The use of arc sound & on-line ultrasonic signal processing on computer technology in welding
conference contributionposted on 31.08.2017, 11:21 by John McCardleJohn McCardle, S. Swallow, Raymond Stroud, K.T. Burge
Monitoring the welding process on-line with ultrasound is problematic, but promises great rewards. A fast classifier is required to exploit the redundancy available in ultrasonic interrogation and ensure an adequate signal / noise ratio. TARDIS is such a classifier, using logical neural network techniques and dedicated hardware. The classification performance of TARDIS alone is noisy, but exceptionally fast. This speed of operation can be used to offset the fuzziness of individual classifications, using higher order correlations. The expert manual welder is capable of simultaneously monitoring visual and acoustic data and, coupled with a knowledge of the process and past experience, is able to attempt an optimum weld. Observations of skilled manual welders has shown a subconscious tendency to change the angle of the electrode and length of arc by listening to adverse fluctuations in the process noise in addition to visual assessment. This has resulted in much research into the analysis of airborne acoustic emissions (AEs) of welding processes. It is evident that to artificially copy these skills requires a fast, robust signal processing and pattern recognition technique similar the known architecture and operation of the brain. The Department of Design, Brunel University, has been researching the possibilities of including the monitoring of airborne acoustic emissions as an additional correcting factor in automated weld process control. Salient relationships between acoustic emissions and process parameters using off-line statistical techniques has been established, however, real time application remains problematic due to the computational intensity of such methods. Statistical approaches to the interpretation of arc sounds relies on the direct correlation observable between the acquired signal or its' various transforms and the monitored process parameters. The method is a time consuming and often mathematically gruelling . Artificial neural networks (ANNs) provide an alternative. By the construction of different architectures and the application of various learning algorithms ANNs can provide a noise tolerant adaptive knowledge acquisition system. The work discussed in this paper illustrates the methods of signal preprocessing and utilisation of artificial neural networks to interpret arc sounds. Techniques are used to filter and compress high dimensional erratic data patterns to form classifiable representations of the process state. Real time scenarios are discussed together with commercially viable hardware solutions.