Reduced Training Data for Robust Ultrasound Signal Interpretation by Neural Networks

Abstract

Abstract We introduce an experimental setup to generate large datasets of ultrasonic signals relevant for quality inspection. A reconfigurable defect is induced by a heating laser projecting a short line on a shape memory polymer foil, which has a special property that its Young’s modulus and ultrasound attenuation level can locally be controlled by its temperature field. Ultrasound is generated by a laser pulse at one fixed position and detected by a laser vibrometer at another fixed position for 64 different defect positions and 3 different configurations of the specimen. The obtained diversified datasets are used to optimize the network architecture for the interpretation of ultrasound signals. We study the robustness of the model in cases of reduced and dissimilar training datasets. In our first study, we classify the specimen configurations with the defect position being the disturbing parameter. The model shows high performance on a dataset of signals obtained at all the defect positions, even if only trained on a completely different dataset containing signals obtained at few defect positions. In our second study, we perform precise defect localization. The model becomes robust to the changes in the specimen configuration when a reduced dataset, containing signals obtained at two different specimen configurations, is used for the training process. These conclusions show the great potential of the demonstrated machine learning algorithm for industrial quality control. High-volume products (simulated by a reconfigurable specimen in our work) can be rapidly tested on the production line using this single-point and contact-free ultrasonic method.