Neural Network Based Inspection of Machined Surfaces Using Multiple Sensors

Abstract

The dominant idea of this research is that the measure of the topology or the quality of surfaces is best obtained by the utilization of a set of different sensors whose outputs are collectively examined by a neural network. Features are extracted from the output of each of these sensors which are then used as inputs to a hierarchical neural net. The net is “trained” by a selected training set of machined or otherwise prepared surfaces whose quality has already been independently established. These samples are repeatedly presented to the sensors and the network will, each time, make a decision about the surface roughness which is then compared to the correct answer, and the error used to modify the connection weights. Following this training period, the net will be able to identify the quality of new surfaces presented to it. Indeed, it should be able to do so even when some information is suppressed by a faulty sensor, or in the presence of noise which is likely to be caused by poor illumination. Examination of the internal connections of the neural network will then reveal the relative importance of each of the sensors used in the inspection process. Thus, from a generic inspection system it is expected that an optimum one may be obtained for a specific production line. Results, using a laser scattering technique, from a set of prepared surfaces are discussed with regard to fusion of different features in order to obtain an adequate measure of surface roughness using the hierarchical neural network.