Manufacturing Process Monitoring Using Two-step Feature Selection and Classifier Fusion.

[摘要] Two-step feature selection and classifier fusion methods have been investigated for coroning and ultrasonic welding processes.First, an acoustic emission (AE) based feature selection is studied for monitoring the coroning process. This involves significant data reduction since the AE signal requires a high sampling rate in order to capture useful information. Features, which are spectral components in the frequency domain, are processed using the class mean scatter (CMS) criterion. The features that are selected in the first step are then combined to reduce dimensions without averaging in the second step. With these features, classification is then performed and the results compared with those obtained using a conventional feature selection method.Next, a classifier fusion method is developed to enhance the reliability and robustness in decision making. This is based on state performance weighting, which incorporates information on performance of classifiers for each state. A penalty voting concept is also investigated to further enhance classifier performance. Using equal weighting for each of the classifiers investigated, the overall classification rate achieved is 87.7%, while with state performance weighting, the classification rate improves to 98.5%. Using penalty voting further enhances the performance to 99.7%.The signal processing techniques developed is then further used to investigate the feasibility of real time monitoring of ultrasonic weld quality using audible sound. A two degree-of-freedom model of the system is developed to help provide better understanding of the process characteristics. A series of experiments is also conducted to define the robust weld quality range for metal welding using the T-peel test. The relationship between weld quality and sound signals generated is then analyzed, and a strong correlation is obtained between the stiffness variation and spectral components of the audible sound. The results show that a good weld is dominated by one frequency component in the audible sound range near 10 kHz, which is half the vibration frequency of 20 kHz, for the condition used, while that for a cold weld is characterized by two frequency components at 9 and 11 kHz. Over weld conditions do not generate unique frequency components.

[发布机构] University of Michigan

[效力级别] Ultrasonic Welding

[关键词] Monitoring;Ultrasonic Welding;Coroning;Two-step Feature Selection;Classifier Fusion;Mechanical Engineering;Engineering;Mechanical Engineering