Dynamic Compensation Control Method for Ship Section Welding Deformation Based on Error Accumulation Principle

Abstract

The present invention discloses a dynamic compensation control method for ship section welding deformation based on the principle of error accumulation, comprising the following steps: collecting multi-pass welding process parameters during the ship section welding process and the corresponding welding deformation amount for each parameter; based on the principle of error accumulation of welding processes and using a deep neural network, constructing a layer-by-layer cumulative deformation error compensation model for each process of the section welded part; obtaining compensation data by using the compensation model, and through the model error compensation method, continuously and dynamically adjusting the process parameters of each welding process in a cumulative compensation manner to obtain the optimal welding process parameters. The present invention can effectively perform dynamic parameter adjustment for the whole process of ship section welding, realize the compensation of welding errors, obtain the optimal welding process parameters, and effectively control the final deformation of the welded part.

Description

Technical Field

[0001] The present invention belongs to the field of ship section welding, and particularly relates to a dynamic compensation control method for ship section welding deformation based on the error accumulation principle. Background Art

[0002] The forming process of weld seams is affected by various factors such as environment, equipment, process, and materials. Among them, the influence of process parameters is the most obvious, and the parameters will interact with each other, thus affecting the geometric dimensions of the weld seams. For the optimization of welding process parameters, the traditional method is to use DOE (Design of Experiment). With the development of technologies such as neural networks, artificial intelligence, and big data, they have gradually been applied to the establishment of welding relationship models. The combination with computer technology provides new methods and ideas for studying the relationship between the two.

[0003] Due to the non-linear characteristics of the artificial neural network (ANN), it can centrally process a large amount of complex information. In recent years, it has been widely applied to aspects such as welding joint performance prediction, welding process parameter optimization, weld seam forming prediction, weld seam tracking, and welding defect detection. Manikya et al. established a relationship model between welding process parameters (wire feeding speed, pulse frequency, plate thickness, and peak current) and weld seam geometric forming in GMAW welding through a BP neural network, and used this model to guide the combination of process parameters in design, so as to obtain the expected ratio of weld width to penetration depth. Parikshit et al. established weld seam forming models for TIG welding based on regression analysis method, BP neural network, and genetic algorithm respectively. By comparison, it was found that the models established by the latter two methods are more excellent than the former in terms of performance and accuracy.

[0004] However, the above method of optimizing welding process parameters through neural network methods only starts from a single welding process. For welded parts with many weld seams and complex welding processes, it will increase the difficulty of optimization and it is difficult to achieve an ideal effect. Summary of the Invention

[0005] Object of the Invention: In order to overcome the problems such as the inevitable irregular deformation of welded parts during the welding process, a dynamic compensation control method for ship section welding deformation based on the error accumulation principle is provided. The deep learning algorithm is used to learn the influence of each welding process on welding deformation, so as to obtain the optimized welding process parameters when the final welding deformation is zero. Through the method of the present invention, the dynamic parameter adjustment of the whole process of ship section welding can be effectively carried out, the compensation of welding errors can be realized, and the final deformation of welded parts can be effectively controlled.

[0006] Technical solution: To achieve the above object, the present invention provides a dynamic compensation control method for ship section welding deformation based on the error accumulation principle, including the following steps:

[0007] S1: Collect multi-pass welding process parameters during the ship section welding process and the welding deformation amount corresponding to each parameter.

[0008] S2: Based on the error accumulation of the welding process and using a deep neural network, construct a layer-by-layer cumulative deformation error compensation model for each process of the section weldment.

[0009] S3: Use the compensation model to obtain compensation data, and through the model error compensation method, continuously and dynamically adjust the process parameters of each welding process in a cumulative compensation manner to obtain the optimal welding process parameters.

[0010] Further, the method for collecting multi-pass welding process parameters in step S1 is as follows:

[0011] The collection is completed through the orthogonal experiment method. The welding process parameters include welding current, welding voltage, welding speed, and wire extension length.

[0012] The method for collecting the welding deformation amount is as follows:

[0013] Analyze the influence of different welding process parameters on the weld forming size, size, and shape during the welding process to obtain experimental data, and use the established deep neural network model to simulate and analyze the experimental data to obtain the welding deformation amount corresponding to each parameter.

[0014] The collection of multi-pass welding process parameters is completed through the orthogonal experiment method. There are many factors involved in the welding process, and each factor contains multiple levels. If the comprehensive experiment method is used to collect sample data, a large number of experiments are required, which consumes too much cost and manpower. If sample data is selected arbitrarily, there is a lack of reasonable basis, which easily leads to large errors in the experimental results. Therefore, the design of the experiment needs to comprehensively consider the cost and manpower of the experiment, and the selection of samples should be representative and able to represent the overall distribution to a certain extent.

[0015] Denote the orthogonal table as L n (S r ), where n represents the number of experiments to be carried out, that is, the total number of samples, S represents the number of levels of each factor, and r represents the maximum number of factors that can be selected for level S. If the interaction between factors is not considered, the S value obtained through analysis in the orthogonal table should be consistent with the number of level factors measured in the experiment, and the maximum factor quantity value that can be selected should theoretically be greater than the actual factor number.

[0016] Further, in step S1, the analysis of variance method is used to analyze the experimental data. Since it is difficult to obtain the welding deformation amount in actual experiments, the data of the weld seam during welding deformation is generated by the orthogonal experiment method. This part is mainly to verify the rationality of the data generated by the orthogonal experiment method. The specific steps are as follows:

[0017] A1: Calculate the sum of squares: The sum of squares is divided into the total sum of squares SS T , the within-group sum of squares SS E and the between-group sum of squares SS A . The calculation formulas are as follows:

[0018] SS T =∑X 2 -(G 2 / N)

[0019] where G represents the sum of all data values, N represents the total number of data, and X is the generated welding deformation amount;

[0020]

[0021] is the total data mean, T i is the sum of data in each group, n i is the number of data in this group;

[0022] SS A =SS T -SS E

[0023] A2: Calculate the degrees of freedom: The df T , df E , df A of the three sums of squares are the total degrees of freedom, error degrees of freedom, and degrees of freedom of each factor respectively, where k is the number of levels of this factor,

[0024] df T =N - 1

[0025] df E =df T -(∑df A )

[0026] df A =k - 1

[0027] A3: Calculate the mean square error:

[0028] MS E =SS E / df E

[0029] MS A =SSA / df A

[0030] A4: Calculate the F value: The F value obtained by comparing the calculated F ratio with the corresponding critical value in the F distribution table, and the calculation formula is as follows:

[0031] F = MS E / MS A 。

[0032] Compare the calculated F with the critical value F0, and it is found that the influence of each factor level on the experimental results is relatively significant during the experiment, which is in line with the existing experience, proving the rationality of the experiment. At the same time, it also proves that the orthogonal experiment method can reduce the number of experiments on the premise of meeting the specific requirements of the experiment, saving manpower and costs to the greatest extent.

[0033] The welding deformation amount is generated by the orthogonal experiment method. The steps A1 - A4 are for calculating and verifying the rationality of the generated data to verify the effectiveness of the subsequent deformation control method with the generated welding deformation data.

[0034] Furthermore, the construction method of the layer - by - layer cumulative deformation error compensation model for each process of the segmented weldment in step S2 is as follows:

[0035] Through network parameter training of the cutting machine parameters and the corresponding welding deformation amount based on the deep learning network DNN, an error compensation model is constructed. The network structure of the error compensation model is multiple linear DNN models. By changing the input and output parameters of the DNN model and the transfer parameters between different DNN models, a complete error compensation model is constructed. The calculation formula from the input layer to the middle layer is:

[0036]

[0037] The activation function is the Sigmoid function:

[0038]

[0039] Furthermore, the construction of the deep learning network in step S2 includes: For the adaptive compensation control function, it is completed through the feedback calculation of the DNN algorithm, and the output minimum deformation amount is controlled by defining the loss function between the input and output, where the loss function is:

[0040]

[0041] Among them, F is the welding machine parameters, including welding current, welding voltage, welding speed, and wire extension length, Y is the weld formation size, including weld width, weld depth, and reinforcement height, and ||x - y||2 is the L2 norm of x - y.

[0042] Furthermore, the update method for the network parameters of the deep learning network and the welding machine parameters in step S2 is as follows:

[0043] The parameters between processes are associated by the forward activation function. To achieve the function of reverse regulating the intermediate process parameters from the final deformation, a loss function is defined, and the compensation network parameters and the welding machine parameters are updated by the loss values between the output deformation and each layer of parameters;

[0044] The update process is shown in the following formula:

[0045]

[0046] where ⊙ is the Hadamard product, and are the partial derivative terms of the compensation network weights and biases respectively, and the updated network weights and bias terms can be calculated from the above formula.

[0047] Furthermore, the transfer method of the compensation network parameters between multiple cascaded DNN networks in the deep learning network in step S2 is as follows:

[0048] The transfer terms are the neural network weights w and biases b between welding processes. Based on the existing network, the training speed of the compensation network for the next welded part is improved, and the compensation network obtained through the parameter training of multiple welded parts will have high applicability and robustness.

[0049] Furthermore, step S3 is specifically as follows:

[0050] The difference between the sum of the deformations of all welds of any welded part and the target deformation amount is used as the cost function of the compensation network, and the process deformation amount value when the cost function obtains the optimal value is solved through the stochastic gradient update function. The relevant functions and the solution process are shown in the following formula:

[0051]

[0052] where f s is the cumulative deformation of all welds, and f(x i ) is the deformation corresponding to each weld respectively;

[0053] f(x i ) = g(θ i )x i

[0054] g(θ i ) is the corresponding function between each weld and the deformation of the previous welding process, and θ i are the welding parameters;

[0055]

[0056] Among them, J(θ i ) is the cost function of the compensation algorithm, and y (j) is the true deformation amount.

[0057] Furthermore, in the step S3, the cost function J(θ i ) of the compensation algorithm has the following gradient update method during its implementation process:

[0058]

[0059] As shown in the above formula, the value of g(θ i ) that meets the conditions is obtained simultaneously by the gradient descent method.

[0060] Beneficial effects: Compared with the prior art, the present invention starts from overall optimization. For multi-pass welding processes and cases with a large number of welds in a single-pass process, by constructing a welding model compensation network through the proposed deep learning network, under the proposed compensation algorithm, the welding parameters and related deformations in the continuous welding task of the welding machine are dynamically adjusted. With the final deformation error of the welded part as the objective function, the welding parameters of each weld are dynamically adjusted under this objective, so as to obtain the goal of minimizing the cumulative deformation of all welds and the deformation of the welded part, and finally achieve the overall optimization of the welding process and the best welding result. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is the flow chart of the dynamic compensation deformation control method of the present invention;

[0062] Figure 2 is the construction and structural schematic diagram of the compensation model;

[0063] Figure 3 is the physical diagram of the multi-pass weld process of the workpiece;

[0064] Figure 4 is the schematic diagram of dynamically adjusting the welding process parameters of each pass from the total deformation of the welded part. DETAILED DESCRIPTION OF THE INVENTION

[0065] The following further clarifies the present invention in conjunction with the drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, various equivalent forms of modification of the present invention by those skilled in the art all fall within the scope defined by the appended claims of this application.

[0066] The present invention provides a dynamic compensation control method for ship section welding deformation based on the error accumulation principle. Referring to Figure 1 , it includes the following steps:

[0067] S1: Collect the multi-pass welding process parameters and the corresponding welding deformation amount during the ship section welding process;

[0068] S2: Based on the error accumulation of the welding process and the deep neural network, construct a layer-by-layer cumulative deformation error compensation model for each process of the segmented welded parts;

[0069] S3: Use the compensation model to obtain compensation data, and through the model error compensation method, continuously adjust the process parameters of each welding process dynamically in a cumulative compensation manner to obtain the optimal welding process parameters.

[0070] Regarding step S1:

[0071] The method for collecting multi-pass welding process parameters is as follows:

[0072] The collection is completed through the orthogonal experiment method. The welding process parameters include welding current, welding voltage, welding speed, and wire extension length.

[0073] The method for collecting the welding deformation amount is as follows:

[0074] Analyze the influence of different welding process parameters on the weld forming size, size, and shape during the welding process to obtain experimental data, and perform simulation analysis on the experimental data through the established deep neural network model to obtain the welding deformation amount corresponding to each parameter. In this embodiment, the welding deformation amount data is the weld forming size: weld width B, penetration depth H, and reinforcement height a.

[0075] In this embodiment, the collection of multi-pass welding process parameters is completed through the orthogonal experiment method. There are many factors involved in the welding process, and each factor contains multiple levels. If the comprehensive experiment method is used to collect sample data, a large number of experiments are required, and the cost and manpower consumed are too much. If sample data is selected arbitrarily, there is a lack of reasonable basis, which is likely to lead to large errors in the experimental results. Therefore, the design of the experiment needs to comprehensively consider the cost and manpower of the experiment, and the selection of samples should be representative and able to represent the overall distribution to a certain extent.

[0076] Denote the orthogonal table as L n (S r ), where n represents the number of experiments to be carried out, that is, the total number of samples, S represents the number of levels of each factor, and r represents the maximum number of factors that can be selected for level S. If the interaction between factors is not considered, the S value obtained through the analysis in the orthogonal table should be consistent with the number of level factors measured in the experiment, and the maximum factor quantity value that can be selected should theoretically be greater than the actual number of factors.

[0077] Through process analysis, it can be seen that different process parameters will affect the formation of robot welds. By referring to relevant materials and previous experimental analyses, in this embodiment, four factors that have a greater impact on welding are selected, namely welding current, welding voltage, welding speed, and wire extension length, which are represented by I, U, V, and L respectively. These four process parameters are used as experimental factors, and the level values of each factor are shown in Table 1.

[0078] Table 1

[0079] Horizontal I (A) V (cm / min) U (v) L (mm) 1 180 40 18 12 2 190 50 19 18 3 200 60 20 4 210 70 21

[0080] In this embodiment, the welding material selected for the experiment is Q235 carbon steel commonly used in port machinery and shipbuilding. The size of the steel plate used in the experiment is 400mm×250mm×5mm, the welding shielding gas is CO2, and the gas flow rate is 20L / min. The welding angle of the welding torch is 90°, a DC welding power source is used as the arc source, and a fixture is used to position and clamp the workpiece to be welded before welding. During the welding process, the control variable method can be used to change several parameter values that need to be verified, namely the four process parameters I, U, V, and length L, and other parameters are defaulted to quantitative values and will not be changed during the experiment. After the experiment, the weld width B, weld depth H, and reinforcement height a need to be compared to judge the strength of the influence of the process parameters on the results. Figure 3 The formed weld in the welding experiment is shown as follows. The weld formation dimensions obtained from the above experiment are shown in Table 2 in the appendix.

[0081] Table 2

[0082]

[0083]

[0084] In this embodiment, the analysis of variance method is used to analyze the experimental data. Since it is difficult to obtain the welding deformation amount in actual experiments, the data of the weld during welding deformation is generated by the orthogonal experiment method. This part is mainly to verify the rationality of the data generated by the orthogonal experiment method. The specific steps are as follows:

[0085] A1: Calculate the sum of squares: The sum of squares is divided into the total sum of squares SS T , the within-group sum of squares SS E and the between-group sum of squares SS A . The calculation formulas are as follows:

[0086] SS T =ΣX 2 -(G 2 / N)

[0087] where G represents the sum of all data values, N represents the total number of data, and X is the generated welding deformation amount;

[0088]

[0089] is the overall data mean, T i is the sum of data in each group, n i is the number of data in this group;

[0090] SS A = SS T - SS E

[0091] A2: Calculate the degrees of freedom: the df of the three types of sum of squares T , df E , df A are the total degrees of freedom, error degrees of freedom, and degrees of freedom of each factor respectively, where k is the number of levels of this factor,

[0092] df T = N - 1

[0093] df E = df T -(∑df A )

[0094] df A = k - 1

[0095] A3: Calculate the mean square error:

[0096] MS E = SS E / df E

[0097] MS A = SS A / df A

[0098] A4: Calculate the F value: The F value obtained by comparing the calculated F ratio with the corresponding critical value in the F distribution table, and the calculation formula is as follows:

[0099] F = MS E / MS A .

[0100] Compare the calculated F with the critical value F0, and it is found that the levels of each factor have a relatively significant impact on the experimental results during the experiment, which is in line with the existing experience, proving the rationality of the experiment. At the same time, it also proves that the orthogonal experiment method can reduce the number of experiments on the premise of meeting the specific requirements of the experiment, saving manpower and costs to the greatest extent.

[0101] The welding deformation amount is generated by the orthogonal experiment method. The steps A1 - A4 are for calculating and verifying the rationality of the generated data, and the generated welding deformation data is used to verify the effectiveness of the subsequent deformation control method.

[0102] Regarding step S2:

[0103] Referring to Figure 2 , the construction method of the layer - by - layer cumulative deformation error compensation model for segmented welded parts is as follows:

[0104] By training the network parameters of the cutting machine parameters and the corresponding welding deformation amount through the deep - learning network DNN, an error compensation model is constructed. The network structure of the error compensation model is multiple linear DNN models. By changing the input and output parameters of the DNN model and the transfer parameters between different DNN models, a complete error compensation model is constructed. The calculation formula from the input layer to the middle layer is:

[0105]

[0106] The activation function is the Sigmoid function:

[0107]

[0108] The construction of the deep - learning network includes: for the adaptive compensation control function, it is completed through the feedback calculation of the DNN algorithm. By defining the loss function between the input and the output, the minimum deformation amount of the output is controlled. The loss function is:

[0109]

[0110] Among them, F is the welding machine parameters, including welding current, welding voltage, welding speed, and wire extension length; Y is the weld formation size, including weld width, penetration depth, and reinforcement; ||x - y||2 is the L2 norm of x - y.

[0111] The update method of the network parameters of the deep - learning network and the welding machine parameters is:

[0112] Such as Figure 2 the adaptive compensation neural network of the welded part x1 composed of n welding processes. The parameters between processes are associated by the above - mentioned forward activation function. To realize the function of reverse regulating the intermediate process parameters from the final deformation, a loss function is defined, and the compensation network parameters and welding machine parameters are updated by the loss values between the output deformation and each layer of parameters;

[0113] The update process is shown in the following formula:

[0114]

[0115] Among them, ⊙ is the Hadamard product, and are the partial derivative terms for compensating the network weight parameters and biases respectively, and the updated network weights and bias terms can be calculated from the above formulas.

[0116] The method for transferring the compensation network parameters between multiple cascaded DNN networks in a deep learning network is as follows:

[0117] The transfer terms are the neural network weights w and biases b between welding processes. Based on the existing network, the training speed of the compensation network for the next welded part is improved. The compensation network obtained through the parameter training of multiple welded parts will have high applicability and robustness.

[0118] Regarding step S3:

[0119] The method of minimizing the overall deformation of a welded part by accumulating the deformation errors of multiple welds on the welded part, as shown in the compensation network model shown in Figure 4 Take the difference between the sum of the deformations of all welds of any welded part and the target deformation as the cost function of the compensation network, and solve the process deformation value when the cost function obtains the optimal value through the stochastic gradient update function. The relevant functions and solution processes are shown in the following formulas:

[0120]

[0121] where f s is the accumulated deformation of all welds, and f(x i ) is the deformation corresponding to each weld respectively;

[0122] f(x i ) = g(θ i )x i

[0123] g(θ i ) is the corresponding function between each weld and the deformation of the previous welding process, and θ i are the welding parameters;

[0124]

[0125] where J(θ i ) is the cost function of the compensation algorithm, and y (j) is the true deformation amount.

[0126] In the implementation process of the cost function J(θ i ) of the compensation algorithm, there is the following gradient update method:

[0127]

[0128] As shown in the above formula, the g(θ i) value.

[0129] The segmented welding deformation dynamic compensation network obtained through the above implementation steps can dynamically adjust the values of each welding process parameter during the segmented welding process, so as to achieve the purpose of obtaining the minimum welding deformation size.

Claims

1. A dynamic compensation control method for ship block welding deformation based on the error accumulation principle, characterized in that, It includes the following steps: S1: Collect multi-pass welding process parameters during the ship section welding process and the corresponding welding deformation amount for each parameter; S2: Based on the principle of error accumulation in the welding process, construct a layer-by-layer cumulative deformation error compensation model for each process of the section welded part based on a deep neural network; S3: Use the compensation model to obtain compensation data, and through the model error compensation method, continuously and dynamically adjust the process parameters of each welding process in a cumulative compensation manner to obtain the optimal welding process parameters; The method for collecting the welding deformation amount in step S1 is as follows: Analyze the influence of different welding process parameters on the weld forming size, size, and shape during the welding process to obtain experimental data, and perform simulation analysis on the experimental data through the established deep neural network model to obtain the welding deformation amount corresponding to each parameter; In step S1, the variance analysis method is used to analyze the experimental data, and the specific steps are as follows: A1: Calculate the sum of squares: The sum of squares is divided into the total sum of squares SS T , the sum of squares within groups SS E and the sum of squares between groups SS A . The calculation formula is as follows: SS T = ∑X 2 - (G 2 / N) Among them, G represents the sum of all data values, N represents the total number of data, and X is the generated welding deformation amount; is the total data mean, T i is the sum of data for each group, n i is the number of data in this group; SS A = SS T - SS E A2: Degrees of freedom for calculation: df for the three sums of squares T , df E , df A are the total degrees of freedom, error degrees of freedom, and degrees of freedom for each factor respectively, where k is the number of levels of the factor df T = N - 1 df E = df T -(∑df A ) df A = k - 1 A3: Calculate the mean square error: MS E = SS E / df E MS A = SS A / df A A4: Calculate the F value: The calculation formula is as follows: F = MS E / MS A ; The method for constructing the layer-by-layer cumulative deformation error compensation model for each process of the section welded part in step S2 is as follows: Through network parameter training of the cutting machine parameters and the corresponding welding deformation amount based on the deep learning network DNN, construct an error compensation model. The network structure of the error compensation model is multiple linear DNN models. By changing the input and output parameters of the DNN model and the transfer parameters between different DNN models, construct a complete error compensation model. The calculation formula from the input layer to the middle layer is: The activation function is the Sigmoid function: The construction of the deep learning network in step S2 includes: For the adaptive compensation control function, it is completed through the feedback calculation of the DNN algorithm, and the output minimum deformation amount is controlled by defining the loss function between the input and output.

2. The dynamic compensation control method for ship section welding deformation based on the error accumulation principle according to claim 1, characterized in that The method for collecting multi-pass welding process parameters in step S1 is as follows: The collection is completed through the orthogonal experiment method. The welding process parameters include welding current, welding voltage, welding speed, and wire extension length.

3. The dynamic compensation control method for ship section welding deformation based on the error accumulation principle according to claim 1, characterized in that The method for updating the network parameters of the deep learning network in step S2 is as follows: The parameters between processes are associated by the forward activation function. To realize the function of reverse regulating the intermediate process parameters from the final deformation, a loss function is defined, and the compensation network parameters are updated by the loss value between the output deformation and each layer of parameters; The update process is shown in the following formula: where ⊙ is the Hadamard product, and are the partial differential terms for compensating the network weight parameters and the bias, respectively, and the updated network weights and bias terms can be calculated from the above formulas.

4. The dynamic compensation control method for ship section welding deformation based on the error accumulation principle according to claim 3, characterized in that The method for transferring the compensation network parameters of the deep learning network in step S2 is as follows: The transfer items are the neural network weights w and bias terms b between welding processes.

5. The dynamic compensation control method for ship block welding deformation based on the error accumulation principle according to claim 1, characterized in that Step S3 is specifically as follows: Take the difference between the sum of all weld deformations of any welded part and the target deformation amount as the cost function of the compensation network, and solve the process deformation amount value when the cost function obtains the best value through the stochastic gradient update function. The relevant functions and solution processes are shown in the following formula: Among them, f s is the cumulative deformation of all welds, and f(x i ) is the deformation corresponding to each weld respectively; f(x i ) = g(θ i )x i g(θ i ) is the corresponding function between each weld seam and the deformation of the previous welding process, and θ i is the welding parameter; Among them, J(θ i ) is the cost function of the compensation algorithm, and y (j) is the true deformation amount.

6. The dynamic compensation control method for ship section welding deformation based on the error accumulation principle according to claim 5, characterized in that, In the step S3, the cost function J(θ i ) of the compensation algorithm has the following gradient update method during the implementation process: As shown in the above formula, the value of g(θ i ) that satisfies the condition is obtained simultaneously by the gradient descent method.

Images