| As one of the concomitant signals in welding process, the arc sound isn't expected by people, but is avoidless. And its essence is vibration signal, of which the frequency falls in between 20Hz and 20kHz. Factually, an experienced welder can roughly judge the forming quality and stability in welding process, depending on hearing and combining the individual experience. Therefore, the changing information of the relevant state in welding process should be contained in the arc sound. Aiming at doing the pattern recognition of penetration state in welding process, the dynamic arc sound signal in MIG butt welding with spray transfer was selected as study object and the modern digital signal analysis methods were adopted to analyze the relationship between it and penetration state. Then the characteristics of the arc sound characterizing penetration state were extracted and quantified, the united character vector was constructed, and dimensionality reducting was carried out by principal component analysis(PCA). BP and RBF network were used to build the mapping model between the arc sound and penetration state The test results shown that it is feasible to classify and distinguish the penetration state in welding process by the arc sound. The main works were done as follows.Based on data acquisition(DAQ) hardware platform, a set of software system was developed by means of packaging function modules and calling dynamic link programs of the hardware driver with virtual instrument programming language(LabVIEW), which can be used to collect and save data, verify client user status, record experiment parameters, collect parameter settings on DAQ card, echo historic waveform, denoise the signal, extract the characteristics and so on.Many of signal analysis methods were used to research on the relationship between the arc sound signal and penetration state in welding process, such as short-time Fourier transform(STFT) and wavelet transform(WT) etc. The methods enriched the recognition of the welding process from the point of view of the signal analysis. The results indicate that using the db04 as wavelet basis to denoise the arc sound based on wavelet decompostion at level 4 can effectively eliminate the noise from the original signals without leading the distortion. The sound's power spectrum mainly focused in 0~7.5kHz, and its envelope is relatively concentrated. The variation of signal energy, lying in 1.5kHz～4.5kHz, can accurately reflect the changes of penetraion in welding process. It can be seen that the arc sound and the penetration state is closely-related, and the explanation to the change of power and frequency spectrum characteristics along with the variation of penetration state.Based on short-time window technology, the characteristics of the arc sound were extracted from time-domain, frequency-domain, cepstrum-domain and geometric-domain, including the short-time energy E_n , the short-time average magnitude M_n, the short-time average zero-crossing ratio Z_n, the short-time zero-to-energy ratio ZER_n , the short-time power spectrum Pn , linear prediction(LP) coefficient{ a_i} , kurtosis coefficient KU, Mel frequency cepstrum coefficient (MFCC) and first-order difference△MFCC, second-order difference△△MFCC. The united character vector A Tat 60 dimension was established by integrating the wavelet packet frequency-band energy E5 kand was compressed by the PCA method, which overcomes multiple correlation among the parameters and realizes the dimentionality reduction of original charater space.The BP and RBF neural networks were applied for pattern recognition of the penetration state in MIG welding process via constructing the train and test samples, in which the input vectors were the sound's united character vector at 10 dimension reduced by PCA analysis and the output vectors ware the penetration state, such as non-penetration, little-penetration, full-penetration, excessive-penetration and burn through.
|
PDF Full Text Request