Multi-block deep regression with reinforced gradients: Modelling the estimation of direct weld parameters in automated welding

Abstract

This paper presents a deep regression model to estimation of the weld bead parameters in welding tasks. It is an aggregate of deepregression blocks where number of these blocks is proportional to the cardinality of the weld parameters. These blocks are trainedsimultaneously and share an identical structure with four-hidden-layer Sigmoid activation functions and a linear transformation attheir outputs. Moreover, they incorporate a new meta-parameter, shared by all the hidden layers of a given block, to maintain thequality of the gradients of their respective weight matrices. This allows the model to further reduce the deviation of its estimatesfrom the expected values of the weld parameters to significantly minimize its estimation error. The evaluation of the performanceof this approach in contrast to state-of-the-art techniques in the literature shows a significant improvement in estimating thesevalues for different welding processes. Furthermore, the proposed deep regression network is capable of retaining its performancewhen presented with combined data of different welding techniques. This is a nontrivial result in attaining an scalable modelwhose quality of estimation is independent of adopted welding techniques.