Neural Network Training Techniques for Enhancing the Performance of a Distributive Tactile Sensor

Abstract

This work describes neural network training techniques for enhancing the performance of a onedimensional beam for determining load positions. The system was a distributive system which relied on the detection of changes in the surface properties that can be seen across the surface. The demonstrated distributive system was 400 mm in length. The applied load positions within the range 60 -300 mm could be determined with an average percentage error of 0.2% of the beam length which corresponded to a position error of 0.8 mm using a network trained with 10 training positions. It was found that the errors were higher for the load applied near the edges of the beam, leading to an average percentage error of 3.6% for the whole length of the beam. The normalization of the network output can be employed to reduce the average percentage error by approximately 1% for a given number of training positions. The performance was improved by introducing more training positions in the less sensitive area. The described training technique not only reduced the prediction error but also enlarged the areas where the prediction errors were satisfactorily small. Using the described technique, the overall prediction error could be decreased by 0.13 -0.24% of the beam length for 30 -10 training positions.The area where the prediction errors were within 2 mm was increased by 14.5 -7.3% of the beam length for a network trained with 4 training positions in the middle portion using 10 -30 training positions.