Prediction of magnetic interference from resistance spot welding processes on implantable cardioverter-defibrillators

Abstract

The purpose of this study is to create a tool capable of reliably predicting the magnetic fields (MFs) generated in resistance spot welding (RSW) processes. If the MF is greater than 10 G, the MF activates a reed switch in the implantable cardioverter-defibrillator (ICD) and temporarily inhibits detection of ventricular tachycardia and ventricular fibrillation. The MF is predicted from two variables: (1) distance between ICD and electrodes; (2) welding current. Firstly, a linear regression model is attempted. The MF is the dependent variable, 1/distance and welding current are the independent variables. The ANOVA table shows that the regression is significant at the 0.01 level but the residuals analysis reveals nonlinear behaviour. An artificial neural network (ANN) is proposed because the ANNs are capable of mapping nonlinear systems. The inputs are two-component vectors, the first component is the distance value and the second component is the welding current value. The training of the ANN uses supervised learning mechanism. Therefore, each input must come with its respective desired output (target) that is the recorded MF. The available data set is randomly split into a training subset (to update weight values) and a validation subset (to guard against overfitting by means of cross validation). The number of neurons in the hidden layers is selected considering the overfitting phenomenon: the number of neurons in the hidden layers that minimizes the validation mean square error (MSE) is 4. With the selected ANN (2-4-4-1) the aim of the present study (to reliably predict the MF) is achieved because this ANN produces good results in prediction from input vectors non-used in the training.