Intelligent prediction method for tensile strength of FSW joints considering welding temperature and axial force

An intelligent prediction method for tensile strength of friction stir welding (FSW) joints considering welding temperature and axial force, wherein, the time series data of temperature on the advancing side (AS) and retreating side (RS) of the outside of the shoulder of the weldment surface and axial force are obtained by infrared thermal imager and force sensor;

• the one-dimensional convolutional neural network (1D CNN) is trained and tested with the acquired time series data to realize the prediction of the tensile strength of FSW joints;
• the specific steps are as follows:
• Step 1:
• during FSW process, the temperature of feature points on the AS and RS of the outside of the shoulder of the weldment surface and axial force are obtained by infrared thermal imager and force sensor;
• Step 2:
• the sampling frequency of the infrared thermal imager is different from that of the force sensor;
• the common multiple of the sampling frequency of the infrared thermal imager and the force sensor is taken as the common sampling frequency fs;
• By resampling method, the sampling frequency of the two devices is controlled to be the same, and the time series data of the temperature of the feature points on the AS and RS and the axial force are obtained;
• Meanwhile, the number of sampling points N is determined according to the welding duration;
• Step 3:
• the length of the welding interval is calculated according to the welding speed, sampling frequency, and the number of sampling points, and the tensile strength of this interval is obtained through the tensile test;
• Step 4:
• the temperature and axial force data processed in step 2 are processed into an N*3 matrix, and the tensile strength obtained in step 3 is processed into a 1*1 matrix, and the data set is prepared for the training and testing of one-dimensional convolutional neural network;
• Step 5:
• 1D CNN is established:
• the model has three inputs, including the time series data of the temperature of the feature points on the AS and RS and axial force, and the output is the tensile strength;
• According to step 4, the input matrix of the 1D CNN is N*3 and the output matrix is 1*1;
• 1D CNN is established, and the parameters of 1D CNN are preset, including the number of convolutional layers, the number and size of convolutional nuclei, the number and size of pooling layers, the number of fully connected layers and the number of units, and the activation function;
• Step 6:
• Huber robust loss function is used as an evaluation criterion, the N*3 temperature and axial force matrix processed in Step 4 is used as the input, and the 1*1 tensile strength matrix is used as the output to train the model;
• the parameters in Step 5 are optimized by the Huber loss function, And the 1D CNN prediction model with the parameter combination with the highest prediction accuracy is finally saved;
• Step 7:
• during the actual welding process, the temperature of feature points on the AS and RS of the outside of the shoulder of the weldment surface and axial force obtained by the infrared thermal imager and the force sensor are processed in step 2 and step 4, and used as the input of the optimized 1D CNN model in Step 6 to predict the tensile strength.