Challenges in Design and Development of Electron Beam Weld Predictor Software

Electron beam welding (EBW) is high power density welding process used predominantly for fabrication of launch vehicles and spacecrafts made of aerospace grade materials. EBW process at LPSC-ISRO is a space qualified process and the cost and time involved in imbibing quality into the EBW process amounts to 65% of the total fabrication costs and 80 % of the total welding time. The weld quality is ensured by evaluating a pre-weld WTS (Weld Test Specimen), each time a weld is done.. The LPSC-ISRO-Bangalore, houses many variants of EBW machines and in order to tap, utilise the huge potential of empirical data available in-house, this software has been designed and developed to predict the EBW machine parameters for a given material and geometry details. The software developed is based on machine learning of the empirical data. The accuracy of prediction results have been validated using EBW experiments and presently the first version of the software is available for deployment in ISRO and continual improvement of prediction accuracies is ensured by enhancing the database with additional reference points. The software aids in reduction of total manufacturing cost and lead time as the WTS (Weld Test Specimen) requirements have been substantially reduced with aid of this software. Unlike software development for electrical and electronic systems, the challenges in design and development of software for mechanical systems (EBW machines in this design) commences with limited data availability in assessable form; data collection and assimilation are humongous as it is based on almost 20 nos of parameters related to EBW machine and weld quality and not all parameters are in measurable or assessable format. Despite this limitation, the software has been developed in two phases including feature extraction studies in phase2 thus enhancing the prediction accuracies achieved in phase1(a primitive and draft software design). Also the other challenges are input and output parameters for the software are in analog format & knowledge based (interpretation needs skill); the empirical data is predominantly dependent on physics of EBW such as geometry, material properties and software accuracy to be validated by real-time physical EBW experiments. The challenges have been overcome by ensuring precision in empirical data collection, assimilation and cleansing and also feature extraction studies indicated the relevance in selection of input parameters. The selection of machine learning algorithm is also based on the applicability of the selected algorithm to address all the above challenges endured. Also the experimental results validated the prediction accuracies achieved via the Phase2 developed software. Despite the challenges endured, the indigenously designed, developed and validated EBW predictor software can predict EBW machine parameters (beam current and focus current) without any need to weld a WTS and hence, the software developed aids in reduction of the total manufacturing cost and also the lead time involved in realisation of launch vehicles and spacecrafts for various ISRO missions have been substantially reduced with induction of this EBW Predictor software.