Method for predicting mechanical properties of friction stir welded joint based on welding process shape and performance characteristics

Abstract

The present application belongs to the field of friction stir welding quality prediction, and relates to a FSW joint mechanical property prediction method based on welding process shape and performance characteristics. The present application uses a combination of theory and experiment to carry out friction stir welding orthogonal test, uses a force meter, an infrared thermal imager and a line laser sensor to measure the welding process upset force, welding surface feature point temperature, butt joint gap and step difference time sequence data, realizes multi-source heterogeneous data fusion through data processing, establishes a time sequence correlation relationship model of welding process shape and performance characteristics and joint mechanical property based on a one-dimensional convolutional neural network-long short-term memory neural network joint model, and realizes joint tensile strength and microhardness prediction. The present application can realize in-situ prediction of the mechanical property of a friction stir welding joint, provide a reference for real-time control of welding process parameters and ensure welding quality.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of friction stir welding (FSW) quality prediction, and particularly relates to a FSW joint mechanical property prediction method based on welding process shape and performance characteristics, which comprehensively uses a force meter, an infrared thermal imager, and a line laser sensor to obtain upset force, welding surface feature point temperature, butt joint gap, and step difference time sequence data, and predicts joint tensile strength and microhardness based on welding process shape and performance characteristics and a one-dimensional convolution neural network-long and short term memory (1DCNN-LSTM) combined model. BACKGROUND

[0002] FSW technology is a solid-state connection method that realizes material welding through the high-speed rotation of a stir head, frictional heat generated by the stir head and material, and material plastic deformation heat, so as to soften the connection part of the welding piece and realize material welding under the dual action of friction and heat. Compared with traditional melting welding, FSW technology has the advantages of small welding joint organization grain size, good connection strength and tensile strength, no smoke generated during welding, small residual stress after welding, and no welding deformation, etc., and has become the preferred method for welding 2-7 series aluminum alloy structural parts. After more than thirty years of development, FSW technology has become relatively mature, and has been widely used in the fields of aerospace, railway transportation, ships, automobiles, and power electronics.

[0003] In actual production, the welding quality of FSW is usually evaluated by the mechanical properties such as the tensile strength and microhardness of the joint after welding. The mechanical properties of the joint are not only affected by the coupling effect of welding force and welding temperature, but also by the geometric quantities (gap and step difference) of the butt joint. Due to the machining and assembly errors of the welding piece and the deformation during the welding process, the butt joint will have a gap and a step difference during the FSW process, which will affect the welding quality. Wang Tao et al. studied the influence of the gap and misalignment of the butt joint on the mechanical properties of 5A05 aluminum alloy FSW joints in "Influence of Assembly Factors on Welding Quality of 5A05 Aluminum Alloy Thin Plate Friction Stir Welding". The test results show that when the misalignment is greater than 0.5 mm or the gap is greater than 0.8 mm, the mechanical properties of the joint decrease and defects occur.

[0004] In the FSW process, the changes in thermal cycle state, upset force, and step difference and gap, etc. make it difficult to predict the mechanical properties of the joint after welding, and thus the welding process parameters cannot be adjusted in real time, and the welding quality cannot be guaranteed. Therefore, it is urgent to establish the mapping relationship between the mechanical properties of the joint and the upset force, welding temperature, and geometric quantities of the butt joint, so as to realize the in-situ prediction of the tensile strength and microhardness of the joint.

[0005] Kamal et al. established a neural network model for mechanical properties of AA2219 alloy FSW joints in Parameter optimization of friction stir welding of cryorolled AA2219 alloy using artificial neural network modeling with genetic algorithm, which takes welding speed, rotation speed and tilt angle as input to predict the tensile strength and microhardness of the joints. Elaziz et al. conducted dissimilar FSW experiments of AA5083 and AA2024 aluminum alloys in Utilization of Random Vector Functional Link integrated with MarinePredators Algorithm for tensile behavior prediction of dissimilar friction stir welded aluminum alloy joints, and established a tensile strength prediction model of the joints with the input of upset force by using random vector function chain neural network. De et al. considered the influence of peak temperature on the performance of joints in the process of FSW of AA5754 H111 aluminum alloy in Prediction of the Vickers Microhardness and Ultimate Tensile Strength of AA5754 H111 Friction Stir Welding Butt Joints Using Artificial Neural Network, and predicted the tensile strength and microhardness of the joints by using artificial neural network.

[0006] Effective prediction of the mechanical properties of FSW joints is the premise to ensure the welding quality. The above researches show that neural network intelligent algorithm is an effective method to realize the performance prediction of FSW joints. However, the existing performance prediction models of FSW joints do not consider the change of the geometric quantity of the butt joint surface, usually only consider the influence of welding process parameters or single physical quantity temperature or upset force on the performance of the joints, and more do not comprehensively consider the influence of the time sequence characteristics and coupling effect of the three on the mechanical properties of FSW joints. SUMMARY

[0007] The present application is directed to the problem that it is difficult to predict the mechanical properties of the welded joint after welding by real-time detection of the shape and performance characteristic quantities of the welding process, and proposes an intelligent prediction method for the mechanical properties of the FSW joint based on the shape and performance characteristics of the welding process. First, the FSW experiment is performed to measure the upsetting force, welding temperature, gap and step difference time series data set of the butt joint during the welding process; then the correlation model of the upsetting force, temperature on the advancing side, temperature on the retreating side, gap, step difference and the tensile strength and microhardness of the joint is established by the 1D CNN-LSTM combined neural network; finally, the data collected online during the FSW process is taken as the input to realize the in-situ prediction of the mechanical properties of the joint and provide a reference for the welding process parameter control.

[0008] The technical scheme of the present application:

[0009] A method for predicting the mechanical properties of the FSW joint based on the shape and performance characteristics of the welding process, which combines theory and experiment to perform the orthogonal test of friction stir welding, uses a force gauge, an infrared thermal imager and a line laser sensor to measure the upsetting force, welding temperature, gap and step difference time series data of the welding process, realizes data fusion through data processing, makes a data set, establishes the correlation model of the shape and performance characteristics of the welding process and the mechanical properties of the welded joint after welding based on the 1D CNN-LSTM combined neural network model, realizes the prediction of the tensile strength and microhardness of the joint, and the specific steps are as follows:

[0010] Step 1: Consider the rotational speed of the stirring head, welding speed, gap and step difference of the butt joint to design an orthogonal test;

[0011] Step 2: Perform the FSW experiment according to step 1, and use a force gauge, an infrared thermal imager and a line laser sensor to measure the upsetting force, temperature on the advancing side, temperature on the retreating side and gap and step difference time series data during the welding process;

[0012] Step 3: After welding, the tensile strength and microhardness values of the welded joint are obtained by tensile test and hardness test;

[0013] Step 4: Extract the time series data collected by the force gauge, infrared thermal imager and line laser sensor in step 2. Determine the number of sampling points according to the welding speed, sample length and sampling frequency, and use the resample function of python to perform resampling processing on the data. Group the upsetting force, temperature at the characteristic points on the advancing side, temperature at the characteristic points on the retreating side, gap and step difference according to the cutting position of the tensile sample, and fuse them with the tensile strength and microhardness data in step 3 to complete the data set making;

[0014] Step 5: Use MySQL to establish a welding database, and store the data set in step 4 into the database;

[0015] Step 6: The dataset in step 5 is called by python, then the data is normalized, and the dataset is divided by five-fold cross-validation, and finally the time sliding window is added to slice the dataset into Mx5 forging force, forward side temperature, backward side temperature, gap and step difference input matrix and Nx2 tensile strength and microhardness output matrix, where M and N are the number of corresponding sampling points.

[0016] Step 7: Import the third-party deep learning framework Keras in python, and establish a 1DCNN-LSTM combined neural network model;

[0017] The 1DCNN calculation formula is:

[0018] (1)

[0019] In the formula, w y and b y are the weight matrix and bias term of the convolution kernel respectively; x t is the tthinput sample value; is the activation function; * represents convolution operation; y t represents the output value of the current convolution layer.

[0020] The LSTM update formula is:

[0021] (2)

[0022] (3)

[0023] (4)

[0024] (5)

[0025] (6)

[0026] (7)

[0027] In the formula, w f , b f , w i , b i , w o , b o , w c , b c are the weight matrix and bias term of the forgetting gate f t , the update gate i t , the output gate O t , and the state unit ; x t is the input value of LSTM at t time; h t-1is the output value of the hidden layer at t-1 time; is the updated state unit, is the state unit before updating.

[0028] The structure parameters of the preset 1DCNN-LSTM combined neural network are set, the model takes the M*5 matrix of step 6 as input, and the N*2 matrix as output.

[0029] Step 8: Set the model training parameters, perform five-fold cross-validation, optimize the parameters through the Huber loss function curve convergence and the prediction accuracy of the test set, and save the optimized 1DCNN-LSTM model;

[0030] Step 9: Call the optimized model in step 8 to predict the test set and save the prediction results. Root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) are used to evaluate the model;

[0031] Step 10: During the FSW process, use the force gauge, infrared thermal imager and line laser sensor to collect the upset forging force, welding temperature, butt joint gap and step difference data online, process steps 4 and 6, and use them as the input of the optimized 1DCNN-LSTM combined neural network model in step 8 to realize the in-situ prediction of the mechanical properties of the FSW joint.

[0032] The beneficial effects of the application are: based on the 1DCNN-LSTM combined neural network, an intelligent prediction model of joint tensile strength and microhardness considering the upset forging force, welding temperature, butt joint gap and step difference in the welding process is established, the in-situ prediction of the mechanical properties of the friction stir welding joint is realized, and the real-time control of the welding process parameters is provided for reference to ensure the welding quality. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is the time sequence curve after fusion processing and grouping;

[0034] Figure 2 is a one-dimensional convolutional neural network-long short-term memory neural network model structure diagram;

[0035] Fig. 3 (a)~Fig. 3 (e) are the error loss function curves of the training set and the validation set of the five-fold cross-validation of the model;

[0036] Fig. 4 (a)~Fig. 4 (e) are the prediction results of the five-fold cross-validation of the model;

[0037] Figure 5 is the relative error comparison of 100 groups of prediction results. DETAILED DESCRIPTION

[0038] The specific embodiments of the present application are described in detail below with reference to the technical solutions and drawings, but the present application is not limited by the embodiments.

[0039] This embodiment uses 2219 aluminum alloy as the welding material, uses an FSW-5M type friction stir welding experimental machine tool to perform FSW experiments, uses a force meter, an infrared thermal imager, and a line laser sensor to obtain upset force, welding surface temperature, and aluminum plate butt joint gap and step difference data, obtains joint tensile strength and microhardness through post-weld tensile testing and hardness testing. Python language is used to write code, a 1DCNN-LSTM joint neural network is used to establish a time sequence correlation relationship model of upset force, front side temperature, back side temperature, gap, step difference, and joint tensile strength and microhardness, and the mechanical properties of the joint are predicted, and the specific steps are as follows:

[0040] Step 1: Consider the rotational speed of the stirring head, the welding speed, the butt joint gap, and the step difference to design a four-factor five-level orthogonal experiment, L 25 (5 4 )orthogonal table as shown in Table 1.

[0041] Table 1 L 25 (5 4 )orthogonal experiment

[0042]

[0043] Step 2: According to the orthogonal table in step 1, 25 groups of 2219 aluminum alloy FSW experiments are performed, wherein the size of the aluminum alloy plate is 298 mm x 120 mm x 18 mm, the shoulder diameter of the stirring head is 32 mm, and the stirring needle length is 17.8 mm. A C2 type force meter of the German HBM company is installed at the bottom of the friction stir welding machine tool workbench to measure the upset force during welding; a FLIR A615 infrared thermal imager is arranged in front of the welded part at an angle of 30° with the main shaft, and the relative position of the machine tool main shaft and the infrared thermal imager is kept unchanged during welding, and the temperatures of the characteristic points on the advancing and retreating sides 28 mm away from the weld center are collected; the HD6 type line laser sensor of the CZLSLASER company is used to obtain the butt joint gap and step difference value.

[0044] Step 3: The post-weld samples in step 2 are cut into standard samples by wire cutting, 4 samples are taken from each experiment, and a total of 100 samples are obtained. A DNS-300 electronic universal testing machine is used to perform tensile testing to obtain the tensile strength value of the joint, and a HV-1000A Vickers hardness tester is used to measure the microhardness of the joint.

[0045] Step 4: Extract the time series data of the top forging force, forward and backward side feature point temperature, gap and step difference collected by the dynamometer, infrared thermal imager and line laser sensor. According to the welding speed, sample length and sampling frequency, determine the number of sampling points as 30, use the resample function of python to resample the data, group the top forging force, forward side feature point temperature, backward side feature point temperature, gap and step difference according to the cutting position of the tensile sample, and align with the tensile strength and microhardness data, realize the fusion of multi-source heterogeneous data, and the time series curve after fusion and grouping is shown in Figure 1 Table 2 shows a group of training data samples. The actual data set contains 100 samples.

[0046] Table 2 shows a group of training data samples. The actual data set contains 100 samples.

[0047]

[0048] Step 5: Use MySQL to establish FSW welding database, store the data in step 4 into the database.

[0049] Step 6: Call the data set in step 5 through the fetchall function of python, then use the MinMaxScaler function in the sklearn library of python to perform maximum and minimum value normalization on the data, then use the five-fold cross-validation method to divide the data set, divide the 100 samples in the data set into 5 subsets and train the model 5 times. For each iteration of training the network, 4 subsets, i.e. 80 samples, are used as training data, and the remaining subset is used as test data. Finally, add a time sliding window to slice the data into a 30x5 top forging force, forward side temperature, backward side temperature, gap and step difference input matrix and a 1x2 tensile strength and microhardness output matrix.

[0050] Step 7: Program hardware facilities using Intel(R) Core(TM) i9-10900K CPU @ 3.70GHz server, Window 11-64 bit operating system, 64 GB running memory. The programming software is PyCharm, which imports the deep learning framework Keras based on TensorFlow as support, and uses its deep neural network Sequential model to realize the construction of 1DCNN-LSTM joint neural network model:

[0051] The 30x5 matrix in step 6 is used as the input of the input layer Input Layer of the 1DCNN-LSTM model, then it is subjected to one-dimensional convolution operation through the convolution layer Conv1, realizing local feature extraction and fusion. The one-dimensional convolution calculation formula is:

[0052] (1)

[0053] where w y and b y are the weight matrix and bias term of the convolution kernel respectively; x t is the value of the t-th input sample; f is the activation function; * represents the convolution operation; y t represents the output value of the current convolution layer.

[0054] After that, the maximum pooling layer Max pooling is used for maximum pooling operation, and the maximum pooling calculation formula is:

[0055] (2)

[0056] where is the output of the (k+1)-th convolution layer in the j-th pooling domain after the pooling operation of the feature map corresponding to the i-th convolution kernel; is the j-th pooling domain of the k-th convolution layer; is the value of the n-th neuron in the feature map corresponding to the i-th convolution kernel of the k-th convolution layer.

[0057] Then, the Batch normalization layer is used for batch standardization processing, and then the convolution layer Conv2 is used for the second convolution operation. Then, the Batch normalization layer is used for the second batch standardization processing, and then the convolution layer Conv3 is used for the third convolution operation. Then, the average pooling layer Average pooling is used for average pooling operation, and the average pooling calculation formula is:

[0058] (3)

[0059] where the symbols are consistent with formula (2).

[0060] Then, the LSTM1 layer is used to save the historical input information. The LSTM is composed of three logical gate structures, including the forgetting gate, the update gate and the output gate. The LSTM forgetting gate update formula is as follows:

[0061] (4)

[0062] (5)

[0063] where w f and b f are the weight matrix and bias term of the forgetting gate f t respectively; x t is the input value of the LSTM at time t; h t-1is the output value of the hidden layer at time t-1; the forget gate adopts a sigmoid activation function.

[0064] The LSTM update gate calculation formula is:

[0065] (6)

[0066] (7)

[0067] (8)

[0068] (9)

[0069] where w i and w c are the weight matrices of the update gate i t and the state unit , respectively; b i and b c are their bias terms; tanh is the hyperbolic tangent activation function; the product of i t and is the degree of preservation of new information; and C t is the updated state unit.

[0070] The LSTM output gate calculation formula is:

[0071] (10)

[0072] (11)

[0073] where O t is the output value of the LSTM at time t; w o and b o are the weight matrix and bias term of the output gate O t , respectively; C t is scaled by the tanh function and multiplied by O t , i.e., the value h t of the hidden state of the current unit at time t.

[0074] Then add the Dropout layer, set the retention probability to 0.5, then go through LSTM2 and LSTM3, then flatten the output matrix h t of the LSTM through the Flatten layer, then go through the fully connected layer Dense, add L1 regularization, set L1 to 0.01, and finally output the tensile strength and microhardness values by the Output Layer, as shown in Figure 1 . The parameter configurations of each layer of the 1DCNN-LSTM model are shown in Table 3.

[0075] Table 3 Parameter configuration of each layer of the 1DCNN-LSTM model

[0076]

[0077] Step 8: Set the batch size to 8, the learning rate to 0.001, use the Adam optimizer, and use the Huber robust loss function as the loss function. Perform 1000 iterations of training, and draw the error loss function curve. Optimize the parameters by the convergence of the Huber loss function curve and the prediction accuracy of the test set. The learning curve results of the five-fold cross-validation training process of the prediction model after parameter optimization are shown in FIGS. 3(a)-3(e). The error loss function curve converges, and the optimized 1DCNN-LSTM model is saved.

[0078] Step 9: Call the optimized model in step 8 to predict the test set and save the prediction results. The five-fold cross-validation prediction results are shown in FIGS. 4(a)-4(e). The 1DCNN-LSTM prediction model has 100 groups of prediction results, and the relative error comparison of each group of prediction results is shown in FIG. 5. Figure 5

[0079] Table 4 Prediction accuracy of tensile strength and microhardness

[0080]

[0081] It can be seen that the average relative error between the predicted values of tensile strength and microhardness and the true values is less than 5%, and the accuracy is high, which can prove that the 1DCNN-LSTM neural network model can effectively predict the mechanical properties of FSW joints.​

Claims

1. A method for predicting mechanical properties of FSW joints based on welding process characteristics, characterized in that, The specific steps are as follows: Step 1: Consider the stirring head speed, welding speed, butt joint gap and step difference design orthogonal test; Step 2: According to step 1, the FSW experiment is carried out, and the force meter, infrared thermal imager and line laser sensor are used to measure the upsetting force, advancing side temperature, retreating side temperature and butt joint gap and step difference time series data during welding; Step 3: After welding, the tensile strength and microhardness value of the welded joint are obtained by tensile test and hardness test; Step 4: Extract the time series data collected by the force meter, infrared thermal imager and line laser sensor in step 2: according to the welding speed, sample length and sampling frequency, determine the number of sampling points, use the resample function of python to resample the data, group the upsetting force, advancing side feature point temperature, retreating side feature point temperature, gap and step difference according to the cutting position of the tensile sample, and fuse with the tensile strength and microhardness data in step 3 to complete the data set preparation; Step 5: Use MySQL to establish a welding database and store the data set in step 4 into the database; Step 6: Call the data set in step 5 through python, then normalize the data, divide the data set by five-fold cross validation, and finally add a time sliding window to slice the data set into Mx5 upsetting force, advancing side temperature, retreating side temperature, gap and step difference input matrix and Nx2 tensile strength and microhardness output matrix; Where M and N are the corresponding number of sampling points; Step 7: Import the third-party deep learning framework Keras in python and establish a 1DCNN-LSTM joint neural network model; The 1DCNN calculation formula is: （1） where w y and b y are the weight matrix and bias term of the convolution kernel, respectively; x t is the value of the t-th input sample; is the activation function; * denotes the convolution operation; y t denotes the output value of the current convolution layer; The LSTM update formula is: （2） （3） （4） （5） （6） （7） wherein w f , b f , w i , b i , w o , b o , w c , b c are weight matrices and bias terms of the forget gate f t , the update gate i t , the output gate O t , and the state cell , respectively; x t is the input value of the LSTM at time t; h t-1 is the output value of the hidden layer at time t-1; C t is the updated state cell, is the state cell before updating. The structure parameters of the preset 1DCNN-LSTM joint neural network are set, the model takes the Mx5 matrix in step 6 as input and the Nx2 matrix as output; Step 8: Set the training parameters, perform five-fold cross validation, optimize the parameters through Huber loss function curve convergence and test set prediction accuracy, and save the optimized 1DCNN-LSTM model; Step 9: Call the optimized model in step 8 to predict the test set and save the prediction results; The root mean square error, mean absolute error and mean absolute percentage error are used to evaluate the model; Step 10: Use the force meter, infrared thermal imager and line laser sensor to collect the upsetting force, welding temperature, butt joint gap and step difference data during FSW process, process them according to steps 4 and 6, and use them as the input of the optimized 1DCNN-LSTM joint neural network model in step 8 to realize the on-site prediction of the mechanical properties of FSW joint.

Images