Monitoring and diagnosis of resistance spot welding process.

Resistance spot welding is a predominant sheet metal joining method in industry. Although widely used for decades, it has a great weakness that is the inability of consistently producing quality welds. As a result, extra welds have to be made to ensure the confidence of welded structures. The objective of this research is to improve the quality of the resistance spot welding process through on-line monitoring and diagnosis. Three significant accomplishments have been achieved. They are expulsion detection and estimation, on-line nugget size estimation, and diagnosis of multiple simultaneous faults. Expulsions cause major defect in the weld quality. On-line expulsion detection and estimation can feedback useful information such that the process parameters can be adjusted or other remedial actions can be taken. In this study, the on-line algorithm is developed based on a generalized likelihood ratio test (GLRT) using electrode force signals. An adaptive noise cancellation (ANC) technique is introduced to eliminate the induced noise. The performance of this algorithm has been demonstrated with the experimental results, where a 100% detection rate is achieved with a properly chosen threshold. A multi-variate process model is developed for on-line nugget size estimation. The features used consist of those extracted from both the controllable process variables and the on-line signals. A systematic feature selection procedure is developed based on the principal component analysis (PCA). The three commonly used on-line signals, dynamic resistance, force, and displacement, have been proven to carry similar information. Thus, only dynamic resistance is sufficient for the model. In model training, the effect of electrode wear on the process is explicitly considered. The obtained nugget size estimation model has been demonstrated to be robust over a wide range of welding conditions. The average relative estimation error is less than ten percent for the conditions tested. Expulsions and undersized welds are often caused by process faults. In this research, typical process faults have been identified and their levels are defined. Due to the fact that multiple faults often occur simultaneously and their interactions are significant, fault patterns cannot be assumed to be linearly additive. On the other hand, the number of tests can easily become intimidating if all the possible fault patterns were to be obtained experimentally. A new fault diagnosis approach is proposed in this study to deal with multiple simultaneous faults. Both linear and non-linear effects on the fault patterns are considered. However, only some of the patterns need to be obtained through the designed experiments. Others can be predicted using a functional regression model developed based on a Bootstrap model-selection procedure. Given a complete set of the fault patterns, a minimum distance classifier can be designed to classify the multiple simultaneous faults. The proposed approach has been demonstrated with both the simulation and experimental results. An 88% success rate is achieved for an example from the resistance spot welding process.Ph.D.Applied SciencesEngineering, Materials scienceMechanical engineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/131708/2/9929881.pd