ACOUSTIC SENSING FOR METAL TRANSFER MODE AND PENETRATION STATE CLASSIFICATION OF GMAW USING ARTIFICIAL NEURAL NETWORKS

ACOUSTIC SENSING FOR METAL TRANSFER MODE AND PENETRATION STATE CLASSIFICATION OF GMAW USING ARTIFICIAL NEURAL NETWORKS: An investigation of the acoustic response of GMAW's capabilities within machine learning-based weld quality monitoring

Bidstrup, Anders

4. term

Manufacturing Technology, Master

2017

2017-06-02

79

Dette studie undersøger, om lyden fra Gas Metal Arc Welding (GMAW) kan bruges sammen med maskinlæring til at overvåge svejsekvalitet. Vi trænede kunstige neurale netværk (ANN'er) til at genkende to metaloverførselsformer—globulær overførsel og kortslutningsoverførsel—samt tre penetrationsstadier: manglende penetration, fuld penetration og overdreven penetration. Metaloverførsel beskriver, hvordan smeltet metal bevæger sig fra tråden til svejsen, og penetration beskriver, hvor dybt materialerne smelter sammen. Fra korte tidsvinduer af svejselyden beregnede vi 1.166 numeriske kendetegn (features) for signalet, herunder tidslige mål, spektral form, harmonisk indhold, perceptuelle mål og statistikker fra en wavelet-pakke-dekomposition, som opdeler et signal i tids-frekvens-komponenter. Data blev indsamlet i en robotiseret GMAW-celle ved bevidst at fremkalde hver ønsket tilstand. Herefter forbehandlede vi signalerne og trænede 110 ANN-konfigurationer på tværs af 15 kombinationer af vinduesstørrelse og overlap ved hjælp af to træningsmetoder: gradientnedstigning med adaptiv læringsrate (GDA) og skaleret konjugeret gradient (SCG). Resultaterne viser, at ANN’er trænet med SCG kan klassificere de tre penetrationsstadier, mens GDA kun delvist lykkedes. For de to metaloverførselsformer var resultaterne ikke entydige, men de viste en tendens til korrekte forudsigelser.

This study explores whether the sound produced during Gas Metal Arc Welding (GMAW) can be used with machine learning to monitor weld quality. We trained artificial neural networks (ANNs) to recognize two metal transfer modes—globular transfer and short-circuit transfer—and three penetration states: lack of penetration, full penetration, and excessive penetration. Metal transfer mode describes how molten metal moves from the wire to the weld, and penetration describes how deeply the weld fuses the materials. From short time windows of the welding sound, we computed 1,166 numerical descriptors (features), covering timing information, spectral shape, harmonic content, perceptual measures, and statistics derived from a wavelet packet decomposition, which breaks a signal into time–frequency components. We collected data in a robotic GMAW cell by deliberately producing each target condition, then preprocessed the signals and trained 110 ANN configurations across 15 window sizes and overlaps using two training methods: gradient descent with adaptive learning rate (GDA) and scaled conjugate gradient (SCG). Results show that ANNs trained with SCG can classify the three penetration states, while GDA achieved only partial success. For the two metal transfer modes, results were inconclusive, though they showed a tendency toward correct predictions.

[This abstract was generated with the help of AI]

Maskinlæring ; Neuralt netværk ; monitorering ; feature extraction ; GMAW ; MIG ; MAG

Download
View record in AAU Student Projects