Variance stabilised optimisation of neural networks: a case study in additive manufacturing

A new framework is presented for training neural networks that is based on the characterisation and stabilisation of measurement variations. The framework results in a number of useful properties that maximises the use of data as well as aiding in the interpretation of results in a principled manner. This is achieved via variance stabilisation and a subsequent standardisation step. The method is a general approach that may be used in any context where repeatability data is available. Standardisation in this manner allows goodness of fit to be quantified and measurement data to be interpreted from a statistical perspective. We demonstrate the utility of this framework in the analysis of advanced manufacturing data.