A progressive prediction method for the cross-sectional profile of the entire bend of metal pipes based on Bo-LSTM

Abstract

The present invention discloses a method for progressively predicting the cross-sectional profile of a full bend section of a metal pipe fitting based on Bo-LSTM. The bending forming process parameters are sampled, simulation is performed according to the process parameters, cross-sections are extracted from the simulation result data, sequence representation is performed on the cross-sectional data of the full bend section, and the dimension of the process parameter features is expanded for the cross-sectional sequence representation; then, the hyperparameters of the Bo-LSTM network are initialized and the processed data is trained, the network hyperparameters are updated according to the RMSE on the validation set, and the final prediction model is iteratively generated; finally, the trained model is used to progressively predict the cross-sectional profile data of the full bend section during the forming process, and the cross-sectional profile of the full bend section of the metal pipe fitting at subsequent moments is obtained. The present invention can improve the prediction accuracy of the cross-sectional profile of the full bend section during the bending forming process of metal pipe fittings, and provide a basis for effectively intervening in the forming process and achieving high-precision forming.

Description

Technical Field

[0001] The present invention relates to the technical field of high-precision metal pipe bending and forming, and in particular to a method for progressively predicting the cross-sectional profile of a full-bend section of a metal pipe based on Bo-LSTM. Background Art

[0002] Metal pipes, with their hollow interior and circumferentially curved cross-section, are ideal for conveying fluids and are widely used in major projects such as aerospace. Defects such as cross-sectional flattening and wall thickness variation during the pipe bending process directly affect the cross-sectional geometry of the pipe, causing pressure loss and flow pulsation within the pipe. Accurate and comprehensive cross-sectional profile prediction allows for effective intervention in the bending process, reducing cross-sectional defects and improving forming accuracy.

[0003] Currently, traditional research on the prediction of cross-sectional defects—cross-sectional distortion and wall thickness variation—is primarily based on theoretical derivations based on assumptions and approximations. These methods are limited in their applicability, often only predicting the mid-section of a bend and severing the coupling between cross-sectional distortion and wall thickness variation. Therefore, a deep learning-based, accurate and comprehensive method for predicting the cross-sectional profile of metal pipes is needed. Summary of the Invention

[0004] In order to solve the problems existing in the background technology, the purpose of the present invention is to provide a method for progressive prediction of the cross-sectional profile of the entire bend section of a metal pipe based on Bo-LSTM, which aims to achieve accurate and comprehensive cross-sectional profile prediction in response to the problem of limited applicable conditions of traditional cross-sectional defect prediction.

[0005] This method can obtain a representation form of the cross-sectional profile of the entire bend section of the pipe fitting that is suitable for neural network input through the vector description of the cross-section and the sequence representation of the bend section. It can also accurately predict the cross-sectional profile of the entire bend section of the metal pipe fitting during the bending process through the sequence prediction of Bo-LSTM, overcoming the limitation of the applicable conditions of traditional prediction research on cross-sectional defects of metal pipe fittings. This allows us to effectively intervene in the bending process, reduce cross-sectional defects, and improve forming accuracy.

[0006] The technical solution adopted by the present invention comprises the following steps:

[0007] 1) Perform Latin hypercube sampling on the adjustable parameters in the metal pipe forming process to obtain a sampled process parameter table to ensure full coverage of each parameter range;

[0008] 2) According to the process parameter table in step 1, the forming process is simulated under the corresponding parameters for each set of adjustable parameters to obtain simulation result data of the entire bending section of the metal pipe;

[0009] 3) Perform cross-section extraction on the simulation result data, including node grouping, nearest neighbor node search, and obtaining cross-section contour points through interpolation method;

[0010] 4) Based on the vector description of the cross-sectional profile, the cross-sectional data of the entire bend is sequentially represented and converted into a form suitable for neural network input;

[0011] 5) By introducing process parameter features, the dimension of the sequence representation of the full-bend section cross-section data is expanded to obtain a cross-section profile vector description that integrates the process parameters;

[0012] 6) Design a deep learning network based on Bo-LSTM and initialize its hyperparameters. Train the network using the data processed in step 5. Update the network hyperparameters based on the root mean square error (RMSE) on the validation set, and iterate to generate the final prediction model.

[0013] 7) For the metal pipe to be tested, input the process parameters and the cross-sectional profile data of the entire bend at a certain bend angle, and use the model trained in step 6 to progressively predict the cross-sectional profile of the entire bend of the metal pipe at subsequent bend angles during the metal pipe forming process.

[0014] The adjustable parameters in step 1 include the outer diameter of the pipe, the wall thickness of the pipe, the bending radius, the friction coefficient between the pipe and the bending die, the friction coefficient between the pipe and the pressing die, the friction coefficient between the pipe and the anti-wrinkle die, the boosting speed, the starting position of the pressing die, and the bending speed.

[0015] In step 2, the simulation result data corresponding to each set of parameters includes simulation result data of bending angles from 0° to α at every set angle Δθ.

[0016] In step 3, the node coordinate information in the simulation result data is imported into Matlab for point cloud processing:

[0017] 3.1) Node grouping:

[0018] The nodes on the same busbar before the pipe is bent are divided into the same group. The nodes in the same group are arranged monotonically along the busbar according to the size of the serial number, because the node serial numbers generated by the simulation during mesh division are monotonically arranged along the busbar.

[0019] 3.2) Nearest neighbor node search:

[0020] For each group of nodes, all adjacent nodes are traversed in order according to the node sequence within the group. When the following inequality holds true, the corresponding adjacent nodes S1 and S2 are the nearest neighboring points A and B in the same group of nodes to the target section plane:

[0021] [tan(OS1)-tan(θ p )]*[tan(OS2)-tan(θ p)]<0

[0022] Among them, S1 and S2 are two adjacent nodes in the same group; tan(OS1) and tan(OS2) are the tangent values of the angles between OS1 and OS2 and the bending starting plane, respectively. p ) is the tangent value of the angle between the target section plane and the bending starting plane;

[0023] The target section plane is at an angle θ from the bending starting plane. p plane, θ p =p*Δθ, 0≤p≤n, n=α / Δθ, α is the angle between the bending start plane and the bending end plane, p is the section number, and n is the maximum section number of the entire bending section of the pipe fitting;

[0024] 3.3) Obtain the cross-sectional contour point Q by interpolation method:

[0025] OQ=(DB*OA+DA*OB) / AB

[0026] Among them, O is the bending center of the metal pipe, Q is the cross-sectional contour point of the metal pipe; A and B are the two points closest to the target cross-sectional plane in each group of nodes and are located on both sides of the plane; D is the intersection of the target cross-sectional plane and the line AB.

[0027] In step 4, the sequence of the full-bend section data of the inner and outer contours of the cross section is characterized as follows:

[0028] l T =(L1 T , L2 T ,...L N T )

[0029] Where T represents the current bending moment; l represents the entire bending segment, L represents the vector description of the cross section on the bending segment, and N represents the number of cross sections in the entire bending segment;

[0030] The vector description L of the cross-section profile on the curved segment is from each sampling point on the right half of the cross-section profile to the cross-section center O mid The multidimensional vector composed of the distances arranged in sequence;

[0031] Among them, the cross-section center O mid Take it as the midpoint between the leftmost and rightmost points of the cross-section profile;

[0032] The sampling points on the cross-section profile are centered on the cross-section center and are spaced according to the angles. Evenly distributed cross-sectional contour points, obtained through step 3.

[0033] In step 5, the feature dimension is expanded to add the adjustable parameters of step 1 to the multidimensional vector description L of the cross section to obtain the cross section profile vector description of the fusion process parameters.

[0034] In step 6:

[0035] The Bo-LSTM deep learning network includes prediction model A and prediction model B, both of which are composed of a sequence input layer, an LSTM layer, a fully connected layer, and a regression layer. The Bayesian optimization algorithm is used to obtain the optimal configuration of hyperparameters such as the number of hidden units in the LSTM layer, the number of neurons in the fully connected layer, the initialization learning rate, and the L2 regularization parameter.

[0036] The input of both prediction models is the cross-sectional profile vector description of the entire bending segment fused with process parameters at the current bending moment. The number of cross-sectional profiles on the entire bending segment input is N. The output of prediction model A is the cross-sectional profile vector description of the first N fused process parameters at the next bending moment. The output of prediction model B is the cross-sectional profile vector description of the last N fused process parameters at the next bending moment.

[0037] The final prediction result of the deep learning network is the combination of the outputs of the two prediction models, and the average of the overlapping data of the two prediction models' outputs is taken.

[0038] In step 7:

[0039] The progressive prediction mode predicts the cross-sectional profile of the entire bend section of the next corner based on the cross-sectional profile data of the entire bend section of the previous corner, and then corrects the predicted value according to actual feedback as input data for subsequent predictions, thereby predicting the cross-sectional profile of the entire bend section of the subsequent corner.

[0040] The beneficial effects of the present invention are:

[0041] The vector description of the cross section and the sequence characterization of the bending segment proposed by the method of the present invention can accurately and comprehensively characterize the cross-sectional profile characteristics of the entire bending segment, realize the unified characterization of the two cross-sectional defects of cross-sectional distortion and wall thickness variation, and convert the cross-sectional profile data into a form suitable for neural network input.

[0042] The method of the present invention realizes sequence prediction of the cross section of the entire bend section through a deep learning network based on LSTM, and adopts a Bayesian optimization algorithm to obtain the optimal configuration of hyperparameters such as the number of hidden units in the LSTM layer, the number of neurons in the fully connected layer, the initialization learning rate, and the L2 regularization parameter, thereby improving the prediction accuracy.

[0043] The method of the present invention adopts a progressive prediction mode, amends the prediction result according to real-time feedback data, and then uses the amended prediction result as input for the next prediction, thereby ensuring the robustness of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of the present invention;

[0045] Figure 2 It is a cross-sectional sequence construction of the entire bending segment during the bending process of the present invention.

[0046] Figure 3 It is a vector description of the inner and outer contours of the cross section of the present invention.

[0047] Figure 4 Schematic diagram of the deep learning network model structure implemented in the present invention. DETAILED DESCRIPTION

[0048] The present invention will be described in further detail below with reference to the accompanying drawings. Figure 1 As shown, the present invention includes the following steps:

[0049] 1. Latin hypercube sampling is performed on the adjustable parameters in the metal pipe forming process. Latin hypercube sampling can ensure full coverage of each variable range by maximizing the stratification of each edge distribution.

[0050] The adjustable parameters involved in CNC mandrel-free rotary bending include the outer diameter D of the pipe, the wall thickness T of the pipe, the bending radius R, the friction coefficient f between the pipe and the bending die b , the friction coefficient f between the pipe fitting and the die p , the friction coefficient f between the pipe fitting and the anti-wrinkle mold w , boost speed V p , Die starting position L p0 , bending speed w. The header of the process parameter table is shown in Table 1:

[0051] Table 1 Process parameter table header

[0052]

[0053]

[0054] Among them, V r It is the linear speed of metal pipe feeding.

[0055] 2. According to the sampled process parameter table, Abaqus 2016 / Explicit was used to simulate the forming process under each group of corresponding parameters to obtain the simulation result data of the full bending section of the metal pipe.

[0056] The simulation result data includes simulation result data of bending angles ranging from 0° to α at regular intervals of angles Δθ.

[0057] 3. Perform cross-section extraction on the simulation result data and import the node coordinate information in the simulation result data into Matlab for point cloud processing, including node grouping, nearest neighbor node search and obtaining cross-section contour points through interpolation method.

[0058] The interpolation method uses linear interpolation to determine Q, and further uses inverse distance interpolation to correct it to Q′. The calculation formula is:

[0059] OQ′=(QB*OA+QA*OB) / AB

[0060] Where O is the bending center of the metal pipe, Q′ is the cross-sectional contour point of the metal pipe; A and B are the two points closest to the target cross-sectional plane in each group of nodes and are located on both sides of the plane; Q is the intersection of the target cross-sectional plane and the line AB.

[0061] A and B are the two nearest neighbor points in each group of nodes to the target cross-section plane. The nearest neighbor node search is calculated by the following formula:

[0062] [tan(OP1)-tan(θ i )]*[tan(OP2)-tan(θ i )]<0

[0063] Among them, P1 and P2 are two adjacent nodes in the same group; tan(OP1) and tan(OP2) are the tangent values of the angles between OP1 and OP2 and the bending starting plane, respectively. i ) is the tangent value of the angle between the target section plane and the bending start plane. Nodes are grouped based on whether they are on the same generatrix before bending.

[0064] 4. Based on the vector description of the cross-sectional profile, the cross-sectional profile is represented as a multidimensional vector. Then, the evenly spaced cross-sections on the curved segment are represented by vectors in sequence and arranged into a sequence, thereby converting it into a form suitable for neural network input.

[0065] like Figure 2 As shown in Figure 2, the sequence of cross-section data for the entire curved section can be characterized as follows:

[0066] l T =(L1 T , L2 T ,...L N T )

[0067] Among them, T represents the current bending moment; l represents the entire bending segment, L represents the vector description of the cross section on the bending segment, and N represents the number of cross sections in the entire bending segment.

[0068] like Figure 3As shown in the figure, L represents the vector description of the cross-section profile on the curved segment, which is a multidimensional vector composed of the distances from each sampling point on the right half of the cross-section profile to the cross-section center arranged in sequence. The distance L from the sampling point P to the cross-section center is OP The center of the section is taken as the midpoint between the leftmost and rightmost points of the section profile. The sampling points are centered on the section center and are spaced according to the angle. The sampling points are obtained by performing similar interpolation on the cross-sectional contour points in step 3. The sequence representation of the full-bend cross-sectional data of the inner contour and outer contour is obtained respectively according to the above process.

[0069] 5. Introduce the 9 process parameter features described in step 1, expand the dimension of the cross-section sequence representation, and introduce 9 features to record the process parameters of the forming process.

[0070] Feature dimension expansion adds process parameter dimension to the multidimensional vector description L of the cross section to record process parameter information and obtain the cross section profile vector description of the fusion process parameters

[0071] 6. Design a deep learning network based on Bo-LSTM and initialize its hyperparameters. Train the network on the data processed in step 5. Update the network hyperparameters based on the RMSE on the validation set, and iterate to generate the final prediction model.

[0072] like Figure 4 As shown in the figure, the Bo-LSTM deep learning network includes two networks, A and B, both of which are composed of a sequence input layer, an LSTM layer, a fully connected layer, and a regression layer. The Bayesian optimization algorithm is used to obtain the optimal configuration of hyperparameters such as the number of hidden units in the LSTM layer, the number of neurons in the fully connected layer, the initialization learning rate, and the L2 regularization parameter.

[0073] Prediction model A is used to predict the first N cross-sectional profiles at the next moment from the N cross-sectional profiles at the previous moment, as shown in the following formula:

[0074] (L1 T+1 , L2 T+1 ,...L N T+1 )=f(l T )=f(L1 T , L2 T ,...L N T )

[0075] Prediction model B is used to predict the next N cross-sectional profiles at the next moment from the N cross-sectional profiles at the previous moment, as shown in the following formula:

[0076] (L2 T+1 , L3 T+1 ,…L N+1T+1 )=g(l T )=g(L1 T , L2 T ,...L N T )

[0077] The comprehensive prediction models A and B obtain the predicted values of the cross-sectional profile sequence of the entire bend at the next moment (taking the average of the overlapping cross-sectional parts), calculate the root mean square error (RMSE) on the validation set, and perform Bayesian optimization to update the network hyperparameters until the termination condition of the iterative optimization is met, that is, the optimization time or the number of optimization iterations reaches the set value.

[0078] 7. Input the process parameters and the cross-sectional profile data of the entire bend at a certain bend angle, and use the trained model to progressively predict the cross-sectional profile of the entire bend section of the metal pipe at subsequent bend angles during the forming process.

[0079] The full cross-sectional profile of the subsequent bend is predicted based on the full cross-sectional profile data of the previous bend. This cross-sectional profile is obtained by combining the prediction results of prediction models A and B. The progressive prediction mode means that the predicted value is then modified based on actual feedback and used as input data for subsequent predictions. The full cross-sectional profile of the subsequent bend is then predicted until the predicted bend angle reaches the bending end angle, which means the process is complete.

[0080] The prediction accuracy measurement indicator is as follows:

[0081]

[0082]

[0083] Where q represents the number of sequence samples in the data set, N represents the number of sections in the sequence sample, and M represents the number of sampling points on the section; L xy It represents the actual value of the distance from the yth sampling point on the xth section in the curved segment to its section center, and Indicates the corresponding predicted value; Sample s represents the sth real cross-sectional sequence data in the data set, Represents the corresponding cross-sectional series data for the forecast.

[0084] Table 2 shows the prediction accuracy of the method proposed in this invention on the test set and validation set. Outer A and Outer B represent the outer contour prediction models A and B respectively, and Inner A and Inner B represent the inner contour prediction models A and B respectively. test 、SPE val denote the sequence prediction errors on the test set and validation set, respectively.

[0085] Comprehensive prediction model A, B, the final prediction error for the entire curve segment, the SPE of the outer contour test The average and maximum values are 0.6% and 2.6% respectively; the SPE of the inner contour test The average value and the maximum value are 0.83% and 2.97% respectively. It can be seen that the method proposed in the present invention has a high accuracy in predicting the cross-sectional profile of the entire bend segment.

[0086] Table 2 Prediction accuracy of the method proposed in this paper on the test set and validation set

[0087]

[0088] Table 3 compares the proposed method with other methods, showing the prediction performance of a standard LSTM, Bo-LSTM, a standard BiLSTM, Bo-BiLSTM, an LSTM with a dropout layer, and a Bo-LSTM with a dropout layer.

[0089] By comparison, it can be seen that the method proposed in the present invention is superior to other methods and has higher prediction accuracy for the entire bend segment.

[0090] Table 3 Comparison results of the method proposed in the present invention with other methods

[0091]

[0092] Table 4 compares the results of the progressive prediction model proposed in this paper with the chain prediction model. The progressive prediction model uses real-time feedback data to correct the previous prediction result before using it as input for the next prediction. Chain prediction, on the other hand, uses the previous prediction result directly as input for the next prediction without correction.

[0093] By comparison, it can be seen that the chain prediction model proposed in the present invention has better robustness for the prediction of the entire curve segment.

[0094] Table 4 Comparison results between the progressive prediction model proposed by the present invention and the chain prediction model

[0095]

Claims

1. A method for progressively predicting the cross-sectional profile of a metal pipe in a full bend section based on Bo-LSTM, characterized in that: The following steps are involved: 1) Perform Latin hypercube sampling on the adjustable parameters in the metal pipe forming process to obtain a sampled process parameter table; 2) According to the process parameter table in step 1, the forming process is simulated under the corresponding parameters for each set of adjustable parameters to obtain simulation result data of the entire bending section of the metal pipe; 3) Perform cross-section extraction on the simulation result data, including node grouping, nearest neighbor node search, and obtaining cross-section contour points through interpolation method; 4) Based on the vector description of the cross-sectional profile, the cross-sectional data of the entire bend is sequentially represented and converted into a form suitable for neural network input; 5) By introducing process parameter features, the dimension of the sequence representation of the full-bend section cross-section data is expanded to obtain a cross-section profile vector description that integrates the process parameters; 6) Design a deep learning network based on Bo-LSTM and initialize its hyperparameters. Train the network using the data processed in step 5. Update the network hyperparameters based on the root mean square error (RMSE) on the validation set, and iterate to generate the final prediction model. 7) For the metal pipe to be tested, input the process parameters and the cross-sectional profile data of the entire bend at a certain bend angle, and use the model trained in step 6 to progressively predict the cross-sectional profile of the entire bend of the metal pipe at subsequent bend angles during the metal pipe forming process.

2. The method for progressively predicting the cross-sectional profile of a metal pipe according to claim 1, which is based on Bo-LSTM, is characterized in that: The adjustable parameters in step 1 include the outer diameter of the pipe, the wall thickness of the pipe, the bending radius, the friction coefficient between the pipe and the bending die, the friction coefficient between the pipe and the pressing die, the friction coefficient between the pipe and the anti-wrinkle die, the boosting speed, the starting position of the pressing die, and the bending speed.

3. The method for progressively predicting the cross-sectional profile of a metal pipe according to claim 1, characterized in that: In step 2, the simulation result data corresponding to each set of parameters includes simulation result data of bending angles from 0° to α at every set angle Δθ.

4. The method for progressively predicting the cross-sectional profile of a metal pipe according to claim 1, characterized in that: In step 3, the node coordinate information in the simulation result data is imported into Matlab for point cloud processing: 3.1) Node grouping: The nodes on the same busbar before the pipe is bent are divided into the same group, and the nodes in the same group are arranged monotonically along the busbar according to the size of the sequence number; 3.2) Nearest neighbor node search: For each group of nodes, all adjacent nodes are traversed in order according to the node sequence within the group. When the following inequality holds true, the corresponding adjacent nodes S1 and S2 are the nearest neighboring points A and B in the same group of nodes to the target section plane: [time(OS1)-time(θ p )]*[tan(OS2)-tan(θ p )]:0 Among them, S1 and S2 are two adjacent nodes in the same group; tan(OS1) and tan(OS2) are the tangent values of the angles between OS1 and OS2 and the bending starting plane, respectively. p ) is the tangent value of the angle between the target section plane and the bending starting plane; The target section plane is at an angle θ from the bending starting plane. p plane, θ p =p*Δθ, 0≤p≤n, n=α / Δθ, α is the angle between the bending start plane and the bending end plane; 3.3) Obtain the cross-sectional contour point Q by interpolation method: OQ=(DB*OA+DA*OB) / AB Among them, O is the bending center of the metal pipe, Q is the cross-sectional contour point of the metal pipe; A and B are the two points closest to the target cross-sectional plane in each group of nodes and are located on both sides of the plane; D is the intersection of the target cross-sectional plane and the line AB.

5. The method for progressively predicting the cross-sectional profile of a metal pipe according to claim 1, which is based on Bo-LSTM, is characterized in that: In step 4, the sequence of the full-bend section data of the inner and outer contours of the cross section is characterized as follows: l T =(L1 T ,L2 T ,...L N T ) Where T represents the current bending moment; l represents the entire bending segment, L represents the vector description of the cross section on the bending segment, and N represents the number of cross sections in the entire bending segment; The vector description L of the cross-section profile on the curved segment is from each sampling point on the right half of the cross-section profile to the cross-section center O mid The multidimensional vector composed of the distances arranged in sequence; Among them, the cross-section center O mid Take it as the midpoint between the leftmost and rightmost points of the cross-section profile; The sampling points on the cross-section profile are centered on the cross-section center and are spaced according to the angles. Evenly distributed cross-sectional contour points, obtained through step 3.

6. The method for progressively predicting the cross-sectional profile of a metal pipe according to claim 1, based on Bo-LSTM, is characterized in that: In step 5, the feature dimension is expanded by adding the adjustable parameters of step 1 to the multi-dimensional vector description L of the cross section to obtain the cross-sectional profile vector description L of the fusion process parameters.

7. The method for progressively predicting the cross-sectional profile of a metal pipe according to claim 1, based on Bo-LSTM, is characterized in that: In step 6: The Bo-LSTM deep learning network includes prediction model A and prediction model B, both of which are composed of a sequence input layer, an LSTM layer, a fully connected layer, and a regression layer. The Bayesian optimization algorithm is used to obtain the optimal configuration of the number of hidden units in the LSTM layer, the number of neurons in the fully connected layer, the initialization learning rate, and the L2 regularization parameter. The input of both prediction models is the cross-sectional profile vector description of the entire bending segment fused with process parameters at the current bending moment. The number of cross-sectional profiles on the entire bending segment input is N. The output of prediction model A is the cross-sectional profile vector description of the first N fused process parameters at the next bending moment. The output of prediction model B is the cross-sectional profile vector description of the last N fused process parameters at the next bending moment. The final prediction result of the deep learning network is the combination of the outputs of the two prediction models, and the average of the overlapping data of the two prediction models' outputs is taken.

8. The method for progressively predicting the cross-sectional profile of a metal pipe according to claim 1, based on Bo-LSTM, is characterized in that: In step 7: The progressive prediction mode predicts the cross-sectional profile of the entire bend section of the next corner based on the cross-sectional profile data of the entire bend section of the previous corner, and then corrects the predicted value according to actual feedback as input data for subsequent predictions, thereby predicting the cross-sectional profile of the entire bend section of the subsequent corner.

Images