Data augmentation by an electro-vibratory feedback improves signal interpretation by neural networks

Data-driven machine learning algorithms represent an alternative to numerical or analytical modeling of vibrating objects for the purpose of extracting their material or geometric properties. Their performance is conditioned by the quality of the training dataset, which should contain all possible production configurations of the object and measurement conditions. In this work, we propose a data-augmentation method based on manipulations of the vibration properties, rather than numerical signal post-processing. Instead of modifying the sample's material or geometry, we connect an additional sensor and actuator to the sample and use low-latency electronics to control the feedback signal. This delivers us an electro-mechanical system with tunable oscillation properties. This signal diversification method is demonstrated on the classification of correct and incorrect adhesion of two aluminum plates. The labeling accuracy of 67 % on nine unknown test objects, when trained on a single unchangeable training object, increases to 83 % when the variety of the training vibration signals is augmented by varying the parameters of the electro-vibratory feedback. This provides advantages for industrial quality control, where a large number of reference training samples are not available.