Research On The Methods For Quality Monitoring Based On Arc Sound Signal In CO₂ Arc Welding

The arc sound signal in the GMAW contains plenty of welding information, which is relevant closely to arc behavior, melting metal transition mode, arc stability, and it is an important signal source of welding quality control. Aiming at on-line monitoring of welding quality, the modern signals analysis methods were adopted to analyze the time and frequency characteristics of the arc sound signal in short circuit GMAW, and the relationship between arc sound and welding states. Then the concept of tone channel and its equivalent electrical model were advanced. The energy characters of sound signal in the different frequency bands which were extracted by wavelet packet analysis and LPC parameters of sound signal constructed the character vectors. BP, RBF artificial neural network and support vector machine were used to predict the spatter, shield gas flux and wire extension in CO₂ short-circuiting arc welding. The results shown that it is available to recognise welding states by arc sound signal. The main works were done as follows.In allusion to the character of GMAW process, a synchronous signal collection system was developed with AC6115.The signal collected software programmed with Visual Basic can be used to collect, display or pretreat signals. The MATLAB interface also was designed to analyze the welding signal.Many of signal analysis methods were used to analyze the time-domain and frequency domain characters of welding signals, Such as correlation analysis, Fourier spectrum, little spectrum, power spectrum and wavelet method etc. these methods enriched the recognition of the welding process from the point of view of the signal analysis. The research indicates that the arc sound presents a periodic "ringing" form, which mainly occurs at the end of short circuit transfer or the moment of arc re-ignition .The sound wave form is some similar to the power difference of arc power, the variation of arc power presents the source of arc sound;The sound's main frequency spectrum focused in 0-12kHz, and there areseveral formants in this frequency band. The frequency characters of sound are associated with the change of welding parameters;the wavelet analysis can eliminate the noise from original signals without leading the distortion, it also used to extract information if welding signal in different frequency bands, and the strange points of welding voltage in higher frequency ranges by wavelet packet analysis can be used to divide different wending periods.The Fourier spectrum and wavelet methods were respectively used to extract energy characters of sound signal in the different frequency bands, which extract by wavelet were made up the character vectors. Correlation between vector elements and spatter in CO2 was evaluated by statistic test theory, and reduced dimensions of vectors and get simplest characteristics subclasses which were used as neural networks input vectors. Then the BP neural networks model were applied and successfully predict spatter in CO2 welding process.By analysis based on time-domain and frequency-domain, the producing and formative mechanism of acoustic signal is approached, and the tone channel is put forward. The research indicates that variation of arc power is the exciting source, and tone channel is made up of the shielded gas, arc column, magnetic field and temperature field during the welding process. They are combined to generate the arc sound. For thoroughly studying characteristic of arc sound, Linear Prediction Coding (LPC) forecast model was constructed to describe the system of tone channel in welding process which is described with a all-poles time dependent coefficients digital filter, which basically conform the assumption of LPC model. Therefore, the LPC model of sound can be factually considered a parametric estimation of the tone channel. The RBF neural networks and SVM model were applied for pattern recognition of the flux of the shield gas and wire extension in welding process, in which the input vectors were constructed by the LPC parameters and reflectance of sound signal.