular node, it is very unrealistic to find the optimal parameter combination through experimental means. Rapidly and accurately establishing a tubular node structural model is the basis and prerequisite for the analysis and research of tubular nodes.

[0003] Among the methods of stress analysis of offshore platform pipe nodes, the finite element method is the most widely used today. It can replace experiments to a certain extent, so that researchers or designers can effectively complete the optimization work at a lower cost. In the optimization process, although the finite element method can be used to predict the strength of pipe nodes, each design parameter combination needs to go through the process of modeling, analysis and result extraction. Obviously, the optimization based on the finite element method requires multiple numerical simulations and result comparison analysis, which takes a lot of time to find a relatively good combination. Especially when conducting three-dimensional model analysis, this disadvantage is more prominent.

[0004] To sum up, it is necessary to make further innovations on existing technologies. Summary of the invention

[0005] In view of the technical problems existing in the above-mentioned background technology, the present invention proposes a method for predicting the strength of pipe nodes based on error back propagation, which has a reasonable concept and realizes direct prediction of the strength range by using the numerical values ​​of the design parameters of the pipe node of the jacket, thereby avoiding a large amount of finite element analysis.

[0006] In order to solve the above technical problems, the present invention provides a method for predicting the strength of a jacket K-type pipe joint based on error back propagation, which mainly includes the following steps:

[0007] (1) Establish the finite element model of K-type tubular node;

[0008] (2) Write command streams to implement finite element parametric batch modeling of pipe nodes;

[0009] (3) Perform data preparation and preprocessing;

[0010] (4) Establish a pipe joint strength prediction model based on neural network;

[0011] (5) Use programming to implement the pipe node strength prediction model.

[0012] The method for predicting the strength of the K-type pipe node of the jacket based on error back propagation, wherein the specific process of the step (1) is as follows: first, according to the K-type pipe node commonly used in the jacket platform, the geometric model of the pipe node is constructed using finite element software, and the chord and the strut are established, and the angle between the chord and the strut is θ; then, on the basis of the completed construction of the geometric model, the finite element model mesh is divided, and according to the actual situation and calculation requirements, a regular quadrilateral mesh is manually divided and analyzed to obtain the model mesh division result; finally, the material properties, unit type, boundary conditions and load application conditions are set in the finite element software.

[0013] The method for predicting the strength of the K-type pipe node of the pipe rack based on error back propagation, wherein the specific process of the step (2) is: using a set of parameters to define the geometric dimensions of the pipe node model, and constraining the relationship between the geometric dimensions, realizing the topological relationship between the geometric dimensions, determining the relevant parameters of the geometric model that need to be changed, and displaying the stress cloud map after the finite element model is established and Solve is completed.

[0014] The method for predicting the strength of a jacket K-type tubular node based on error back propagation, wherein the geometric dimensions of the tubular node model can be defined according to dimensionless parameters:

[0015]

[0016] In the above formulas (1)-(4), L represents the length of the chord tube; D represents the outer diameter of the chord tube; d represents the outer diameter of the support tube; T represents the wall thickness of the chord tube; t represents the wall thickness of the support tube; α is the ratio of the length of the chord tube to the outer diameter, which reflects the flexibility of the chord tube; β is the ratio of the outer diameter of the chord tube to the diameter of the support tube, which reflects the load transfer and stress distribution; γ is the ratio of the outer diameter of the chord tube to the wall thickness of the chord tube, which reflects the radial stiffness of the chord tube; τ is the ratio of the wall thickness of the support tube to the wall thickness of the chord tube, which reflects the relative bending stiffness of the chord tube and the support tube.

[0017] The method for predicting the strength of K-type tubular nodes of a conductor rack based on error back propagation, wherein the specific process of step (3) is as follows: after batch running the example model, extracting the finite element results calculated by the finite element analysis of the model, i.e., different design parameter combinations of the tubular nodes and the obtained maximum stress values, and arranging the extracted finite element results; normalizing the finite element results and assigning a smaller initial value to avoid the data being 0.

[0018] The method for predicting the strength of the K-type pipe node of the conductor rack based on error back propagation, wherein: the maximum stress value is divided into three categories: "safe", "uncertain" and "dangerous", and 0, 1 and 2 represent "safe", "uncertain" and "dangerous" respectively.

[0019] The method for predicting the strength of a K-type pipe node of a pipe rack based on error back propagation, wherein: the step (4) is to predict the structural strength of the pipe node according to the finite element results, and then to establish a pipe node strength prediction model using the error back propagation algorithm, and the architecture of the pipe node strength prediction model includes an input layer, a hidden layer, and an output layer;

[0020] The input layer selects the structural parameters of the pipe nodes in combination with the structural characteristics of the pipe nodes in actual engineering, and uses the structural parameters as the input nodes;

[0021] The hidden layer uses a single hidden layer error back propagation algorithm to predict the safety of the pipe node; the number of hidden layer nodes is determined by combining multiple tests and reference empirical formulas; the reference empirical formula is as follows:

[0022]

[0023] m<n-1 (6);

[0024] m = log2 n (7);

[0025] In the above formulas (5)-(7), m represents the number of hidden layer nodes, n represents the number of input layer nodes, l represents the number of output layer nodes, and the constant a = 1 to 10;

[0026] The output layer selects three safety states, namely "safe", "uncertain" and "dangerous", as output.

[0027] The method for predicting the strength of K-type pipe joints of a pipe rack based on error back propagation, wherein: the pipe joint strength prediction model in step (5) can be divided into six parts: initializing the network, training the network, querying the network, assigning variables, importing data and calculating the accuracy;

[0028] The specific process of step (5) is as follows: first, the neural network model is trained, and in the process of training the neural network model, the network weights are continuously adjusted by back propagation of the result errors based on the learned sample data, so as to complete the training, and finally, the trained model is verified and tested;

[0029] The initialized network needs to set the number of input layer nodes, hidden layer nodes and output layer nodes of the neural network model, and these node numbers define the shape and size of the neural network model; in addition, the initialized network also needs to set the learning rate, initialize the link weights and define the activation function; the initial link weights are randomly generated and the random numbers are normally distributed, with a mean of 0 and a standard deviation of -0.5 power of the number of hidden layer nodes, i.e., the initial weights between the input layer and the hidden layer, and -0.5 power of the number of output layer nodes, i.e., the initial weights between the hidden layer and the output layer;

[0030] The activation function is a sigmoid function, and the formula is as follows:

[0031]

[0032] The training network is divided into two parts. The first part is to calculate the output for a given training sample; the second part is to compare the calculated output with the expected output, and improve the network weights based on the difference optimization, by calculating the loss function L with respect to each weight parameter W. (l) and bias b (l) The gradient of , to update each weight parameter W (l) and bias b (l) To minimize losses;

[0033] For each layer l, first calculate the loss function relative to the activation value A (l) The partial derivative of :

[0034]

[0035] Next, calculate the linear transformation output Z (l) The partial derivative of :

[0036] δ (l) =δ (l+1) ·g′(Z (l) ) (10);

[0037] In the above formula (10), g′(Z (l) ) is the derivative of the activation function;

[0038] Then, update the weights and biases of each layer:

[0039]

[0040] The weights and biases are updated using the gradient descent method:

[0041]

[0042] In the above formula (13), η is the learning rate;

[0043] The query network needs to transmit input signals from input layer nodes, combine all inputs with all correct link weights, generate a combined and adjusted signal matrix, transmit it to each hidden layer node, use the link weights to adjust the signal through the hidden layer and then transmit it to the output layer, the output layer gradually updates the link weights according to the back propagation algorithm to minimize the prediction error, and finally outputs it from the output layer; when the data transmitted by the input layer is fed to a given hidden layer node or output layer node, the link weights are used to adjust the data transmitted by the input layer and the sigmoid activation function is applied to suppress the signals from these nodes;

[0044] The variables are assigned values ​​according to the error back propagation network structure. The number of input layer nodes of the neural network model is 5, the number of hidden layer nodes is 4, the number of output layer nodes is 3, and the learning rate is 0.1;

[0045] The data import needs to create a neural network object, import a training data set to train the network and import a test data set to test the network. When the training data is insufficient, the method of iterating all sample data multiple times can be adopted, that is, increasing the number of training rounds in the neural network model, and performing network training and learning to improve the training effect;

[0046] The output result of the pipe node strength prediction model is the safety status of the pipe node under different design parameters, and the maximum value represents the predicted state; since the output result is a one-dimensional array of three numbers in the range of 0 to 1, a part of the code is added to intuitively represent the training effect by using accuracy measurement.

[0047] The method for predicting the strength of the K-type pipe node of the conductor rack based on error back propagation, wherein the process of training the neural network model is: first, initialize the weight of the network and the bias value of the neuron, and select an input sample; then, perform forward propagation, calculate the output of the hidden layer and the output layer, and then calculate the error according to the difference between the expected output and the actual output; pass the error to each hidden layer through back propagation, calculate the error of the hidden layer, and adjust the weight and bias value; then check whether the training of the current sample is completed, if not, continue to adjust, otherwise increase the number of iterations and select the next sample, and repeat the above process; finally, when all samples are trained, the entire training process ends.

[0048] By adopting the above technical solution, the present invention has the following beneficial effects:

[0049] The tubular node strength prediction method based on error back propagation of the present invention is reasonably conceived, realizes safe prediction of tubular node strength, is used in the tubular node design stage, simulates multiple parameters, reduces the number of experiments, and provides a faster evaluation of the tubular node structural strength. It can effectively solve the current cumbersome problem of finite element analysis in the tubular node design stage. At the same time, it can effectively improve efficiency, reduce costs, and accelerate the design and development of tubular nodes.

[0050] The present invention aims at the large number of design parameters that need to be considered in the process of tube node design, and uses BP neural network to establish a tube node strength prediction model. This model is suitable for processing information processing problems that need to consider multiple factors at the same time and have complex and fuzzy conditions. The working process is to learn based on a small number of representative samples and continuously adjust its own weights. A large number of adjustable parameters within the network are highly flexible and have strong fault tolerance, so it can analyze and process more complex nonlinear systems, such as prediction and control problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1 Schematic diagram of the flow of a method for predicting pipe node strength based on error back propagation in an embodiment of the present invention;

[0053] Figure 2 Schematic diagram of a K-type tube node model in an embodiment of the present invention;

[0054] Figure 3 The influence of the number of hidden layer nodes on the prediction accuracy in the embodiment of the present invention;

[0055] Figure 4 Schematic diagram of the BP network training process in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The present invention is further explained below in conjunction with specific implementation modes.

[0058] like Figure 1As shown, this embodiment provides a method for predicting the strength of a pipe node based on error back propagation, comprising the following steps:

[0059] S100, Establish the finite element model of K-type tube node

[0060] According to the K-type tube node commonly used in the jacket platform, the geometric model of the tube node is constructed using finite element software, and the chord and the strut are established. The angle between the chord and the strut is θ, as shown in Figure 2 The pipe node model established in this embodiment is as shown in Figure 2 As shown, the length is 600 mm extending outward from the intersection of the chord and the strut axis, and the angle between the chord and the strut is θ.

[0061] The K-type node symmetrical structure established in this embodiment, that is, the two brace angles and brace wall thickness and other parameters have the same value, and the geometric parameter settings are shown in Table 1-1:

[0062]

[0063] Then, the finite element model is meshed based on the completed geometric model. According to the actual situation and calculation requirements, regular quadrilateral meshes are manually generated for analysis. The calculation efficiency is effectively improved while ensuring the calculation accuracy, and the model meshing results are obtained.

[0064] Finally, set the material properties (including yield strength, elastic modulus, Poisson's ratio, stress-strain relationship, material yield criterion, etc., ignoring the influence of residual stress of weld in node area on pipe node), unit type (select SOLID185), boundary conditions, and load application in the finite element software to achieve state simulation of pipe node. In this embodiment, CCSD36 is selected as the material of pipe node finite element model, and yield strength f y =355MPa, the stress-strain relationship uses the ideal elastic-plastic model. The elastic modulus, Poisson's ratio, and yield criterion of the material are defined to comply with the Von Mise yield criterion and related flow laws, and the influence of residual stress of the weld in the node area on the tube node is not considered. SOLID185 is selected for relevant analysis and research. The unit is defined by 8 nodes, and each node has 3 degrees of freedom for translation along the x, y, and z directions. The above boundary conditions are specifically as follows: the left end of the chord is a fixed hinge support, and the right end is a sliding hinge support, which can slide along the axis of the chord; one end of the brace is welded to the chord, and the other end is a cantilever end and load application. The two cantilever ends apply axial force and bending moment respectively, and the state simulation of the tube node is realized. MASS21 is used to establish a node at the center of the end face of the brace tube, which is rigidly coupled with all nodes on the end face of the brace. The load transmitted to the tube node from the upper part is simulated by applying a load to this point.

[0065] S200, write command flow to realize finite element parametric batch modeling of pipe nodes

[0066] A set of parameters is used to define the geometric dimensions of the pipe node model, and the relationship between the dimensions is constrained to achieve the topological relationship between the geometric dimensions, determine the relevant parameters of the geometric model that need to be changed, and display the stress cloud map after the finite element model is established and Solve is completed.

[0067] The geometric dimensions of the tube node model can be defined according to dimensionless parameters:

[0068]

[0069] In the above formulas (1)-(4), L represents the length of the chord, in mm; D represents the outer diameter of the chord, in mm; d represents the outer diameter of the support tube, in mm; T represents the wall thickness of the chord, in mm; t represents the wall thickness of the support tube, in mm; α is the ratio of the length of the chord to the outer diameter, reflecting the flexibility of the chord; β is the ratio of the outer diameter of the chord to the diameter of the support tube, reflecting the load transfer and stress distribution; γ is the ratio of the outer diameter of the chord to the wall thickness of the chord, reflecting the radial stiffness of the chord; τ is the ratio of the wall thickness of the support tube to the wall thickness of the chord, reflecting the relative bending stiffness of the chord and the support tube.

[0070] With reference to the offshore steel structure design specification "FATIGUE DESIGNOF OFFSHORE STEEL STRUCTURES" published by the Norwegian Classification Society (DNV) in 2010, the effective range of the geometric dimensions of offshore platform pipe nodes is: 4≤α≤40; 4≤β≤40; 4≤γ≤40; 4≤τ≤40.

[0071] SOLID185 units are selected for modeling. A top-down approach is adopted to directly create the body. Then, the geometric model of the tube node is generated through various Boolean operations. After appropriate meshing, the finite element model is obtained.

[0072] The specific parameterization process is as follows:

[0073] S201, determining relevant parameters of the geometric model that need to be changed, and assigning values ​​to the variables;

[0074] S202, set up the analysis environment, define the unit type as SOLID185, real constants and material properties, etc.;

[0075] S203, establishing chords and struts, and obtaining a geometric model of the tube node using Boolean operations;

[0076] S204, grid division;

[0077] S205, applying constraints to the left and right ends of the chord, applying full constraints to the left end, applying constraints in the X and Y directions to the right end, and applying loads after rigid coupling of the cantilever end faces of the struts;

[0078] S206. After Solve is completed, the stress cloud diagram is displayed.

[0079] Furthermore, after extracting the complete command stream, the command stream is parameterized, and all command stream files that need to calculate the model are stored in a folder in the form of inp. Then, the corresponding batch modeling command stream is written and run in the finite element software command window to implement batch modeling calculations. By changing the path command stream, the establishment and analysis calculations of all models are completed.

[0080] S300, data preparation and preprocessing

[0081] After batch running the calculation model, the different design parameter combinations of the pipe nodes calculated by the finite element analysis of the model and the obtained maximum stress values ​​are extracted, and the finite element results and data are sorted, totaling 110 groups. Specifically, the strength influencing factors considered involve the chord diameter, chord wall thickness, strut diameter, strut wall thickness and strut angle; the maximum stress value is divided into three categories according to "safe", "uncertain" and "dangerous". According to the reduction coefficient of 7%, the maximum stress less than 330MPa is taken as "safe", between 330MPa and 355MPa is "uncertain", and greater than 355MPa is dangerous, and 0, 1, and 2 are used to represent "safe", "uncertain" and "dangerous" respectively, which is convenient for the data reading of the neural network model. The extracted data is normalized and a smaller initial value is given to avoid the situation where the data is 0 so that the neural network has no effect due to weights. The original data is processed into a comparable standard sequence as much as possible, and the training effect is significantly improved compared with the original method, and the prediction result accuracy is higher.

[0082] S400, Establishing a model for predicting the strength of pipe joints based on neural networks

[0083] Design a neural network model, the architecture of the neural network model includes:

[0084] In the input layer, the structural parameters of the tube nodes are selected in combination with the structural characteristics of the tube nodes in the actual project, and the structural parameters are used as the input nodes. Specifically, in combination with the symmetrical structural characteristics of the K-type tube node in this embodiment, several parameters that have a greater impact on the safety of the tube node are obtained, namely, the chord diameter, the chord wall thickness, the brace diameter, the brace wall thickness and the brace angle, and these five parameters are used as the input nodes.

[0085] Hidden layer, in theory, the three-layer error back propagation algorithm model can obtain any m-dimensional to n-dimensional mapping, and the calculation results are better in practical applications. The present invention uses a single-layer hidden layer error back propagation algorithm to predict the safety of the pipe node; the number of hidden layer nodes is generally determined by combining multiple tests and reference empirical formulas; the number of hidden layer nodes in the present invention ranges from 3 to 12; the reference empirical formula is as follows:

[0086]

[0087] m<n-1 (6);

[0088] m = log2n (7);

[0089] In the above formulas (5)-(7), m represents the number of hidden layer nodes, n represents the number of input layer nodes, l represents the number of output layer nodes, and the constant a=1~10.

[0090] Specifically, the optimal number of hidden layer nodes is determined by experimental method. As the number of hidden layer nodes increases, the training time increases, and the highest peak value appears when the accuracy is 4, such as Figure 3 Therefore, in the current case, the optimal number of hidden layer nodes is 4.

[0091] Output layer, for the output layer nodes, according to the invented method, three safety states, namely "safe", "uncertain" and "dangerous", are selected as output;

[0092] The algorithm adopted by the pipe joint strength prediction model of the present invention is the error back propagation algorithm (BP network), and the activation function is the sigmoid function. The formula is as follows:

[0093]

[0094] S500, using programming to implement a pipe node strength prediction model; the pipe node strength prediction model can be divided into six parts, namely initializing the network, training the network, querying the network, assigning variables, importing data, and calculating accuracy; the specific process is:

[0095] First, train the neural network model, such as Figure 4 As shown in the figure, first calculate the state and activation value of each layer until the last layer (that is, the signal is forward propagated); calculate the error of each layer, and the error calculation process is advanced from the last layer to the front (that is, the error is back propagated); calculate the gradient of the weight of each neuron connection; update the parameters according to the gradient descent rule (the goal is to reduce the error, iterate the above steps until the stopping criterion is met; in the process of training the neural network model, based on the sample data of learning, the network weights are continuously adjusted through the back propagation of the result error to complete the training, and finally the trained model is verified and tested.

[0096] Specifically, the initialization network needs to set the number of input layer nodes, hidden layer nodes and output layer nodes of the neural network model, and these node numbers define the shape and size of the neural network model; In addition, the initialization network also needs to set the learning rate, initialize the link weights and define the activation function (using sigmoid function). The most important part of the network is the link weights, which play a role in calculating feedforward signals and back propagation errors, and continuously optimize the link weights themselves based on the gradient descent method when trying to improve the network. Wherein the initial link weights of the present invention are randomly generated and the random number adopts a normal distribution, with a mean of 0, and the standard deviation is respectively -0.5 power (initial weight between input layer and hidden layer) of the number of hidden layer nodes and -0.5 power (initial weight between hidden layer and output layer) of the number of output layer nodes.

[0097] Specifically, the training network is divided into two parts. The first part is to calculate the output for a given training sample. That is, let the input sample be X, the model parameter be W, and the target output be Y. The calculation process of the neural network is as follows:

[0098] For each layer l:

[0099] Z (l) =W (l) A (l-1) +b (l) ;

[0100] Among them, A in the above formula (l-1) is the activation value of the previous layer, and the initial input is A (0) =X;W (l) is the weight matrix of the lth layer, b (l) is the bias vector; Z (l) is the linear transformation output of the lth layer;

[0101] The second part is to compare the calculated output with the expected output and improve the network weights based on the difference optimization, that is, back propagation is the process of optimizing network weights based on the gradient descent method. The goal of back propagation is to calculate the loss function L with respect to each weight parameter W. (l) and bias b (l) The gradient of , to update each weight parameter W (l) and bias b (l) To minimize losses.

[0102] For each layer l, first calculate the loss function relative to the activation value A (l) The partial derivative of :

[0103]

[0104] Next, calculate the linear transformation output Z (l) The partial derivative of :

[0105] δ (l) =δ (l+1) ·g′(Z (l) )(10);

[0106] Among them, g′(Z (l) ) is the derivative of the activation function.

[0107] Then, update the weights and biases of each layer:

[0108]

[0109] The weights and biases are updated using the gradient descent method:

[0110]

[0111] Among them, η is the learning rate (Learning Rate);

[0112] Specifically, the query network needs to transmit the input signal from the input layer node, combine all the inputs with all the correct link weights, generate a combined adjusted signal matrix, transmit it to each hidden layer node, and pass it to the output layer after adjusting the signal through the hidden layer using the link weights. The output layer uses the link weights to gradually update according to the back propagation algorithm to minimize the prediction error, and each link weight will be gradually updated according to the back propagation algorithm to minimize the prediction error, and finally output from the output layer; when the data passed by the input layer is fed to a given hidden layer node or output layer node, the link weights are used to adjust the data passed by the input layer, and the sigmoid activation function is applied to suppress the signals from these nodes.

[0113] Specifically, the variables are assigned according to the error back propagation network structure. The number of nodes in the input layer of the neural network model is 5, the number of nodes in the hidden layer is 4, the number of nodes in the output layer is 3, and the learning rate is 0.1.

[0114] Specifically, importing data requires creating a neural network object, importing a training data set to train the network, and importing a test data set to test the network. When research conditions are limited and training data is insufficient, you can iterate all sample data multiple times, that is, increase the number of training rounds (epochs) in the neural network model, and perform network training to improve the training effect.

[0115] Specifically, the output result of the pipe node strength prediction model is the safety status of the pipe node under different design parameters, and the maximum value represents the predicted state; since the output result is a one-dimensional array of three numbers in the range of 0 to 1, it is not convenient for researchers to study and analyze, so a part of the code is added to adopt the accuracy measurement method to intuitively represent the training effect.

[0116] like Figure 4 As shown in the figure, the above training process of the neural network model is as follows: first, initialize the network weights and neuron bias values, and select an input sample; then, perform forward propagation, calculate the outputs of the hidden layer and the output layer, and then calculate the error based on the difference between the expected output and the actual output; pass the error to each hidden layer through back propagation, calculate the error of the hidden layer, and adjust the weight and bias values; then check whether the training of the current sample is completed, if not, continue to adjust, otherwise increase the number of iterations and select the next sample, and repeat the above process; finally, when all samples are trained, the entire training process ends.

[0117] In this embodiment, an artificial neural network is established based on strength analysis and safety prediction based on different design parameters of K-type pipe nodes. 110 sample data in random order are used as training data and 30 test data. Through analysis and optimization of the artificial neural network, the artificial neural network is trained with 5 input nodes, 3 output layer nodes, 4 hidden layer nodes and a learning rate of 0.1. The results show that the safety prediction of K-type pipe nodes can reach up to 92.86%, which meets the 10% error requirement of the project. In actual engineering, the use of BP neural network model to predict safety instead of finite element analysis with finite element software can greatly improve the efficiency of design and construction of pipe nodes.

[0118] The present invention has a reasonable concept and realizes direct prediction of the strength range by using the numerical values ​​of the design parameters of the K-type tube node of the conductor rack, thereby avoiding the optimization based on the finite element method, which requires multiple numerical simulations and result comparison and analysis, and consumes a lot of time to find a relatively good combination. The present invention can effectively improve efficiency, reduce costs, and accelerate the design and development of K-type tube nodes.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the strength of a K-type pipe joint of a jacket based on error back propagation, characterized in that: The main steps include: (1) Establish the finite element model of K-type tubular node; (2) Write command streams to implement finite element parametric batch modeling of pipe nodes; (3) Perform data preparation and preprocessing; (4) Establish a pipe joint strength prediction model based on neural network; (5) Use programming to implement the pipe node strength prediction model.

2. The method for predicting the strength of a K-type pipe joint of a pipe rack based on error back propagation according to claim 1, characterized in that: The specific process of step (1) is as follows: first, according to the K-type tube node commonly used in the jacket platform, the geometric model of the tube node is constructed using finite element software, and the chord and the strut are established, and the angle between the chord and the strut is θ; then, on the basis of the completed construction of the geometric model, the finite element model is meshed, and a regular quadrilateral mesh is manually generated according to the actual situation and calculation requirements for analysis to obtain the model meshing result; finally, the material properties, unit type, boundary conditions and load application conditions are set in the finite element software.

3. The method for predicting the strength of a K-type pipe joint of a pipe rack based on error back propagation according to claim 1, characterized in that: The specific process of step (2) is as follows: using a set of parameters to define the geometric dimensions of the pipe node model, constraining the relationship between the geometric dimensions, realizing the topological relationship between the geometric dimensions, determining the relevant parameters of the geometric model that need to be changed, and displaying the stress cloud map after the finite element model is established and Solve is completed.

4. The method for predicting the strength of a K-type pipe joint of a pipe rack based on error back propagation according to claim 3, characterized in that: The geometric dimensions of the tube node model can be defined according to dimensionless parameters: In the above formulas (1)-(4), L represents the length of the chord tube; D represents the outer diameter of the chord tube; d represents the outer diameter of the support tube; T represents the wall thickness of the chord tube; t represents the wall thickness of the support tube; α is the ratio of the length of the chord tube to the outer diameter, which reflects the flexibility of the chord tube; β is the ratio of the outer diameter of the chord tube to the diameter of the support tube, which reflects the load transfer and stress distribution; γ is the ratio of the outer diameter of the chord tube to the wall thickness of the chord tube, which reflects the radial stiffness of the chord tube; τ is the ratio of the wall thickness of the support tube to the wall thickness of the chord tube, which reflects the relative bending stiffness of the chord tube and the support tube.

5. The method for predicting the strength of a K-type pipe joint of a jacket based on error back propagation according to claim 1, characterized in that: The specific process of step (3) is as follows: after batch running the example model, extract the finite element results of the model finite element analysis calculation, i.e., different design parameter combinations of the pipe node and the obtained maximum stress value, and sort out the extracted finite element results; normalize the finite element results and assign a smaller initial value to avoid the data being 0.

6. The method for predicting the strength of a K-type pipe joint of a jacket based on error back propagation according to claim 5, characterized in that: The maximum stress values ​​are divided into three categories: "safe", "uncertain" and "dangerous", and 0, 1 and 2 represent "safe", "uncertain" and "dangerous" respectively.

7. The method for predicting the strength of a jacket K-type joint based on error back propagation according to claim 5, characterized in that: The step (4) is to predict the structural strength of the tube node according to the finite element results, and then establish a tube node strength prediction model using an error back propagation algorithm, and the architecture of the tube node strength prediction model includes an input layer, a hidden layer, and an output layer; The input layer selects the structural parameters of the pipe nodes in combination with the structural characteristics of the pipe nodes in actual engineering, and uses the structural parameters as the input nodes; The hidden layer uses a single hidden layer error back propagation algorithm to predict the safety of the pipe node; the number of hidden layer nodes is determined by combining multiple tests and reference empirical formulas; the reference empirical formula is as follows: m<n-1 (6); m = log2 n (7); In the above formulas (5)-(7), m represents the number of hidden layer nodes, n represents the number of input layer nodes, l represents the number of output layer nodes, and the constant a = 1 to 10; The output layer selects three safety states, namely "safe", "uncertain" and "dangerous", as output.

8. The method for predicting the strength of a K-type pipe joint of a jacket based on error back propagation according to claim 1, characterized in that: The pipe node strength prediction model in step (5) can be divided into six parts: initializing the network, training the network, querying the network, assigning variables, importing data and calculating accuracy; The specific process of step (5) is as follows: first, the neural network model is trained, and in the process of training the neural network model, the network weights are continuously adjusted by back propagation of the result errors based on the learned sample data, so as to complete the training, and finally, the trained model is verified and tested; The initialized network needs to set the number of input layer nodes, hidden layer nodes and output layer nodes of the neural network model, and these node numbers define the shape and size of the neural network model; in addition, the initialized network also needs to set the learning rate, initialize the link weights and define the activation function; the initial link weights are randomly generated and the random numbers are normally distributed, with a mean of 0 and a standard deviation of -0.5 power of the number of hidden layer nodes, i.e., the initial weights between the input layer and the hidden layer, and -0.5 power of the number of output layer nodes, i.e., the initial weights between the hidden layer and the output layer; The activation function is a sigmoid function, and the formula is as follows: The training network is divided into two parts. The first part is to calculate the output for a given training sample; the second part is to compare the calculated output with the expected output, and improve the network weights based on the difference optimization, by calculating the loss function L with respect to each weight parameter W. (l) and bias b (l) The gradient of , to update each weight parameter W (l) and bias b (l) To minimize losses; For each layer l, first calculate the loss function relative to the activation value A (l) The partial derivative of : Next, calculate the linear transformation output Z (l) The partial derivative of : d (l) =d (l+1) ·g′(Z (l) ) (10); In the above formula (10), g′(Z (l) ) is the derivative of the activation function; Then, update the weights and biases of each layer: The weights and biases are updated using the gradient descent method: In the above formula (13), η is the learning rate; The query network needs to transmit input signals from input layer nodes, combine all inputs with all correct link weights, generate a combined and adjusted signal matrix, transmit it to each hidden layer node, use the link weights to adjust the signal through the hidden layer and then transmit it to the output layer, the output layer gradually updates the link weights according to the back propagation algorithm to minimize the prediction error, and finally outputs it from the output layer; when the data transmitted by the input layer is fed to a given hidden layer node or output layer node, the link weights are used to adjust the data transmitted by the input layer and the sigmoid activation function is applied to suppress the signals from these nodes; The variables are assigned values ​​according to the error back propagation network structure. The number of input layer nodes of the neural network model is 5, the number of hidden layer nodes is 4, the number of output layer nodes is 3, and the learning rate is 0.1; The data import needs to create a neural network object, import a training data set to train the network and import a test data set to test the network. When the training data is insufficient, the method of iterating all sample data multiple times can be adopted, that is, increasing the number of training rounds in the neural network model, and performing network training and learning to improve the training effect; The output result of the pipe node strength prediction model is the safety status of the pipe node under different design parameters, and the maximum value represents the predicted state; since the output result is a one-dimensional array of three numbers in the range of 0 to 1, a part of the code is added to intuitively represent the training effect by using accuracy measurement.

9. The method for predicting the strength of a K-type pipe joint of a jacket according to claim 8, characterized in that: The process of training the neural network model is as follows: First, initialize the network weights and neuron bias values, and select an input sample; Next, forward propagation is performed to calculate the output of the hidden layer and the output layer, and then the error is calculated based on the difference between the expected output and the actual output; The error is passed to each hidden layer through back propagation, the error of the hidden layer is calculated, and the weight and bias values ​​are adjusted; Then check whether the training of the current sample is completed. If not, continue to adjust. Otherwise, increase the number of iterations and select the next sample to repeat the above process. Finally, when all samples are trained, the entire training process ends.

Images