^{1}

^{*}

^{1}

^{1}

^{1}

Weld seam deviation prediction is the key to weld seam tracking control, which is of great significance for realizing welding automation and ensuring welding quality. Aiming at the problem of weld seam deviation prediction in GMAW (gas metal arc welding), a method of weld seam deviation prediction based on arc sound signal is proposed. By analyzing the feature of the arc sound signal waveform, the time domain feature of the arc sound signal is extracted. The wavelet packet analysis method is used to analyze the time-fre - quency domain feature of the arc sound signal, and the wavelet packet energy feature is extracted. The time domain feature and wavelet packet energy feature are used to establish the feature vector, and the BP (back propagation) neural network is used to realize the weld seam deviation prediction. The results show that the method proposed in this paper has a good weld seam deviation prediction effect, with a mean absolute error of 0.234 mm, which provides a new method for GMAW weld seam recognition.

The deviation of the weld seam during the welding process will cause defects in the weld seam formation and deterioration of the mechanical properties of the welded structure. The key to solving this problem lies in the weld seam deviation prediction technology. The automatic tracking control of the welding seam is realized by the welding seam deviation prediction technology, which can avoid the deviation of the welding seam from the groove center due to the processing error and thermal deformation of the workpiece [

The arc sound signal has practical application value and contains a large amount of welding information [

The schematic diagram of the experimental setup is shown in

The welding diagram and welding torch trajectory are shown in

Data are collected within a welding torch swing period as a sample of arc sound signal.

Welding Parameter | The parameter value |
---|---|

Welding current/(A) | 120 |

Swing amplitude/(mm) | 6 |

Swing frequency/(Hz) | 1.5 |

Wire diameter/(mm) | 1 |

Welding speed/(mm/s) | 3.2 |

Shielding gas | 82% Ar + 18% CO_{2 } |

E d i f = ∑ i = 1 n x i , r i g h t 2 − ∑ j = 1 n x j , l e f t 2 (1)

V d i f = ∑ i = 1 n ( x i , r i g h t − x ¯ i , r i g h t ) 2 n − 1 − ∑ j = 1 n ( x j , l e f t − x ¯ j , l e f t ) 2 n − 1 (2)

where E d i f represents the energy difference, V d i f represents the variance difference, x i , r i g h t represents the sound pressure of the i-th sampling point at the right limit position, x j , l e f t represents the sound pressure of the j-th sampling point at the left limit position.

The wavelet packet analysis method is a time-frequency domain feature analysis

method established on the theoretical basis of wavelet analysis. It is good at analyzing the features of non-stationary signals. It can not only decompose the low frequency band of the signal, but also decompose the high frequency band, which has strong adaptability and flexibility. The wavelet packet decomposition algorithm was written as follows [

d l j , 2 n = ∑ k g 0 ( k − 2 l ) d k j + 1 , n d l j , 2 n + 1 = ∑ k g 1 ( k − 2 l ) d k j + 1 , n (3)

where d l j , 2 n , d l j , 2 n + 1 , d k j + 1 , n are the wavelet packet decomposition coefficients; g 0 ( k − 2 l ) and g 1 ( k − 2 l ) are conjugate quadrature low-pass and high-pass filters, respectively.

The wavelet packet energy is calculated by summing up the signal squares in each frequency band (following the wavelet packet decomposition). The energy corresponding to each frequency band is shown in Equation (4):

P i = ∑ k = 1 N | d m n ( k ) | 2 (4)

where N represents the original signal length, d m n ( k ) is the m-th frequency band wavelet packet decomposition sequence in the n-th layer; m = 0 , 1 , ⋯ , 2 n − 1 .

Daubechies3 (db3) wavelet base was selected after the comparative testing. The arc sound signal samples with 1 mm to the left, centering, 1 mm to the right of the weld seam were decomposed by three-layer wavelet packets respectively. They were divided into eight frequency bands to obtain the wavelet packet energy feature.

frequency bands account for the largest proportion, and the energy change is symmetrical when the weld seam deviation increases. This shows that the wavelet packet energy features in the seventh and eighth frequency bands are the most stable. The wavelet packet energy feature can well reflect the deviation change of the weld seam, but it cannot reflect the deviation direction. It is necessary to combine the time domain features to jointly identify the weld seam deviation.

The four-dimensional feature vector T = ( E d i f , V d i f , P 7 , P 8 ) is created by extracting time-domain and wavelet packet energy features. The 20 datasets were selected for each case, including the welding seam with 1 mm left, center, 1mm right, resulting in a total of 60 data sets. From each case, 14 sets were randomly selected to form 42 training sets, and the rest as the test set. The data were normalized by mapping it to the interval [0, 1], mostly to improve the model convergence speed and eliminate the feature dimension influence on the model. The normalization method is shown in Equation (5):

y = x − x min x max − x min (5)

where x represents input data, y represents output data, x max and x min represent the maximum and minimum values of input data respectively.

BP neural network is used to establish the weld seam deviation prediction model. Set the number of input neurons to 4, the number of hidden layer neurons to 5, the number of output neurons to 1, the learning rate to 0.1, the maximum training times to 1000, and the target error to 1 × 10^{−5}. BP neural network structure diagram is shown in

has achieved good prediction results. It can be seen that it is feasible to use the arc sound signal to predict the deviation of the GMAW weld seam, which provides a new way for real-time tracking and control of the weld seam.

This paper proposes a method of weld seam deviation prediction based on arc sound signal.

When there is a deviation in the weld seam, the waveform at the limit position of the arc sound signal changes significantly, so the energy and variance difference features at the limit position are extracted. Analyzing the wavelet packet energy feature of the arc sound signal, it is found that the high-frequency band energy decreases significantly with the increase of the weld seam deviation. The seventh and eighth frequency band wavelet packet energy features with the best performance are extracted. A four-dimensional feature vector is constructed using the time domain features and wavelet packet energy features of arc sound signals. The BP neural network is used to predict the weld seam deviation. The mean absolute error of the prediction results is 0.234 mm, which proves the effectiveness of the weld seam deviation prediction method proposed in this paper. Future work includes optimizing the predictive model and developing a real-time control system for seam tracking based on arc sound signals.

The authors declare no conflicts of interest regarding the publication of this paper.

Zhao, W., Yue, J.F., Liu, W.J. and Liu, H.H. (2021) Weld Seam Deviation Prediction of Gas Metal Arc Welding Based on Arc Sound Signal. World Journal of Engineering and Technology, 9, 51-59. https://doi.org/10.4236/wjet.2021.91004