Intelligent decision method for hull plate curved surface forming based on hybrid enhanced intelligence

Abstract

The application discloses a hull plate curved surface forming intelligent decision method based on hybrid enhanced intelligence, carries out finite element simulation on a test hull plate, constructs a support vector machine model, adopts a Sigmoid function to update a convergence factor in the model, determines different weights according to different positions of the top three wolves, optimizes a penalty coefficient and a hyperparameter in the model, and constitutes a proxy model; simulation data, prediction data and artificial experience data are combined to form a multi-source database, which is expressed by a knowledge graph; processing parameters of a to-be-processed ship plate are preliminarily screened based on the knowledge graph to carry out first processing, local deformation error is calculated, a flame path position of second processing is fine-tuned, and the final flame path position is obtained through repetition; the processing parameters of the to-be-processed ship plate are subjected to multi-objective optimization by taking into account the weights of production time and production energy consumption, and optimal processing parameters are obtained; the proxy model is used as a prediction model, and the actual distribution of sample points is considered, so that the processing efficiency and decision accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of shipbuilding, in particular to a ship hull plate curved surface forming decision method for ship hull plate bending forming processing. BACKGROUND

[0002] In the ship hull plate bending forming processing method, water fire bending is an effective method, however, the current water fire bending method mainly relies on artificial experience to determine the processing scheme, and the automation degree is low, and the processing efficiency is poor.

[0003] The concept of hybrid enhanced intelligence refers to introducing human action into artificial intelligence systems to form a new intelligent form. This form is considered a viable and important growth pattern of artificial intelligence. Although the application field of hybrid enhanced intelligence is wide, there is no application of hybrid enhanced intelligence in the field of ship hull plate processing.

[0004] The method disclosed in the document with Chinese patent application publication number 114091304A and the name "Intelligent decision method for hydrogen-oxygen gas heat source processing of ship hull plate" determines the relationship between the processing parameters and the deformation amount of hydrogen-oxygen gas heating; establishes a heat source model for hydrogen-oxygen heat source processing of ship plate, obtains finite element simulation data; improves the grey wolf algorithm to optimize the support vector machine deformation prediction, obtains the prediction result, interpolates the finite element simulation data, obtains the complete deformation amount and processing parameter corresponding relationship, and constructs a hydrogen-oxygen flame processing ship plate database; the drawing of the target plate is obtained by the plate unfolding method, the processing path and the corresponding path deformation amount are obtained, and the processing parameter group is obtained by the corresponding relationship in the database; and the optimal green processing scheme is obtained. However, when establishing the prediction model, the actual distribution of the sample points is not considered, resulting in too concentrated sample points, which may distort the training results and lack representativeness.

[0005] The method disclosed in the document with Chinese patent application publication number CN102682139A and the name "A ship hull plate curved surface forming method" records experience knowledge by a large number of actual operations of the water fire bending numerical control equipment, stores a large number of processing parameters into a database to form a knowledge base of an expert system, establishes a mechanism model of the water fire bending through a neural network when the experience data accumulates to a certain degree, thereby forming a reasoning machine of the expert system, and the water fire bending numerical control equipment automatically gives a processing track according to the reasoning machine to realize full automation of the water fire bending. However, the patent method only discloses the actual forming method, and does not involve the actual decision processing scheme. SUMMARY

[0006] The application aims to solve the problems of the prior art, and provides a hull plate curved surface forming intelligent decision method based on artificial hybrid enhanced intelligence, which considers the actual distribution of sample points, effectively solves the problem of excessive reliance on artificial experience in existing ship plate processing engineering, and improves the ship plate processing efficiency and decision accuracy.

[0007] The hull plate curved surface forming intelligent decision method based on hybrid enhanced intelligence has the following steps:

[0008] Step 1): finite element simulation is performed on the test hull plate to obtain simulation local deformation amounts corresponding to different processing parameters and flame path positions, which form simulation data,

[0009] Step 2): the simulation data is divided into a training set and a test set, and after normalization processing, a support vector machine model f(x)=w·gΦ(x)+c is constructed, the hyperparameter g=2a·r1-a, the penalty coefficient c=2·ar2, w is a weight vector, Φ(x) is a mapping function, r1 and r2 are two random values in the interval [0, 1];

[0010] Step 3): based on the distance D=||c·W p (t)-W(t)|| between the grey wolf and the hunting object and the vector W(t+1)=W P (t)-gD, the Sigmoid function is used to update the convergence factor a, t is the current iteration number; different weights are determined according to the different positions of the top three wolves, and the updated position of the grey wolf is calculated by using the Euclidean distance formula, so that the penalty coefficient c and the hyperparameter g are optimized to obtain the optimal penalty coefficient c and the optimal hyperparameter g;

[0011] Step 4): the optimal penalty coefficient c and the hyperparameter g are substituted into the support vector machine model, the data of the training set is used for training, and the data of the test set is used for testing, and the predicted local deformation amount is output, and a proxy model is constructed;

[0012] Step 5): the simulation data, the predicted data of the proxy model and the artificial experience data are combined to form a multi-source database, and the multi-source database is expressed in the form of a knowledge graph;

[0013] Step 6): the processing method of the to-be-processed ship plate is determined, the plate unfolding method is used to obtain the crack of the to-be-processed ship plate, the local deformation amount of the crack is calculated according to the crack size, the flame path position of the crack is calculated according to the crack position, the processing parameters of the to-be-processed ship plate are preliminarily screened based on the knowledge graph, and the first processing is performed by using the preliminarily screened processing parameters;

[0014] Step 7) After the first processing, the local deformation error of the corresponding position between the processed ship plate and the set target ship plate is calculated, the flame port position is fine-tuned as the flame port position of the second processing, and the final flame port position is obtained by repeating the above steps;

[0015] Step 8) The processing parameters of the to-be-processed ship plate are multi-objective optimized by considering the weights of production time and production energy consumption, and the optimal processing parameters of the to-be-processed ship plate are obtained.

[0016] Step 9) The optimal processing parameters and the final flame port position are combined to form a processing scheme.

[0017] Further, in step 3), the update formula of the Sigmoid function update convergence factor a is: a begin The initial value of a is 2, a final The final value of a is 0.

[0018] Further, the weights E1, E2 and E3 of the three wolves in front are respectively The updated position of the gray wolf is T is the maximum number of iterations.

[0019] Further, in step 9), the processing scheme is provided to the operator, the processing scheme is judged and corrected by manual judgment to form a final processing scheme, and the final processing scheme is stored in the knowledge graph.

[0020] The technical effects of the above technical scheme adopted by the present application are as follows:

[0021] 1) The present application can quickly accumulate processing data, form a multi-source database and a knowledge graph combined with artificial experience, use the plate unfolding method combined with knowledge graph reasoning to determine the best heat source processing parameters and flame port position for one-time processing, calculate the difference between the curvature and angular deformation of the processed steel plate and the target plate, calculate the secondary processing heating parameters, and use the flame port approximation to infer the secondary processing flame port position for secondary processing, and repeat the above steps until the steel plate meets the requirements, thereby improving the accuracy of ship plate processing decision.

[0022] 2) The present application is based on artificial mixed enhanced intelligence in the human loop, the processing scheme determined by decision is adjusted by artificial adjustment, and the processing scheme is corrected by artificial cognition, which is more consistent with the actual processing scene. And the correct processing scheme is returned to the database for next time calling, realizing the updating and improvement of the rule knowledge.

[0023] 3. Compared with the method of patent application publication number 114091304A, the present invention uses a surrogate model as the prediction model, which takes into account the actual distribution of sample points and selects sample points uniformly from the sample space that conforms to the actual situation. Such sample points are more realistic.

[0024] 4. Compared with the method of patent application publication number 114091304A, the present invention updates the position by using the weights of different wolves in the gray wolf algorithm, improves the convergence factor, and enhances the global search capability of the algorithm.

[0025] 5. Compared with the method of patent application publication number CN102682139A, the present invention uses the plate unfolding method to provide a detailed explanation of the reasoning and processing scheme, which improves the efficiency of ship plate processing and the accuracy of decision-making.

[0026] 6. This invention breaks away from the previous reliance on worker experience and introduces hybrid augmented intelligence. It proposes a method that combines the digitization of human experience with computer intelligent algorithms to automatically generate processing plans based on processing requirements. This is of great significance for the establishment of an intelligent decision support system for the automatic generation of hull plate processing plans. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and specific implementation methods.

[0028] Figure 1 This is a flowchart illustrating the intelligent decision-making method for hull plate surface forming based on artificial hybrid enhanced intelligence of the present invention.

[0029] Figure 2 for Figure 1 A schematic diagram of the input and output labels for finite element simulation.

[0030] Figure 3 for Figure 1 A schematic diagram illustrating the construction process of the proxy model in China;

[0031] Figure 4 for Figure 1 A schematic diagram illustrating the construction of a knowledge graph in China;

[0032] Figure 5 for Figure 1 A schematic diagram approximating the location of the middle flame channel. Detailed Implementation

[0033] like Figure 1As shown, the present application is based on a mixed enhanced intelligent hull plate curved surface forming intelligent decision method. Sample points obtained from test hull plates are taken as input. Simulation deformation of the hull plate is obtained through ANSYS finite element simulation. A proxy model is constructed to predict the deformation and quickly accumulate processing data. The predicted data is combined with artificial experience to form a multi-source database. The multi-source database is analyzed to form a knowledge graph. The target plate is unfolded using the plate unfolding method. The optimal heat source processing parameters and flame path position are inferred combined with the knowledge graph. Once processing is performed. The error between the processed plate and the target plate is calculated. The flame path approach is used to infer the secondary processing flame path position. Secondary processing is performed. This is repeated until the plate meets the requirements. Finally, based on mixed enhanced intelligence, the processing scheme is corrected combined with artificial cognition to form a correct processing scheme. The correct processing scheme is added to the knowledge graph for storage for next time. The specific steps are as follows:

[0034] S1: Through ANSYS finite element analysis software, the deformation process of the test hull plate is simulated to obtain hull plate deformation data corresponding to different processing parameters and flame path positions. According to different processing parameters and flame path positions and corresponding hull plate deformation data, the relationship between processing parameters and local deformation is established, as shown in Figure 2 . Wherein, different processing parameters include hydrogen flow Q, heating speed V, heat source efficiency η and heating radius b, and the flame path position is L, and the corresponding hull plate deformation data is simulation local deformation X1.

[0035] In the sample space of different processing parameters (hydrogen flow Q, heating speed V, heat source efficiency η and heating radius b, and flame path position L), sampling is performed using a conventional Latin hypercube test to obtain sample points. In this way, the data of the entire sample space can be considered, the range of processing parameters is expanded, and the accuracy of the prediction result is improved. The processing parameter data represented by these sample points (hydrogen flow Q1, heating speed V1, heat source efficiency η1, heating radius b1, flame path position L1) are taken as input. Through finite element simulation, the simulation local deformation X1 is output. These input and output data constitute the simulation data set (Q1, V1, η1, b1, L1, X1).

[0036] S2: As shown in Figure 3 , an improved grey wolf algorithm optimized support vector machine (IGWO-SVM) is used as a proxy model. This proxy model can quickly predict the deformation of different input parameters, saving time for obtaining deformation data. Specifically:

[0037] S2-1: The data set obtained in step S1 is divided into a training set and a test set in a ratio of 4:1, and normalized, and a support vector machine (SVM) model is constructed as follows:

[0038] f(x) = w*g*Phi(x) + c (1)

[0039] In the formula, w represents a weight vector, g represents a hyperparameter, Phi(x) is a mapping function, the role of Phi(x) is to map each data of the training set from a low dimension to a high dimensional feature space, and c represents a penalty coefficient.

[0040] S2-2: Initialize the wolf group and parameters a, g and c, set the normal wolf group number value and the maximum iteration number T, wherein the mathematical formula of the grey wolf algorithm is as follows,

[0041] D = ||c*W p (t) - W(t) || (2)

[0042] W(t+1) = W P (t) - gD (3)

[0043] g = 2a*r1-a (4)

[0044] c = 2*ar2 (5)

[0045] In the formula, D is the distance between the grey wolf and the hunting object; W P and W represent the vector of the current position of the pth grey wolf and the hunting object after the optimization algorithm is iterated t times; a represents a convergence factor; g is a disturbance factor, ranging from [-a, a]; c is a disturbance factor, which is a random number; r1 and r2 are two random values between 0 and 1.

[0046] The present application adopts the Sigmoid function structure to update the convergence factor a instead of the traditional algorithm from 2 to 0, which can guarantee the global optimization ability and enhance the local solution ability. The update formula is as follows:

[0047]

[0048] In the formula, a begin represents the initial value 2 of a, a final represents the terminal value 0 of a, and t is the current iteration number.

[0049] S2-3: The optimization number t is initially 0, and alpha wolf, beta wolf and delta wolf are defined as the three wolves with the initial position in the front. When calculating the position of omega wolf, the present application determines different weights of the three wolves according to their different positions, and introduces a dynamic strategy of Euclidean distance to calculate the position update formula of the grey wolf. The weight calculation and the position update formula are as follows:

[0050]

[0051] In the formula, E1, E2 and E3 represent the weight coefficients of alpha wolf, beta wolf and delta wolf, respectively, represent the positions of α wolf, β wolf, δ wolf respectively, W (t+1) is the updated ω wolf position.

[0052] S2-4: update parameters a, g, c, calculate the fitness of the wolf pack by a conventional algorithm and update the positions of the top three wolves in S2-3, and optimize the penalty coefficient c and the hyperparameter g in formula (1). Then go to the next iteration until the iteration is completed, and output the optimal target parameters.

[0053] S2-5: bring the optimal parameters c and g into formula (1), first train with the data of the training set, and then test with the data of the test set, the input is hydrogen flow Q2, heating speed V2, heat source efficiency η2, heating radius b2 and flame position L2, the output is predicted local deformation X2, obtain the prediction result, and construct a proxy model.

[0054] S3: organize the data of hydrogen flow Q1, heating speed V1, heat source efficiency η1, heating radius b1, flame position L1 and the simulated local deformation X1 of the flame of the finite element simulation in step S1, and the predicted data of hydrogen flow Q2, heating speed V2, heat source efficiency η2, heating radius b2, flame position L2 and the predicted local deformation X2 of the flame of the proxy model, to form an organized simulation model data set. Combine this simulation model data set with artificial experience data to form a multi-source database. Express the multi-source database in the form of a knowledge graph, specifically:

[0055] First, according to the collected knowledge data, determine the structure of the knowledge graph, including the definition of entities, relationships and attributes. Among them, the entities include sail-shaped plates, saddle-shaped plates, and plates to be processed, the relationships include plate type, plate thickness, processing parameters, etc., and the attributes include angular deformation, linear deformation, etc. Therefore, the relationship between each parameter quantity can be represented and saved by establishing a knowledge graph as shown in Figure 4 Figure 4 ​It can be seen that the center of the atlas is the ship plate. First, according to the plate type, the sail-shaped plate, saddle-shaped plate, twisted plate, Long-shaped plate and conical plate are divided into categories. Further, according to the plate thickness, thick plate and thin plate are distinguished, and then the processing method is determined according to artificial experience, such as heating method, cooling method, etc. Taking the sail-shaped plate as an example: if the plate thickness is less than 15mm, straight-line heating and front water cooling are adopted, and if it is greater than 15mm, spiral heating and back water cooling are adopted. Further, the last arrow points to the local deformation amount X of the ship plate under various processing parameters (hydrogen flow Q, heating speed V, heat source efficiency η, heating radius b, flame path position L). The nodes of the atlas are connected by arrows to represent the relationship between the two. Therefore, by using the layer-by-layer lookup method, the target value can be quickly found in the atlas as the control output, or the value can be modified; and for the increase of data, only the new branch needs to be added behind the corresponding node. In this way, the acquired knowledge is expressed and stored in categories, so as to be called when the processing scheme is decided next time.

[0056] S4: For the ship plate to be processed, first determine the processing method according to the plate type, plate thickness and other factors of the ship plate to be processed, then use the plate unfolding method to obtain the crack and its position, deduce the flame path position L according to the crack position, and inversely deduce the deformation amount according to the crack size, and obtain the processing parameters (hydrogen flow Q, heating speed V, heat source efficiency η, heating radius b, flame path position L) based on the knowledge atlas.

[0057] When the plate unfolding method is used to unfold the ship plate to be processed, the crack position is taken as the flame path position L, and the size of the local deformation amount X of the flame path position L can be calculated by using the conventional algorithm according to the crack size. Then, according to the value of the local deformation amount X, the processing parameters are preliminarily screened from the knowledge atlas, a plurality of sets of processing parameters (Q, V, η, b) are obtained, and the final processing parameters determined by the conventional white whale algorithm are used for the first processing.

[0058] The steps of the plate unfolding method are as follows:

[0059] 1) Select non-uniform B-spline to represent the curved surface shape of the ship plate;

[0060] 2) Determine the unfolding center and the reference unit strip after unfolding of the ship plate;

[0061] 3) According to the different curvatures of the curved surface shapes of different regions of the ship plate, the ship plate is divided into a plurality of unit quadrilaterals by using an adaptive method;

[0062] 4) Triangular mesh is unfolded for each unit quadrilateral;

[0063] 5) : all the triangular meshes are encrypted to determine whether to converge. If it converges, the unfolding ends; if it diverges, the quadrilateral where the triangular mesh is located is encrypted and divided. Repeat this operation on the triangular mesh that does not converge until the gap error after unfolding is less than the set ideal value.

[0064] According to the multi-objective optimization of the processing parameters (Q, V, η, b) preliminarily screened from the knowledge graph, the production time and the production energy consumption of the ship plate curved surface forming are reduced as the optimization objectives, and the optimal processing parameters are found. From the actual situation, the production time and the production energy consumption need to be considered at the same time, so the weight coefficients of the production time and the production energy consumption functions are both determined as 0.5. The conventional white whale algorithm is used to optimize the objective function. At this time, the multi-objective optimization problem can be represented as:

[0065]

[0066] In the formula: X represents a group of processing parameters (Q, V, η, b, L); T * (X) and E * (X) respectively represent the time function value and the energy consumption function value; T(X) and E(X) respectively represent the minimum time processing parameter group and the minimum energy consumption processing parameter group; Optimal(X) represents the optimal processing parameter, that is, the objective function.

[0067] S5: The target ship plate is the expected forming effect of the ship plate processing. After the first processing, the difference between the local deformation of the processed ship plate and the corresponding position of the set target ship plate is calculated, and the calculation formula is as follows.

[0068]

[0069] Referring to Figure 5 As the ship plate cannot be heated at the same point for multiple times, the flame path position of the first processing is taken as the flame path reference line B. The flame path position is fine-tuned, and parallel lines are made at a certain distance near the flame path of the first processing, that is, the flame path approximation line A, as shown by the dotted line in Figure 5 The flame path approximation line A is taken as the flame path position for the second processing. Repeat this process until the ship plate meets the requirements and the final flame path position is obtained.

[0070] Although the crack position after unfolding can calculate the parameters and the flame path position of the first processing, it is often difficult to achieve the requirements in one processing in actual processing. The advantage of this flame path approximation method which simulates the actual processing process is that it can simulate the manual processing process and make the deformation of the ship plate approach the target ship plate.

[0071] S6: The optimal processing parameters obtained in S and the final flame path position obtained after the cycle in S5 are combined to become a processing scheme.

[0072] S7: According to the method of mixed enhanced intelligence, the processing scheme obtained in S6 is provided to the operator, the processing scheme is judged manually, and the processing scheme is corrected to form a final processing scheme as a final decision result, thereby enhancing the accuracy of the intelligent decision method.

[0073] S8: Finally, the final processing scheme is stored in the knowledge graph for later calling.

Claims

1. A smart decision-making method for hull plate surface forming based on hybrid augmented intelligence, characterized by the following steps: step 1): Finite element simulation was performed on the outer plating of the test vessel to obtain the simulated local deformation based on different processing parameters and flame channel positions, forming simulation data. Step 2): Divide the simulation data into training and testing sets, and construct a support vector machine model after normalization. hyperparameters Penalty coefficient , For weight vectors, For mapping functions, The value is randomly generated between two [0, 1]; Initialize the wolf pack and parameters a, g, and c, and set the wolf pack size and maximum number of iterations. The mathematical formula for the Grey Wolf algorithm is as follows: This refers to the distance between the gray wolf and its prey. They represent the iterations of the optimization algorithm. The next The vectors of the current positions of the gray wolf and its prey; Represents the convergence factor; As an interference factor, the range is ; The interference factor is a random number; The value is randomly generated between two [0, 1]; Step 3): Update the convergence factor using the Sigmoid function structure. The updated formula is: represent The initial value is 2. represent The termination value is 0, and t is the current iteration number; Number of optimization attempts The initial value is 0, defined Wolf, Wolf, The wolves are initially positioned as the first three. Different weights are assigned based on the positions of these three wolves. The updated position of the gray wolf is calculated using Euclidean distance. The position update formula is: Represent Wolf, Wolf, The weighting coefficient of wolves, , , Represent Wolf, Wolf, The wolf's position For the updated Wolf's position; Update parameter convergence factor Hyperparameters Penalty coefficient Calculate the wolf pack fitness and update the positions of the first three wolves. Optimize the penalty coefficient c and hyperparameter g in the support vector machine model. Proceed to the next iteration until the iteration is complete and output the optimal target parameters. Step 4): Substitute the optimal penalty coefficient c and hyperparameter g into the support vector machine model formula, first train the model using the training set data, and then test it using the test set data to construct the surrogate model. Step 5): Combine the simulation data, the prediction data of the proxy model, and the human experience data to form a multi-source database, and express the multi-source database in the form of a knowledge graph. Step 6): Determine the processing method of the ship plate to be processed, use the plate unfolding method to obtain the cracks of the ship plate to be processed, deduce its local deformation based on the crack size, deduce its flame channel position based on the crack position, preliminarily screen the processing parameters of the ship plate to be processed based on the knowledge graph, and perform the first processing using the preliminarily screened processing parameters. Step 7): After the first processing, calculate the local deformation error between the processed ship plate and the set target ship plate at the corresponding position, and fine-tune the flame channel position as the flame channel position for the second processing. Repeat this process to obtain the final flame channel position. Step 8): Taking into account the weights of production time and production energy consumption, perform multi-objective optimization of the processing parameters of the ship plate to be processed to obtain the optimal processing parameters of the ship plate to be processed. Step 9): Combine the optimal processing parameters and the final flame channel position to form a processing scheme.

2. The intelligent decision-making method for hull plate curved surface forming according to any one of claims 1, characterized in that: In step 9), the processing plan is provided to the operator, who then manually judges and modifies the processing plan to form the final processing plan. The final processing plan is then stored in the knowledge graph.

3. The intelligent decision-making method for hull plate curved surface forming according to any one of claims 1, characterized in that: The knowledge graph is centered on a ship plate, and its structure includes the definition of entities, relationships, and attributes. The entities include sail-shaped plates, saddle-shaped plates, and plates to be processed. The relationships include plate type, plate thickness, and processing parameters. The attributes include angular deformation and line deformation.

4. The intelligent decision-making method for hull plate curved surface forming according to any one of claims 1, characterized in that: The plate unfolding method is as follows: First, a non-uniform B-spline is selected to characterize the surface shape of the ship plate, and the unfolding center and the reference unit strip after unfolding are determined; then, according to the different curvatures of the surface shape in different regions of the ship plate, the ship plate is divided into several unit quadrilaterals, and triangular mesh unfolding is performed on each unit quadrilateral; finally, all triangular meshes are refined to determine whether they converge. If they converge, the unfolding ends; if they diverge, the quadrilateral containing the triangular mesh is further refined until the gap error after unfolding is less than the set ideal value.

5. The intelligent decision-making method for hull plate curved surface forming according to claim 1, characterized in that: The plate unfolding method is as follows: In step 8), the multi-objective optimization function is: , Represents a set of processing parameters. , These represent the time function value and the energy consumption function value, respectively. , These represent the minimum time processing parameter group and the minimum energy consumption processing parameter group, respectively.

6. The intelligent decision-making method for hull plate curved surface forming according to claim 1, characterized in that: The steps described in step 7) .

7. The intelligent decision-making method for hull plate curved surface forming according to claim 1, characterized in that: The processing parameters include hydrogen flow rate, heating rate, heat source efficiency, and heating radius.

8. The intelligent decision-making method for hull plate curved surface forming according to claim 1, characterized in that: In step 6), multiple sets of processing parameters are initially selected based on the knowledge graph, and then the White Whale algorithm is used to determine the final processing parameters for the first processing.

Images