Development of a predictive maintenance system for electrodes dressing in welding guns

The presented work of thesis arises from the desire of the ISI-Welding Company to have a reliable system of predictive maintenance for the dressing of electrodes in welding guns. The promise of predictive maintenance is to allow a convenient scheduling of corrective maintenance and to prevent unexpected equipment failures with the goal to extend product useful life. Machine learning algorithms lend well for this purpose permitting, based on fairly large data sets, to find common relations between the data and to identify a common pattern for a given unsafe situation. The first step has been to investigate the nature of spot welding phenomenon finding both e polynomial and a physical model. The polynomial model is achieved with ARX, ARMAX and OE methodologies and the best results are chosen after a residuals analysis and an evaluation of RMSE and Best Fit. The physical model is obtained with a gray-box approach: the physics of the welding phenomenon has been studied and adapted to the known data in order to get the best possible approximation. Once acquired a great knowledge on the welding procedure, the study has moved to the recognition of a predictive maintenance system to identify the necessity of the electrodes to be dressed (milling in order to recover the nominal electrical qualities and the geometry needed to work in an optimal state). After several attempts the solution turned out to be the usage of statistical descriptor. As result of a data analysis, some precise statistical descriptors have been chosen in order to give, to machine learning algorithms, a precise description of the current and voltage curves without passing the data millisecond to millisecond. The algorithms applied are: one layer neural network, multi-layers neural network and decision trees. The results during the test phase have been really encouraging, the algorithm are also able to recognize the false-safe conditions.